int sbdsdc_(char *uplo, char *compq, int *n, float *d__,
float *e, float *u, int *ldu, float *vt, int *ldvt, float *q,
int *iq, float *work, int *iwork, int *info)
{
/* System generated locals */
int u_dim1, u_offset, vt_dim1, vt_offset, i__1, i__2;
float r__1;
/* Builtin functions */
double r_sign(float *, float *), log(double);
/* Local variables */
int i__, j, k;
float p, r__;
int z__, ic, ii, kk;
float cs;
int is, iu;
float sn;
int nm1;
float eps;
int ivt, difl, difr, ierr, perm, mlvl, sqre;
extern int lsame_(char *, char *);
int poles;
extern int slasr_(char *, char *, char *, int *,
int *, float *, float *, float *, int *);
int iuplo, nsize, start;
extern int scopy_(int *, float *, int *, float *,
int *), sswap_(int *, float *, int *, float *, int *
), slasd0_(int *, int *, float *, float *, float *, int *
, float *, int *, int *, int *, float *, int *);
extern double slamch_(char *);
extern int slasda_(int *, int *, int *,
int *, float *, float *, float *, int *, float *, int *,
float *, float *, float *, float *, int *, int *, int *,
int *, float *, float *, float *, float *, int *, int *),
xerbla_(char *, int *);
extern int ilaenv_(int *, char *, char *, int *, int *,
int *, int *);
extern int slascl_(char *, int *, int *, float *,
float *, int *, int *, float *, int *, int *);
int givcol;
extern int slasdq_(char *, int *, int *, int
*, int *, int *, float *, float *, float *, int *, float *
, int *, float *, int *, float *, int *);
int icompq;
extern int slaset_(char *, int *, int *, float *,
float *, float *, int *), slartg_(float *, float *, float *
, float *, float *);
float orgnrm;
int givnum;
extern double slanst_(char *, int *, float *, float *);
int givptr, qstart, smlsiz, wstart, smlszp;
/* -- LAPACK routine (version 3.2) -- */
/* Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */
/* November 2006 */
/* .. Scalar Arguments .. */
/* .. */
/* .. Array Arguments .. */
/* .. */
/* Purpose */
/* ======= */
/* SBDSDC computes the singular value decomposition (SVD) of a float */
/* N-by-N (upper or lower) bidiagonal matrix B: B = U * S * VT, */
/* using a divide and conquer method, where S is a diagonal matrix */
/* with non-negative diagonal elements (the singular values of B), and */
/* U and VT are orthogonal matrices of left and right singular vectors, */
/* respectively. SBDSDC can be used to compute all singular values, */
/* and optionally, singular vectors or singular vectors in compact form. */
/* This code makes very mild assumptions about floating point */
/* arithmetic. It will work on machines with a guard digit in */
/* add/subtract, or on those binary machines without guard digits */
/* which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. */
/* It could conceivably fail on hexadecimal or decimal machines */
/* without guard digits, but we know of none. See SLASD3 for details. */
/* The code currently calls SLASDQ if singular values only are desired. */
/* However, it can be slightly modified to compute singular values */
/* using the divide and conquer method. */
/* Arguments */
/* ========= */
/* UPLO (input) CHARACTER*1 */
/* = 'U': B is upper bidiagonal. */
/* = 'L': B is lower bidiagonal. */
/* COMPQ (input) CHARACTER*1 */
/* Specifies whether singular vectors are to be computed */
/* as follows: */
/* = 'N': Compute singular values only; */
/* = 'P': Compute singular values and compute singular */
/* vectors in compact form; */
/* = 'I': Compute singular values and singular vectors. */
/* N (input) INTEGER */
/* The order of the matrix B. N >= 0. */
/* D (input/output) REAL array, dimension (N) */
/* On entry, the n diagonal elements of the bidiagonal matrix B. */
/* On exit, if INFO=0, the singular values of B. */
/* E (input/output) REAL array, dimension (N-1) */
/* On entry, the elements of E contain the offdiagonal */
/* elements of the bidiagonal matrix whose SVD is desired. */
/* On exit, E has been destroyed. */
/* U (output) REAL array, dimension (LDU,N) */
/* If COMPQ = 'I', then: */
/* On exit, if INFO = 0, U contains the left singular vectors */
/* of the bidiagonal matrix. */
/* For other values of COMPQ, U is not referenced. */
/* LDU (input) INTEGER */
/* The leading dimension of the array U. LDU >= 1. */
/* If singular vectors are desired, then LDU >= MAX( 1, N ). */
/* VT (output) REAL array, dimension (LDVT,N) */
/* If COMPQ = 'I', then: */
/* On exit, if INFO = 0, VT' contains the right singular */
/* vectors of the bidiagonal matrix. */
/* For other values of COMPQ, VT is not referenced. */
/* LDVT (input) INTEGER */
/* The leading dimension of the array VT. LDVT >= 1. */
/* If singular vectors are desired, then LDVT >= MAX( 1, N ). */
/* Q (output) REAL array, dimension (LDQ) */
/* If COMPQ = 'P', then: */
/* On exit, if INFO = 0, Q and IQ contain the left */
/* and right singular vectors in a compact form, */
/* requiring O(N log N) space instead of 2*N**2. */
/* In particular, Q contains all the REAL data in */
/* LDQ >= N*(11 + 2*SMLSIZ + 8*INT(LOG_2(N/(SMLSIZ+1)))) */
/* words of memory, where SMLSIZ is returned by ILAENV and */
/* is equal to the maximum size of the subproblems at the */
/* bottom of the computation tree (usually about 25). */
/* For other values of COMPQ, Q is not referenced. */
/* IQ (output) INTEGER array, dimension (LDIQ) */
/* If COMPQ = 'P', then: */
/* On exit, if INFO = 0, Q and IQ contain the left */
/* and right singular vectors in a compact form, */
/* requiring O(N log N) space instead of 2*N**2. */
/* In particular, IQ contains all INTEGER data in */
/* LDIQ >= N*(3 + 3*INT(LOG_2(N/(SMLSIZ+1)))) */
/* words of memory, where SMLSIZ is returned by ILAENV and */
/* is equal to the maximum size of the subproblems at the */
/* bottom of the computation tree (usually about 25). */
/* For other values of COMPQ, IQ is not referenced. */
/* WORK (workspace) REAL array, dimension (MAX(1,LWORK)) */
/* If COMPQ = 'N' then LWORK >= (4 * N). */
/* If COMPQ = 'P' then LWORK >= (6 * N). */
/* If COMPQ = 'I' then LWORK >= (3 * N**2 + 4 * N). */
/* IWORK (workspace) INTEGER array, dimension (8*N) */
/* INFO (output) INTEGER */
/* = 0: successful exit. */
/* < 0: if INFO = -i, the i-th argument had an illegal value. */
/* > 0: The algorithm failed to compute an singular value. */
/* The update process of divide and conquer failed. */
/* Further Details */
/* =============== */
/* Based on contributions by */
/* Ming Gu and Huan Ren, Computer Science Division, University of */
/* California at Berkeley, USA */
/* ===================================================================== */
/* Changed dimension statement in comment describing E from (N) to */
/* (N-1). Sven, 17 Feb 05. */
/* ===================================================================== */
/* .. Parameters .. */
/* .. */
/* .. Local Scalars .. */
/* .. */
/* .. External Functions .. */
/* .. */
/* .. External Subroutines .. */
/* .. */
/* .. Intrinsic Functions .. */
/* .. */
/* .. Executable Statements .. */
/* Test the input parameters. */
/* Parameter adjustments */
--d__;
--e;
u_dim1 = *ldu;
u_offset = 1 + u_dim1;
u -= u_offset;
vt_dim1 = *ldvt;
vt_offset = 1 + vt_dim1;
vt -= vt_offset;
--q;
--iq;
--work;
--iwork;
/* Function Body */
*info = 0;
iuplo = 0;
if (lsame_(uplo, "U")) {
iuplo = 1;
}
if (lsame_(uplo, "L")) {
iuplo = 2;
}
if (lsame_(compq, "N")) {
icompq = 0;
} else if (lsame_(compq, "P")) {
icompq = 1;
} else if (lsame_(compq, "I")) {
icompq = 2;
} else {
icompq = -1;
}
if (iuplo == 0) {
*info = -1;
} else if (icompq < 0) {
*info = -2;
} else if (*n < 0) {
*info = -3;
} else if (*ldu < 1 || icompq == 2 && *ldu < *n) {
*info = -7;
} else if (*ldvt < 1 || icompq == 2 && *ldvt < *n) {
*info = -9;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SBDSDC", &i__1);
return 0;
}
/* Quick return if possible */
if (*n == 0) {
return 0;
}
smlsiz = ilaenv_(&c__9, "SBDSDC", " ", &c__0, &c__0, &c__0, &c__0);
if (*n == 1) {
if (icompq == 1) {
q[1] = r_sign(&c_b15, &d__[1]);
q[smlsiz * *n + 1] = 1.f;
} else if (icompq == 2) {
u[u_dim1 + 1] = r_sign(&c_b15, &d__[1]);
vt[vt_dim1 + 1] = 1.f;
}
d__[1] = ABS(d__[1]);
return 0;
}
nm1 = *n - 1;
/* If matrix lower bidiagonal, rotate to be upper bidiagonal */
/* by applying Givens rotations on the left */
wstart = 1;
qstart = 3;
if (icompq == 1) {
scopy_(n, &d__[1], &c__1, &q[1], &c__1);
i__1 = *n - 1;
scopy_(&i__1, &e[1], &c__1, &q[*n + 1], &c__1);
}
if (iuplo == 2) {
qstart = 5;
wstart = (*n << 1) - 1;
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
if (icompq == 1) {
q[i__ + (*n << 1)] = cs;
q[i__ + *n * 3] = sn;
} else if (icompq == 2) {
work[i__] = cs;
work[nm1 + i__] = -sn;
}
/* L10: */
}
}
/* If ICOMPQ = 0, use SLASDQ to compute the singular values. */
if (icompq == 0) {
slasdq_("U", &c__0, n, &c__0, &c__0, &c__0, &d__[1], &e[1], &vt[
vt_offset], ldvt, &u[u_offset], ldu, &u[u_offset], ldu, &work[
wstart], info);
goto L40;
}
/* If N is smaller than the minimum divide size SMLSIZ, then solve */
/* the problem with another solver. */
if (*n <= smlsiz) {
if (icompq == 2) {
slaset_("A", n, n, &c_b29, &c_b15, &u[u_offset], ldu);
slaset_("A", n, n, &c_b29, &c_b15, &vt[vt_offset], ldvt);
slasdq_("U", &c__0, n, n, n, &c__0, &d__[1], &e[1], &vt[vt_offset]
, ldvt, &u[u_offset], ldu, &u[u_offset], ldu, &work[
wstart], info);
} else if (icompq == 1) {
iu = 1;
ivt = iu + *n;
slaset_("A", n, n, &c_b29, &c_b15, &q[iu + (qstart - 1) * *n], n);
slaset_("A", n, n, &c_b29, &c_b15, &q[ivt + (qstart - 1) * *n], n);
slasdq_("U", &c__0, n, n, n, &c__0, &d__[1], &e[1], &q[ivt + (
qstart - 1) * *n], n, &q[iu + (qstart - 1) * *n], n, &q[
iu + (qstart - 1) * *n], n, &work[wstart], info);
}
goto L40;
}
if (icompq == 2) {
slaset_("A", n, n, &c_b29, &c_b15, &u[u_offset], ldu);
slaset_("A", n, n, &c_b29, &c_b15, &vt[vt_offset], ldvt);
}
/* Scale. */
orgnrm = slanst_("M", n, &d__[1], &e[1]);
if (orgnrm == 0.f) {
return 0;
}
slascl_("G", &c__0, &c__0, &orgnrm, &c_b15, n, &c__1, &d__[1], n, &ierr);
slascl_("G", &c__0, &c__0, &orgnrm, &c_b15, &nm1, &c__1, &e[1], &nm1, &
ierr);
eps = slamch_("Epsilon");
mlvl = (int) (log((float) (*n) / (float) (smlsiz + 1)) / log(2.f)) + 1;
smlszp = smlsiz + 1;
if (icompq == 1) {
iu = 1;
ivt = smlsiz + 1;
difl = ivt + smlszp;
difr = difl + mlvl;
z__ = difr + (mlvl << 1);
ic = z__ + mlvl;
is = ic + 1;
poles = is + 1;
givnum = poles + (mlvl << 1);
k = 1;
givptr = 2;
perm = 3;
givcol = perm + mlvl;
}
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
if ((r__1 = d__[i__], ABS(r__1)) < eps) {
d__[i__] = r_sign(&eps, &d__[i__]);
}
/* L20: */
}
start = 1;
sqre = 0;
i__1 = nm1;
for (i__ = 1; i__ <= i__1; ++i__) {
if ((r__1 = e[i__], ABS(r__1)) < eps || i__ == nm1) {
/* Subproblem found. First determine its size and then */
/* apply divide and conquer on it. */
if (i__ < nm1) {
/* A subproblem with E(I) small for I < NM1. */
nsize = i__ - start + 1;
} else if ((r__1 = e[i__], ABS(r__1)) >= eps) {
/* A subproblem with E(NM1) not too small but I = NM1. */
nsize = *n - start + 1;
} else {
/* A subproblem with E(NM1) small. This implies an */
/* 1-by-1 subproblem at D(N). Solve this 1-by-1 problem */
/* first. */
nsize = i__ - start + 1;
if (icompq == 2) {
u[*n + *n * u_dim1] = r_sign(&c_b15, &d__[*n]);
vt[*n + *n * vt_dim1] = 1.f;
} else if (icompq == 1) {
q[*n + (qstart - 1) * *n] = r_sign(&c_b15, &d__[*n]);
q[*n + (smlsiz + qstart - 1) * *n] = 1.f;
}
d__[*n] = (r__1 = d__[*n], ABS(r__1));
}
if (icompq == 2) {
slasd0_(&nsize, &sqre, &d__[start], &e[start], &u[start +
start * u_dim1], ldu, &vt[start + start * vt_dim1],
ldvt, &smlsiz, &iwork[1], &work[wstart], info);
} else {
slasda_(&icompq, &smlsiz, &nsize, &sqre, &d__[start], &e[
start], &q[start + (iu + qstart - 2) * *n], n, &q[
start + (ivt + qstart - 2) * *n], &iq[start + k * *n],
&q[start + (difl + qstart - 2) * *n], &q[start + (
difr + qstart - 2) * *n], &q[start + (z__ + qstart -
2) * *n], &q[start + (poles + qstart - 2) * *n], &iq[
start + givptr * *n], &iq[start + givcol * *n], n, &
iq[start + perm * *n], &q[start + (givnum + qstart -
2) * *n], &q[start + (ic + qstart - 2) * *n], &q[
start + (is + qstart - 2) * *n], &work[wstart], &
iwork[1], info);
if (*info != 0) {
return 0;
}
}
start = i__ + 1;
}
/* L30: */
}
/* Unscale */
slascl_("G", &c__0, &c__0, &c_b15, &orgnrm, n, &c__1, &d__[1], n, &ierr);
L40:
/* Use Selection Sort to minimize swaps of singular vectors */
i__1 = *n;
for (ii = 2; ii <= i__1; ++ii) {
i__ = ii - 1;
kk = i__;
p = d__[i__];
i__2 = *n;
for (j = ii; j <= i__2; ++j) {
if (d__[j] > p) {
kk = j;
p = d__[j];
}
/* L50: */
}
if (kk != i__) {
d__[kk] = d__[i__];
d__[i__] = p;
if (icompq == 1) {
iq[i__] = kk;
} else if (icompq == 2) {
sswap_(n, &u[i__ * u_dim1 + 1], &c__1, &u[kk * u_dim1 + 1], &
c__1);
sswap_(n, &vt[i__ + vt_dim1], ldvt, &vt[kk + vt_dim1], ldvt);
}
} else if (icompq == 1) {
iq[i__] = i__;
}
/* L60: */
}
/* If ICOMPQ = 1, use IQ(N,1) as the indicator for UPLO */
if (icompq == 1) {
if (iuplo == 1) {
iq[*n] = 1;
} else {
iq[*n] = 0;
}
}
/* If B is lower bidiagonal, update U by those Givens rotations */
/* which rotated B to be upper bidiagonal */
if (iuplo == 2 && icompq == 2) {
slasr_("L", "V", "B", n, n, &work[1], &work[*n], &u[u_offset], ldu);
}
return 0;
/* End of SBDSDC */
} /* sbdsdc_ */
int sbdsqr_(char *uplo, int *n, int *ncvt, int *
nru, int *ncc, float *d__, float *e, float *vt, int *ldvt, float *
u, int *ldu, float *c__, int *ldc, float *work, int *info)
{
/* System generated locals */
int c_dim1, c_offset, u_dim1, u_offset, vt_dim1, vt_offset, i__1,
i__2;
float r__1, r__2, r__3, r__4;
double d__1;
/* Builtin functions */
double pow_dd(double *, double *), sqrt(double), r_sign(float *
, float *);
/* Local variables */
float f, g, h__;
int i__, j, m;
float r__, cs;
int ll;
float sn, mu;
int nm1, nm12, nm13, lll;
float eps, sll, tol, abse;
int idir;
float abss;
int oldm;
float cosl;
int isub, iter;
float unfl, sinl, cosr, smin, smax, sinr;
extern int srot_(int *, float *, int *, float *,
int *, float *, float *), slas2_(float *, float *, float *, float *,
float *);
extern int lsame_(char *, char *);
float oldcs;
extern int sscal_(int *, float *, float *, int *);
int oldll;
float shift, sigmn, oldsn;
int maxit;
float sminl;
extern int slasr_(char *, char *, char *, int *,
int *, float *, float *, float *, int *);
float sigmx;
int lower;
extern int sswap_(int *, float *, int *, float *,
int *), slasq1_(int *, float *, float *, float *, int *),
slasv2_(float *, float *, float *, float *, float *, float *, float *,
float *, float *);
extern double slamch_(char *);
extern int xerbla_(char *, int *);
float sminoa;
extern int slartg_(float *, float *, float *, float *, float *
);
float thresh;
int rotate;
float tolmul;
/* -- LAPACK routine (version 3.2) -- */
/* Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */
/* January 2007 */
/* .. Scalar Arguments .. */
/* .. */
/* .. Array Arguments .. */
/* .. */
/* Purpose */
/* ======= */
/* SBDSQR computes the singular values and, optionally, the right and/or */
/* left singular vectors from the singular value decomposition (SVD) of */
/* a float N-by-N (upper or lower) bidiagonal matrix B using the implicit */
/* zero-shift QR algorithm. The SVD of B has the form */
/* B = Q * S * P**T */
/* where S is the diagonal matrix of singular values, Q is an orthogonal */
/* matrix of left singular vectors, and P is an orthogonal matrix of */
/* right singular vectors. If left singular vectors are requested, this */
/* subroutine actually returns U*Q instead of Q, and, if right singular */
/* vectors are requested, this subroutine returns P**T*VT instead of */
/* P**T, for given float input matrices U and VT. When U and VT are the */
/* orthogonal matrices that reduce a general matrix A to bidiagonal */
/* form: A = U*B*VT, as computed by SGEBRD, then */
/* A = (U*Q) * S * (P**T*VT) */
/* is the SVD of A. Optionally, the subroutine may also compute Q**T*C */
/* for a given float input matrix C. */
/* See "Computing Small Singular Values of Bidiagonal Matrices With */
/* Guaranteed High Relative Accuracy," by J. Demmel and W. Kahan, */
/* LAPACK Working Note #3 (or SIAM J. Sci. Statist. Comput. vol. 11, */
/* no. 5, pp. 873-912, Sept 1990) and */
/* "Accurate singular values and differential qd algorithms," by */
/* B. Parlett and V. Fernando, Technical Report CPAM-554, Mathematics */
/* Department, University of California at Berkeley, July 1992 */
/* for a detailed description of the algorithm. */
/* Arguments */
/* ========= */
/* UPLO (input) CHARACTER*1 */
/* = 'U': B is upper bidiagonal; */
/* = 'L': B is lower bidiagonal. */
/* N (input) INTEGER */
/* The order of the matrix B. N >= 0. */
/* NCVT (input) INTEGER */
/* The number of columns of the matrix VT. NCVT >= 0. */
/* NRU (input) INTEGER */
/* The number of rows of the matrix U. NRU >= 0. */
/* NCC (input) INTEGER */
/* The number of columns of the matrix C. NCC >= 0. */
/* D (input/output) REAL array, dimension (N) */
/* On entry, the n diagonal elements of the bidiagonal matrix B. */
/* On exit, if INFO=0, the singular values of B in decreasing */
/* order. */
/* E (input/output) REAL array, dimension (N-1) */
/* On entry, the N-1 offdiagonal elements of the bidiagonal */
/* matrix B. */
/* On exit, if INFO = 0, E is destroyed; if INFO > 0, D and E */
/* will contain the diagonal and superdiagonal elements of a */
/* bidiagonal matrix orthogonally equivalent to the one given */
/* as input. */
/* VT (input/output) REAL array, dimension (LDVT, NCVT) */
/* On entry, an N-by-NCVT matrix VT. */
/* On exit, VT is overwritten by P**T * VT. */
/* Not referenced if NCVT = 0. */
/* LDVT (input) INTEGER */
/* The leading dimension of the array VT. */
/* LDVT >= MAX(1,N) if NCVT > 0; LDVT >= 1 if NCVT = 0. */
/* U (input/output) REAL array, dimension (LDU, N) */
/* On entry, an NRU-by-N matrix U. */
/* On exit, U is overwritten by U * Q. */
/* Not referenced if NRU = 0. */
/* LDU (input) INTEGER */
/* The leading dimension of the array U. LDU >= MAX(1,NRU). */
/* C (input/output) REAL array, dimension (LDC, NCC) */
/* On entry, an N-by-NCC matrix C. */
/* On exit, C is overwritten by Q**T * C. */
/* Not referenced if NCC = 0. */
/* LDC (input) INTEGER */
/* The leading dimension of the array C. */
/* LDC >= MAX(1,N) if NCC > 0; LDC >=1 if NCC = 0. */
/* WORK (workspace) REAL array, dimension (4*N) */
/* INFO (output) INTEGER */
/* = 0: successful exit */
/* < 0: If INFO = -i, the i-th argument had an illegal value */
/* > 0: */
/* if NCVT = NRU = NCC = 0, */
/* = 1, a split was marked by a positive value in E */
/* = 2, current block of Z not diagonalized after 30*N */
/* iterations (in inner while loop) */
/* = 3, termination criterion of outer while loop not met */
/* (program created more than N unreduced blocks) */
/* else NCVT = NRU = NCC = 0, */
/* the algorithm did not converge; D and E contain the */
/* elements of a bidiagonal matrix which is orthogonally */
/* similar to the input matrix B; if INFO = i, i */
/* elements of E have not converged to zero. */
/* Internal Parameters */
/* =================== */
/* TOLMUL REAL, default = MAX(10,MIN(100,EPS**(-1/8))) */
/* TOLMUL controls the convergence criterion of the QR loop. */
/* If it is positive, TOLMUL*EPS is the desired relative */
/* precision in the computed singular values. */
/* If it is negative, ABS(TOLMUL*EPS*sigma_max) is the */
/* desired absolute accuracy in the computed singular */
/* values (corresponds to relative accuracy */
/* ABS(TOLMUL*EPS) in the largest singular value. */
/* ABS(TOLMUL) should be between 1 and 1/EPS, and preferably */
/* between 10 (for fast convergence) and .1/EPS */
/* (for there to be some accuracy in the results). */
/* Default is to lose at either one eighth or 2 of the */
/* available decimal digits in each computed singular value */
/* (whichever is smaller). */
/* MAXITR INTEGER, default = 6 */
/* MAXITR controls the maximum number of passes of the */
/* algorithm through its inner loop. The algorithms stops */
/* (and so fails to converge) if the number of passes */
/* through the inner loop exceeds MAXITR*N**2. */
/* ===================================================================== */
/* .. Parameters .. */
/* .. */
/* .. Local Scalars .. */
/* .. */
/* .. External Functions .. */
/* .. */
/* .. External Subroutines .. */
/* .. */
/* .. Intrinsic Functions .. */
/* .. */
/* .. Executable Statements .. */
/* Test the input parameters. */
/* Parameter adjustments */
--d__;
--e;
vt_dim1 = *ldvt;
vt_offset = 1 + vt_dim1;
vt -= vt_offset;
u_dim1 = *ldu;
u_offset = 1 + u_dim1;
u -= u_offset;
c_dim1 = *ldc;
c_offset = 1 + c_dim1;
c__ -= c_offset;
--work;
/* Function Body */
*info = 0;
lower = lsame_(uplo, "L");
if (! lsame_(uplo, "U") && ! lower) {
*info = -1;
} else if (*n < 0) {
*info = -2;
} else if (*ncvt < 0) {
*info = -3;
} else if (*nru < 0) {
*info = -4;
} else if (*ncc < 0) {
*info = -5;
} else if (*ncvt == 0 && *ldvt < 1 || *ncvt > 0 && *ldvt < MAX(1,*n)) {
*info = -9;
} else if (*ldu < MAX(1,*nru)) {
*info = -11;
} else if (*ncc == 0 && *ldc < 1 || *ncc > 0 && *ldc < MAX(1,*n)) {
*info = -13;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SBDSQR", &i__1);
return 0;
}
if (*n == 0) {
return 0;
}
if (*n == 1) {
goto L160;
}
/* ROTATE is true if any singular vectors desired, false otherwise */
rotate = *ncvt > 0 || *nru > 0 || *ncc > 0;
/* If no singular vectors desired, use qd algorithm */
if (! rotate) {
slasq1_(n, &d__[1], &e[1], &work[1], info);
return 0;
}
nm1 = *n - 1;
nm12 = nm1 + nm1;
nm13 = nm12 + nm1;
idir = 0;
/* Get machine constants */
eps = slamch_("Epsilon");
unfl = slamch_("Safe minimum");
/* If matrix lower bidiagonal, rotate to be upper bidiagonal */
/* by applying Givens rotations on the left */
if (lower) {
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
work[i__] = cs;
work[nm1 + i__] = sn;
/* L10: */
}
/* Update singular vectors if desired */
if (*nru > 0) {
slasr_("R", "V", "F", nru, n, &work[1], &work[*n], &u[u_offset],
ldu);
}
if (*ncc > 0) {
slasr_("L", "V", "F", n, ncc, &work[1], &work[*n], &c__[c_offset],
ldc);
}
}
/* Compute singular values to relative accuracy TOL */
/* (By setting TOL to be negative, algorithm will compute */
/* singular values to absolute accuracy ABS(TOL)*norm(input matrix)) */
/* Computing MAX */
/* Computing MIN */
d__1 = (double) eps;
r__3 = 100.f, r__4 = pow_dd(&d__1, &c_b15);
r__1 = 10.f, r__2 = MIN(r__3,r__4);
tolmul = MAX(r__1,r__2);
tol = tolmul * eps;
/* Compute approximate maximum, minimum singular values */
smax = 0.f;
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Computing MAX */
r__2 = smax, r__3 = (r__1 = d__[i__], ABS(r__1));
smax = MAX(r__2,r__3);
/* L20: */
}
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Computing MAX */
r__2 = smax, r__3 = (r__1 = e[i__], ABS(r__1));
smax = MAX(r__2,r__3);
/* L30: */
}
sminl = 0.f;
if (tol >= 0.f) {
/* Relative accuracy desired */
sminoa = ABS(d__[1]);
if (sminoa == 0.f) {
goto L50;
}
mu = sminoa;
i__1 = *n;
for (i__ = 2; i__ <= i__1; ++i__) {
mu = (r__2 = d__[i__], ABS(r__2)) * (mu / (mu + (r__1 = e[i__ -
1], ABS(r__1))));
sminoa = MIN(sminoa,mu);
if (sminoa == 0.f) {
goto L50;
}
/* L40: */
}
L50:
sminoa /= sqrt((float) (*n));
/* Computing MAX */
r__1 = tol * sminoa, r__2 = *n * 6 * *n * unfl;
thresh = MAX(r__1,r__2);
} else {
/* Absolute accuracy desired */
/* Computing MAX */
r__1 = ABS(tol) * smax, r__2 = *n * 6 * *n * unfl;
thresh = MAX(r__1,r__2);
}
/* Prepare for main iteration loop for the singular values */
/* (MAXIT is the maximum number of passes through the inner */
/* loop permitted before nonconvergence signalled.) */
maxit = *n * 6 * *n;
iter = 0;
oldll = -1;
oldm = -1;
/* M points to last element of unconverged part of matrix */
m = *n;
/* Begin main iteration loop */
L60:
/* Check for convergence or exceeding iteration count */
if (m <= 1) {
goto L160;
}
if (iter > maxit) {
goto L200;
}
/* Find diagonal block of matrix to work on */
if (tol < 0.f && (r__1 = d__[m], ABS(r__1)) <= thresh) {
d__[m] = 0.f;
}
smax = (r__1 = d__[m], ABS(r__1));
smin = smax;
i__1 = m - 1;
for (lll = 1; lll <= i__1; ++lll) {
ll = m - lll;
abss = (r__1 = d__[ll], ABS(r__1));
abse = (r__1 = e[ll], ABS(r__1));
if (tol < 0.f && abss <= thresh) {
d__[ll] = 0.f;
}
if (abse <= thresh) {
goto L80;
}
smin = MIN(smin,abss);
/* Computing MAX */
r__1 = MAX(smax,abss);
smax = MAX(r__1,abse);
/* L70: */
}
ll = 0;
goto L90;
L80:
e[ll] = 0.f;
/* Matrix splits since E(LL) = 0 */
if (ll == m - 1) {
/* Convergence of bottom singular value, return to top of loop */
--m;
goto L60;
}
L90:
++ll;
/* E(LL) through E(M-1) are nonzero, E(LL-1) is zero */
if (ll == m - 1) {
/* 2 by 2 block, handle separately */
slasv2_(&d__[m - 1], &e[m - 1], &d__[m], &sigmn, &sigmx, &sinr, &cosr,
&sinl, &cosl);
d__[m - 1] = sigmx;
e[m - 1] = 0.f;
d__[m] = sigmn;
/* Compute singular vectors, if desired */
if (*ncvt > 0) {
srot_(ncvt, &vt[m - 1 + vt_dim1], ldvt, &vt[m + vt_dim1], ldvt, &
cosr, &sinr);
}
if (*nru > 0) {
srot_(nru, &u[(m - 1) * u_dim1 + 1], &c__1, &u[m * u_dim1 + 1], &
c__1, &cosl, &sinl);
}
if (*ncc > 0) {
srot_(ncc, &c__[m - 1 + c_dim1], ldc, &c__[m + c_dim1], ldc, &
cosl, &sinl);
}
m += -2;
goto L60;
}
/* If working on new submatrix, choose shift direction */
/* (from larger end diagonal element towards smaller) */
if (ll > oldm || m < oldll) {
if ((r__1 = d__[ll], ABS(r__1)) >= (r__2 = d__[m], ABS(r__2))) {
/* Chase bulge from top (big end) to bottom (small end) */
idir = 1;
} else {
/* Chase bulge from bottom (big end) to top (small end) */
idir = 2;
}
}
/* Apply convergence tests */
if (idir == 1) {
/* Run convergence test in forward direction */
/* First apply standard test to bottom of matrix */
if ((r__2 = e[m - 1], ABS(r__2)) <= ABS(tol) * (r__1 = d__[m], ABS(
r__1)) || tol < 0.f && (r__3 = e[m - 1], ABS(r__3)) <=
thresh) {
e[m - 1] = 0.f;
goto L60;
}
if (tol >= 0.f) {
/* If relative accuracy desired, */
/* apply convergence criterion forward */
mu = (r__1 = d__[ll], ABS(r__1));
sminl = mu;
i__1 = m - 1;
for (lll = ll; lll <= i__1; ++lll) {
if ((r__1 = e[lll], ABS(r__1)) <= tol * mu) {
e[lll] = 0.f;
goto L60;
}
mu = (r__2 = d__[lll + 1], ABS(r__2)) * (mu / (mu + (r__1 =
e[lll], ABS(r__1))));
sminl = MIN(sminl,mu);
/* L100: */
}
}
} else {
/* Run convergence test in backward direction */
/* First apply standard test to top of matrix */
if ((r__2 = e[ll], ABS(r__2)) <= ABS(tol) * (r__1 = d__[ll], ABS(
r__1)) || tol < 0.f && (r__3 = e[ll], ABS(r__3)) <= thresh) {
e[ll] = 0.f;
goto L60;
}
if (tol >= 0.f) {
/* If relative accuracy desired, */
/* apply convergence criterion backward */
mu = (r__1 = d__[m], ABS(r__1));
sminl = mu;
i__1 = ll;
for (lll = m - 1; lll >= i__1; --lll) {
if ((r__1 = e[lll], ABS(r__1)) <= tol * mu) {
e[lll] = 0.f;
goto L60;
}
mu = (r__2 = d__[lll], ABS(r__2)) * (mu / (mu + (r__1 = e[
lll], ABS(r__1))));
sminl = MIN(sminl,mu);
/* L110: */
}
}
}
oldll = ll;
oldm = m;
/* Compute shift. First, test if shifting would ruin relative */
/* accuracy, and if so set the shift to zero. */
/* Computing MAX */
r__1 = eps, r__2 = tol * .01f;
if (tol >= 0.f && *n * tol * (sminl / smax) <= MAX(r__1,r__2)) {
/* Use a zero shift to avoid loss of relative accuracy */
shift = 0.f;
} else {
/* Compute the shift from 2-by-2 block at end of matrix */
if (idir == 1) {
sll = (r__1 = d__[ll], ABS(r__1));
slas2_(&d__[m - 1], &e[m - 1], &d__[m], &shift, &r__);
} else {
sll = (r__1 = d__[m], ABS(r__1));
slas2_(&d__[ll], &e[ll], &d__[ll + 1], &shift, &r__);
}
/* Test if shift negligible, and if so set to zero */
if (sll > 0.f) {
/* Computing 2nd power */
r__1 = shift / sll;
if (r__1 * r__1 < eps) {
shift = 0.f;
}
}
}
/* Increment iteration count */
iter = iter + m - ll;
/* If SHIFT = 0, do simplified QR iteration */
if (shift == 0.f) {
if (idir == 1) {
/* Chase bulge from top to bottom */
/* Save cosines and sines for later singular vector updates */
cs = 1.f;
oldcs = 1.f;
i__1 = m - 1;
for (i__ = ll; i__ <= i__1; ++i__) {
r__1 = d__[i__] * cs;
slartg_(&r__1, &e[i__], &cs, &sn, &r__);
if (i__ > ll) {
e[i__ - 1] = oldsn * r__;
}
r__1 = oldcs * r__;
r__2 = d__[i__ + 1] * sn;
slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
work[i__ - ll + 1] = cs;
work[i__ - ll + 1 + nm1] = sn;
work[i__ - ll + 1 + nm12] = oldcs;
work[i__ - ll + 1 + nm13] = oldsn;
/* L120: */
}
h__ = d__[m] * cs;
d__[m] = h__ * oldcs;
e[m - 1] = h__ * oldsn;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &vt[
ll + vt_dim1], ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
+ 1], &u[ll * u_dim1 + 1], ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
+ 1], &c__[ll + c_dim1], ldc);
}
/* Test convergence */
if ((r__1 = e[m - 1], ABS(r__1)) <= thresh) {
e[m - 1] = 0.f;
}
} else {
/* Chase bulge from bottom to top */
/* Save cosines and sines for later singular vector updates */
cs = 1.f;
oldcs = 1.f;
i__1 = ll + 1;
for (i__ = m; i__ >= i__1; --i__) {
r__1 = d__[i__] * cs;
slartg_(&r__1, &e[i__ - 1], &cs, &sn, &r__);
if (i__ < m) {
e[i__] = oldsn * r__;
}
r__1 = oldcs * r__;
r__2 = d__[i__ - 1] * sn;
slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
work[i__ - ll] = cs;
work[i__ - ll + nm1] = -sn;
work[i__ - ll + nm12] = oldcs;
work[i__ - ll + nm13] = -oldsn;
/* L130: */
}
h__ = d__[ll] * cs;
d__[ll] = h__ * oldcs;
e[ll] = h__ * oldsn;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
nm13 + 1], &vt[ll + vt_dim1], ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u[ll *
u_dim1 + 1], ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &c__[
ll + c_dim1], ldc);
}
/* Test convergence */
if ((r__1 = e[ll], ABS(r__1)) <= thresh) {
e[ll] = 0.f;
}
}
} else {
/* Use nonzero shift */
if (idir == 1) {
/* Chase bulge from top to bottom */
/* Save cosines and sines for later singular vector updates */
f = ((r__1 = d__[ll], ABS(r__1)) - shift) * (r_sign(&c_b49, &d__[
ll]) + shift / d__[ll]);
g = e[ll];
i__1 = m - 1;
for (i__ = ll; i__ <= i__1; ++i__) {
slartg_(&f, &g, &cosr, &sinr, &r__);
if (i__ > ll) {
e[i__ - 1] = r__;
}
f = cosr * d__[i__] + sinr * e[i__];
e[i__] = cosr * e[i__] - sinr * d__[i__];
g = sinr * d__[i__ + 1];
d__[i__ + 1] = cosr * d__[i__ + 1];
slartg_(&f, &g, &cosl, &sinl, &r__);
d__[i__] = r__;
f = cosl * e[i__] + sinl * d__[i__ + 1];
d__[i__ + 1] = cosl * d__[i__ + 1] - sinl * e[i__];
if (i__ < m - 1) {
g = sinl * e[i__ + 1];
e[i__ + 1] = cosl * e[i__ + 1];
}
work[i__ - ll + 1] = cosr;
work[i__ - ll + 1 + nm1] = sinr;
work[i__ - ll + 1 + nm12] = cosl;
work[i__ - ll + 1 + nm13] = sinl;
/* L140: */
}
e[m - 1] = f;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &vt[
ll + vt_dim1], ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
+ 1], &u[ll * u_dim1 + 1], ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
+ 1], &c__[ll + c_dim1], ldc);
}
/* Test convergence */
if ((r__1 = e[m - 1], ABS(r__1)) <= thresh) {
e[m - 1] = 0.f;
}
} else {
/* Chase bulge from bottom to top */
/* Save cosines and sines for later singular vector updates */
f = ((r__1 = d__[m], ABS(r__1)) - shift) * (r_sign(&c_b49, &d__[
m]) + shift / d__[m]);
g = e[m - 1];
i__1 = ll + 1;
for (i__ = m; i__ >= i__1; --i__) {
slartg_(&f, &g, &cosr, &sinr, &r__);
if (i__ < m) {
e[i__] = r__;
}
f = cosr * d__[i__] + sinr * e[i__ - 1];
e[i__ - 1] = cosr * e[i__ - 1] - sinr * d__[i__];
g = sinr * d__[i__ - 1];
d__[i__ - 1] = cosr * d__[i__ - 1];
slartg_(&f, &g, &cosl, &sinl, &r__);
d__[i__] = r__;
f = cosl * e[i__ - 1] + sinl * d__[i__ - 1];
d__[i__ - 1] = cosl * d__[i__ - 1] - sinl * e[i__ - 1];
if (i__ > ll + 1) {
g = sinl * e[i__ - 2];
e[i__ - 2] = cosl * e[i__ - 2];
}
work[i__ - ll] = cosr;
work[i__ - ll + nm1] = -sinr;
work[i__ - ll + nm12] = cosl;
work[i__ - ll + nm13] = -sinl;
/* L150: */
}
e[ll] = f;
/* Test convergence */
if ((r__1 = e[ll], ABS(r__1)) <= thresh) {
e[ll] = 0.f;
}
/* Update singular vectors if desired */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
nm13 + 1], &vt[ll + vt_dim1], ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u[ll *
u_dim1 + 1], ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &c__[
ll + c_dim1], ldc);
}
}
}
/* QR iteration finished, go back and check convergence */
goto L60;
/* All singular values converged, so make them positive */
L160:
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
if (d__[i__] < 0.f) {
d__[i__] = -d__[i__];
/* Change sign of singular vectors, if desired */
if (*ncvt > 0) {
sscal_(ncvt, &c_b72, &vt[i__ + vt_dim1], ldvt);
}
}
/* L170: */
}
/* Sort the singular values into decreasing order (insertion sort on */
/* singular values, but only one transposition per singular vector) */
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Scan for smallest D(I) */
isub = 1;
smin = d__[1];
i__2 = *n + 1 - i__;
for (j = 2; j <= i__2; ++j) {
if (d__[j] <= smin) {
isub = j;
smin = d__[j];
}
/* L180: */
}
if (isub != *n + 1 - i__) {
/* Swap singular values and vectors */
d__[isub] = d__[*n + 1 - i__];
d__[*n + 1 - i__] = smin;
if (*ncvt > 0) {
sswap_(ncvt, &vt[isub + vt_dim1], ldvt, &vt[*n + 1 - i__ +
vt_dim1], ldvt);
}
if (*nru > 0) {
sswap_(nru, &u[isub * u_dim1 + 1], &c__1, &u[(*n + 1 - i__) *
u_dim1 + 1], &c__1);
}
if (*ncc > 0) {
sswap_(ncc, &c__[isub + c_dim1], ldc, &c__[*n + 1 - i__ +
c_dim1], ldc);
}
}
/* L190: */
}
goto L220;
/* Maximum number of iterations exceeded, failure to converge */
L200:
*info = 0;
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
if (e[i__] != 0.f) {
++(*info);
}
/* L210: */
}
L220:
return 0;
/* End of SBDSQR */
} /* sbdsqr_ */
/* Subroutine */ int ssteqr_(char *compz, integer *n, real *d__, real *e,
real *z__, integer *ldz, real *work, integer *info)
{
/* System generated locals */
integer z_dim1, z_offset, i__1, i__2;
real r__1, r__2;
/* Local variables */
real b, c__, f, g;
integer i__, j, k, l, m;
real p, r__, s;
integer l1, ii, mm, lm1, mm1, nm1;
real rt1, rt2, eps;
integer lsv;
real tst, eps2;
integer lend, jtot;
real anorm;
integer lendm1, lendp1;
integer iscale;
real safmin;
real safmax;
integer lendsv;
real ssfmin;
integer nmaxit, icompz;
real ssfmax;
/* -- LAPACK routine (version 3.2) -- */
/* November 2006 */
/* Purpose */
/* ======= */
/* SSTEQR computes all eigenvalues and, optionally, eigenvectors of a */
/* symmetric tridiagonal matrix using the implicit QL or QR method. */
/* The eigenvectors of a full or band symmetric matrix can also be found */
/* if SSYTRD or SSPTRD or SSBTRD has been used to reduce this matrix to */
/* tridiagonal form. */
/* Arguments */
/* ========= */
/* COMPZ (input) CHARACTER*1 */
/* = 'N': Compute eigenvalues only. */
/* = 'V': Compute eigenvalues and eigenvectors of the original */
/* symmetric matrix. On entry, Z must contain the */
/* orthogonal matrix used to reduce the original matrix */
/* to tridiagonal form. */
/* = 'I': Compute eigenvalues and eigenvectors of the */
/* tridiagonal matrix. Z is initialized to the identity */
/* matrix. */
/* N (input) INTEGER */
/* The order of the matrix. N >= 0. */
/* D (input/output) REAL array, dimension (N) */
/* On entry, the diagonal elements of the tridiagonal matrix. */
/* On exit, if INFO = 0, the eigenvalues in ascending order. */
/* E (input/output) REAL array, dimension (N-1) */
/* On entry, the (n-1) subdiagonal elements of the tridiagonal */
/* matrix. */
/* On exit, E has been destroyed. */
/* Z (input/output) REAL array, dimension (LDZ, N) */
/* On entry, if COMPZ = 'V', then Z contains the orthogonal */
/* matrix used in the reduction to tridiagonal form. */
/* On exit, if INFO = 0, then if COMPZ = 'V', Z contains the */
/* orthonormal eigenvectors of the original symmetric matrix, */
/* and if COMPZ = 'I', Z contains the orthonormal eigenvectors */
/* of the symmetric tridiagonal matrix. */
/* If COMPZ = 'N', then Z is not referenced. */
/* LDZ (input) INTEGER */
/* The leading dimension of the array Z. LDZ >= 1, and if */
/* eigenvectors are desired, then LDZ >= max(1,N). */
/* WORK (workspace) REAL array, dimension (max(1,2*N-2)) */
/* If COMPZ = 'N', then WORK is not referenced. */
/* INFO (output) INTEGER */
/* = 0: successful exit */
/* < 0: if INFO = -i, the i-th argument had an illegal value */
/* > 0: the algorithm has failed to find all the eigenvalues in */
/* a total of 30*N iterations; if INFO = i, then i */
/* elements of E have not converged to zero; on exit, D */
/* and E contain the elements of a symmetric tridiagonal */
/* matrix which is orthogonally similar to the original */
/* matrix. */
/* ===================================================================== */
/* Test the input parameters. */
/* Parameter adjustments */
--d__;
--e;
z_dim1 = *ldz;
z_offset = 1 + z_dim1;
z__ -= z_offset;
--work;
/* Function Body */
*info = 0;
if (lsame_(compz, "N")) {
icompz = 0;
} else if (lsame_(compz, "V")) {
icompz = 1;
} else if (lsame_(compz, "I")) {
icompz = 2;
} else {
icompz = -1;
}
if (icompz < 0) {
*info = -1;
} else if (*n < 0) {
*info = -2;
} else if (*ldz < 1 || icompz > 0 && *ldz < max(1,*n)) {
*info = -6;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SSTEQR", &i__1);
return 0;
}
/* Quick return if possible */
if (*n == 0) {
return 0;
}
if (*n == 1) {
if (icompz == 2) {
z__[z_dim1 + 1] = 1.f;
}
return 0;
}
/* Determine the unit roundoff and over/underflow thresholds. */
eps = slamch_("E");
/* Computing 2nd power */
r__1 = eps;
eps2 = r__1 * r__1;
safmin = slamch_("S");
safmax = 1.f / safmin;
ssfmax = sqrt(safmax) / 3.f;
ssfmin = sqrt(safmin) / eps2;
/* Compute the eigenvalues and eigenvectors of the tridiagonal */
/* matrix. */
if (icompz == 2) {
slaset_("Full", n, n, &c_b9, &c_b10, &z__[z_offset], ldz);
}
nmaxit = *n * 30;
jtot = 0;
/* Determine where the matrix splits and choose QL or QR iteration */
/* for each block, according to whether top or bottom diagonal */
/* element is smaller. */
l1 = 1;
nm1 = *n - 1;
L10:
if (l1 > *n) {
goto L160;
}
if (l1 > 1) {
e[l1 - 1] = 0.f;
}
if (l1 <= nm1) {
i__1 = nm1;
for (m = l1; m <= i__1; ++m) {
tst = (r__1 = e[m], dabs(r__1));
if (tst == 0.f) {
goto L30;
}
if (tst <= sqrt((r__1 = d__[m], dabs(r__1))) * sqrt((r__2 = d__[m
+ 1], dabs(r__2))) * eps) {
e[m] = 0.f;
goto L30;
}
}
}
m = *n;
L30:
l = l1;
lsv = l;
lend = m;
lendsv = lend;
l1 = m + 1;
if (lend == l) {
goto L10;
}
/* Scale submatrix in rows and columns L to LEND */
i__1 = lend - l + 1;
anorm = slanst_("I", &i__1, &d__[l], &e[l]);
iscale = 0;
if (anorm == 0.f) {
goto L10;
}
if (anorm > ssfmax) {
iscale = 1;
i__1 = lend - l + 1;
slascl_("G", &c__0, &c__0, &anorm, &ssfmax, &i__1, &c__1, &d__[l], n,
info);
i__1 = lend - l;
slascl_("G", &c__0, &c__0, &anorm, &ssfmax, &i__1, &c__1, &e[l], n,
info);
} else if (anorm < ssfmin) {
iscale = 2;
i__1 = lend - l + 1;
slascl_("G", &c__0, &c__0, &anorm, &ssfmin, &i__1, &c__1, &d__[l], n,
info);
i__1 = lend - l;
slascl_("G", &c__0, &c__0, &anorm, &ssfmin, &i__1, &c__1, &e[l], n,
info);
}
/* Choose between QL and QR iteration */
if ((r__1 = d__[lend], dabs(r__1)) < (r__2 = d__[l], dabs(r__2))) {
lend = lsv;
l = lendsv;
}
if (lend > l) {
/* QL Iteration */
/* Look for small subdiagonal element. */
L40:
if (l != lend) {
lendm1 = lend - 1;
i__1 = lendm1;
for (m = l; m <= i__1; ++m) {
/* Computing 2nd power */
r__2 = (r__1 = e[m], dabs(r__1));
tst = r__2 * r__2;
if (tst <= eps2 * (r__1 = d__[m], dabs(r__1)) * (r__2 = d__[m
+ 1], dabs(r__2)) + safmin) {
goto L60;
}
}
}
m = lend;
L60:
if (m < lend) {
e[m] = 0.f;
}
p = d__[l];
if (m == l) {
goto L80;
}
/* If remaining matrix is 2-by-2, use SLAE2 or SLAEV2 */
/* to compute its eigensystem. */
if (m == l + 1) {
if (icompz > 0) {
slaev2_(&d__[l], &e[l], &d__[l + 1], &rt1, &rt2, &c__, &s);
work[l] = c__;
work[*n - 1 + l] = s;
slasr_("R", "V", "B", n, &c__2, &work[l], &work[*n - 1 + l], &
z__[l * z_dim1 + 1], ldz);
} else {
slae2_(&d__[l], &e[l], &d__[l + 1], &rt1, &rt2);
}
d__[l] = rt1;
d__[l + 1] = rt2;
e[l] = 0.f;
l += 2;
if (l <= lend) {
goto L40;
}
goto L140;
}
if (jtot == nmaxit) {
goto L140;
}
++jtot;
/* Form shift. */
g = (d__[l + 1] - p) / (e[l] * 2.f);
r__ = slapy2_(&g, &c_b10);
g = d__[m] - p + e[l] / (g + r_sign(&r__, &g));
s = 1.f;
c__ = 1.f;
p = 0.f;
/* Inner loop */
mm1 = m - 1;
i__1 = l;
for (i__ = mm1; i__ >= i__1; --i__) {
f = s * e[i__];
b = c__ * e[i__];
slartg_(&g, &f, &c__, &s, &r__);
if (i__ != m - 1) {
e[i__ + 1] = r__;
}
g = d__[i__ + 1] - p;
r__ = (d__[i__] - g) * s + c__ * 2.f * b;
p = s * r__;
d__[i__ + 1] = g + p;
g = c__ * r__ - b;
/* If eigenvectors are desired, then save rotations. */
if (icompz > 0) {
work[i__] = c__;
work[*n - 1 + i__] = -s;
}
}
/* If eigenvectors are desired, then apply saved rotations. */
if (icompz > 0) {
mm = m - l + 1;
slasr_("R", "V", "B", n, &mm, &work[l], &work[*n - 1 + l], &z__[l
* z_dim1 + 1], ldz);
}
d__[l] -= p;
e[l] = g;
goto L40;
/* Eigenvalue found. */
L80:
d__[l] = p;
++l;
if (l <= lend) {
goto L40;
}
goto L140;
} else {
/* QR Iteration */
/* Look for small superdiagonal element. */
L90:
if (l != lend) {
lendp1 = lend + 1;
i__1 = lendp1;
for (m = l; m >= i__1; --m) {
/* Computing 2nd power */
r__2 = (r__1 = e[m - 1], dabs(r__1));
tst = r__2 * r__2;
if (tst <= eps2 * (r__1 = d__[m], dabs(r__1)) * (r__2 = d__[m
- 1], dabs(r__2)) + safmin) {
goto L110;
}
}
}
m = lend;
L110:
if (m > lend) {
e[m - 1] = 0.f;
}
p = d__[l];
if (m == l) {
goto L130;
}
/* If remaining matrix is 2-by-2, use SLAE2 or SLAEV2 */
/* to compute its eigensystem. */
if (m == l - 1) {
if (icompz > 0) {
slaev2_(&d__[l - 1], &e[l - 1], &d__[l], &rt1, &rt2, &c__, &s)
;
work[m] = c__;
work[*n - 1 + m] = s;
slasr_("R", "V", "F", n, &c__2, &work[m], &work[*n - 1 + m], &
z__[(l - 1) * z_dim1 + 1], ldz);
} else {
slae2_(&d__[l - 1], &e[l - 1], &d__[l], &rt1, &rt2);
}
d__[l - 1] = rt1;
d__[l] = rt2;
e[l - 1] = 0.f;
l += -2;
if (l >= lend) {
goto L90;
}
goto L140;
}
if (jtot == nmaxit) {
goto L140;
}
++jtot;
/* Form shift. */
g = (d__[l - 1] - p) / (e[l - 1] * 2.f);
r__ = slapy2_(&g, &c_b10);
g = d__[m] - p + e[l - 1] / (g + r_sign(&r__, &g));
s = 1.f;
c__ = 1.f;
p = 0.f;
/* Inner loop */
lm1 = l - 1;
i__1 = lm1;
for (i__ = m; i__ <= i__1; ++i__) {
f = s * e[i__];
b = c__ * e[i__];
slartg_(&g, &f, &c__, &s, &r__);
if (i__ != m) {
e[i__ - 1] = r__;
}
g = d__[i__] - p;
r__ = (d__[i__ + 1] - g) * s + c__ * 2.f * b;
p = s * r__;
d__[i__] = g + p;
g = c__ * r__ - b;
/* If eigenvectors are desired, then save rotations. */
if (icompz > 0) {
work[i__] = c__;
work[*n - 1 + i__] = s;
}
}
/* If eigenvectors are desired, then apply saved rotations. */
if (icompz > 0) {
mm = l - m + 1;
slasr_("R", "V", "F", n, &mm, &work[m], &work[*n - 1 + m], &z__[m
* z_dim1 + 1], ldz);
}
d__[l] -= p;
e[lm1] = g;
goto L90;
/* Eigenvalue found. */
L130:
d__[l] = p;
--l;
if (l >= lend) {
goto L90;
}
goto L140;
}
/* Undo scaling if necessary */
L140:
if (iscale == 1) {
i__1 = lendsv - lsv + 1;
slascl_("G", &c__0, &c__0, &ssfmax, &anorm, &i__1, &c__1, &d__[lsv],
n, info);
i__1 = lendsv - lsv;
slascl_("G", &c__0, &c__0, &ssfmax, &anorm, &i__1, &c__1, &e[lsv], n,
info);
} else if (iscale == 2) {
i__1 = lendsv - lsv + 1;
slascl_("G", &c__0, &c__0, &ssfmin, &anorm, &i__1, &c__1, &d__[lsv],
n, info);
i__1 = lendsv - lsv;
slascl_("G", &c__0, &c__0, &ssfmin, &anorm, &i__1, &c__1, &e[lsv], n,
info);
}
/* Check for no convergence to an eigenvalue after a total */
/* of N*MAXIT iterations. */
if (jtot < nmaxit) {
goto L10;
}
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
if (e[i__] != 0.f) {
++(*info);
}
}
goto L190;
/* Order eigenvalues and eigenvectors. */
L160:
if (icompz == 0) {
/* Use Quick Sort */
slasrt_("I", n, &d__[1], info);
} else {
/* Use Selection Sort to minimize swaps of eigenvectors */
i__1 = *n;
for (ii = 2; ii <= i__1; ++ii) {
i__ = ii - 1;
k = i__;
p = d__[i__];
i__2 = *n;
for (j = ii; j <= i__2; ++j) {
if (d__[j] < p) {
k = j;
p = d__[j];
}
}
if (k != i__) {
d__[k] = d__[i__];
d__[i__] = p;
sswap_(n, &z__[i__ * z_dim1 + 1], &c__1, &z__[k * z_dim1 + 1],
&c__1);
}
}
}
L190:
return 0;
/* End of SSTEQR */
} /* ssteqr_ */
int slasdq_(char *uplo, int *sqre, int *n, int *
ncvt, int *nru, int *ncc, float *d__, float *e, float *vt,
int *ldvt, float *u, int *ldu, float *c__, int *ldc, float *
work, int *info)
{
/* System generated locals */
int c_dim1, c_offset, u_dim1, u_offset, vt_dim1, vt_offset, i__1,
i__2;
/* Local variables */
int i__, j;
float r__, cs, sn;
int np1, isub;
float smin;
int sqre1;
extern int lsame_(char *, char *);
extern int slasr_(char *, char *, char *, int *,
int *, float *, float *, float *, int *);
int iuplo;
extern int sswap_(int *, float *, int *, float *,
int *), xerbla_(char *, int *), slartg_(float *,
float *, float *, float *, float *);
int rotate;
extern int sbdsqr_(char *, int *, int *, int
*, int *, float *, float *, float *, int *, float *, int *
, float *, int *, float *, int *);
/* -- LAPACK auxiliary routine (version 3.2) -- */
/* Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */
/* November 2006 */
/* .. Scalar Arguments .. */
/* .. */
/* .. Array Arguments .. */
/* .. */
/* Purpose */
/* ======= */
/* SLASDQ computes the singular value decomposition (SVD) of a float */
/* (upper or lower) bidiagonal matrix with diagonal D and offdiagonal */
/* E, accumulating the transformations if desired. Letting B denote */
/* the input bidiagonal matrix, the algorithm computes orthogonal */
/* matrices Q and P such that B = Q * S * P' (P' denotes the transpose */
/* of P). The singular values S are overwritten on D. */
/* The input matrix U is changed to U * Q if desired. */
/* The input matrix VT is changed to P' * VT if desired. */
/* The input matrix C is changed to Q' * C if desired. */
/* See "Computing Small Singular Values of Bidiagonal Matrices With */
/* Guaranteed High Relative Accuracy," by J. Demmel and W. Kahan, */
/* LAPACK Working Note #3, for a detailed description of the algorithm. */
/* Arguments */
/* ========= */
/* UPLO (input) CHARACTER*1 */
/* On entry, UPLO specifies whether the input bidiagonal matrix */
/* is upper or lower bidiagonal, and wether it is square are */
/* not. */
/* UPLO = 'U' or 'u' B is upper bidiagonal. */
/* UPLO = 'L' or 'l' B is lower bidiagonal. */
/* SQRE (input) INTEGER */
/* = 0: then the input matrix is N-by-N. */
/* = 1: then the input matrix is N-by-(N+1) if UPLU = 'U' and */
/* (N+1)-by-N if UPLU = 'L'. */
/* The bidiagonal matrix has */
/* N = NL + NR + 1 rows and */
/* M = N + SQRE >= N columns. */
/* N (input) INTEGER */
/* On entry, N specifies the number of rows and columns */
/* in the matrix. N must be at least 0. */
/* NCVT (input) INTEGER */
/* On entry, NCVT specifies the number of columns of */
/* the matrix VT. NCVT must be at least 0. */
/* NRU (input) INTEGER */
/* On entry, NRU specifies the number of rows of */
/* the matrix U. NRU must be at least 0. */
/* NCC (input) INTEGER */
/* On entry, NCC specifies the number of columns of */
/* the matrix C. NCC must be at least 0. */
/* D (input/output) REAL array, dimension (N) */
/* On entry, D contains the diagonal entries of the */
/* bidiagonal matrix whose SVD is desired. On normal exit, */
/* D contains the singular values in ascending order. */
/* E (input/output) REAL array. */
/* dimension is (N-1) if SQRE = 0 and N if SQRE = 1. */
/* On entry, the entries of E contain the offdiagonal entries */
/* of the bidiagonal matrix whose SVD is desired. On normal */
/* exit, E will contain 0. If the algorithm does not converge, */
/* D and E will contain the diagonal and superdiagonal entries */
/* of a bidiagonal matrix orthogonally equivalent to the one */
/* given as input. */
/* VT (input/output) REAL array, dimension (LDVT, NCVT) */
/* On entry, contains a matrix which on exit has been */
/* premultiplied by P', dimension N-by-NCVT if SQRE = 0 */
/* and (N+1)-by-NCVT if SQRE = 1 (not referenced if NCVT=0). */
/* LDVT (input) INTEGER */
/* On entry, LDVT specifies the leading dimension of VT as */
/* declared in the calling (sub) program. LDVT must be at */
/* least 1. If NCVT is nonzero LDVT must also be at least N. */
/* U (input/output) REAL array, dimension (LDU, N) */
/* On entry, contains a matrix which on exit has been */
/* postmultiplied by Q, dimension NRU-by-N if SQRE = 0 */
/* and NRU-by-(N+1) if SQRE = 1 (not referenced if NRU=0). */
/* LDU (input) INTEGER */
/* On entry, LDU specifies the leading dimension of U as */
/* declared in the calling (sub) program. LDU must be at */
/* least MAX( 1, NRU ) . */
/* C (input/output) REAL array, dimension (LDC, NCC) */
/* On entry, contains an N-by-NCC matrix which on exit */
/* has been premultiplied by Q' dimension N-by-NCC if SQRE = 0 */
/* and (N+1)-by-NCC if SQRE = 1 (not referenced if NCC=0). */
/* LDC (input) INTEGER */
/* On entry, LDC specifies the leading dimension of C as */
/* declared in the calling (sub) program. LDC must be at */
/* least 1. If NCC is nonzero, LDC must also be at least N. */
/* WORK (workspace) REAL array, dimension (4*N) */
/* Workspace. Only referenced if one of NCVT, NRU, or NCC is */
/* nonzero, and if N is at least 2. */
/* INFO (output) INTEGER */
/* On exit, a value of 0 indicates a successful exit. */
/* If INFO < 0, argument number -INFO is illegal. */
/* If INFO > 0, the algorithm did not converge, and INFO */
/* specifies how many superdiagonals did not converge. */
/* Further Details */
/* =============== */
/* Based on contributions by */
/* Ming Gu and Huan Ren, Computer Science Division, University of */
/* California at Berkeley, USA */
/* ===================================================================== */
/* .. Parameters .. */
/* .. */
/* .. Local Scalars .. */
/* .. */
/* .. External Subroutines .. */
/* .. */
/* .. External Functions .. */
/* .. */
/* .. Intrinsic Functions .. */
/* .. */
/* .. Executable Statements .. */
/* Test the input parameters. */
/* Parameter adjustments */
--d__;
--e;
vt_dim1 = *ldvt;
vt_offset = 1 + vt_dim1;
vt -= vt_offset;
u_dim1 = *ldu;
u_offset = 1 + u_dim1;
u -= u_offset;
c_dim1 = *ldc;
c_offset = 1 + c_dim1;
c__ -= c_offset;
--work;
/* Function Body */
*info = 0;
iuplo = 0;
if (lsame_(uplo, "U")) {
iuplo = 1;
}
if (lsame_(uplo, "L")) {
iuplo = 2;
}
if (iuplo == 0) {
*info = -1;
} else if (*sqre < 0 || *sqre > 1) {
*info = -2;
} else if (*n < 0) {
*info = -3;
} else if (*ncvt < 0) {
*info = -4;
} else if (*nru < 0) {
*info = -5;
} else if (*ncc < 0) {
*info = -6;
} else if (*ncvt == 0 && *ldvt < 1 || *ncvt > 0 && *ldvt < MAX(1,*n)) {
*info = -10;
} else if (*ldu < MAX(1,*nru)) {
*info = -12;
} else if (*ncc == 0 && *ldc < 1 || *ncc > 0 && *ldc < MAX(1,*n)) {
*info = -14;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SLASDQ", &i__1);
return 0;
}
if (*n == 0) {
return 0;
}
/* ROTATE is true if any singular vectors desired, false otherwise */
rotate = *ncvt > 0 || *nru > 0 || *ncc > 0;
np1 = *n + 1;
sqre1 = *sqre;
/* If matrix non-square upper bidiagonal, rotate to be lower */
/* bidiagonal. The rotations are on the right. */
if (iuplo == 1 && sqre1 == 1) {
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
if (rotate) {
work[i__] = cs;
work[*n + i__] = sn;
}
/* L10: */
}
slartg_(&d__[*n], &e[*n], &cs, &sn, &r__);
d__[*n] = r__;
e[*n] = 0.f;
if (rotate) {
work[*n] = cs;
work[*n + *n] = sn;
}
iuplo = 2;
sqre1 = 0;
/* Update singular vectors if desired. */
if (*ncvt > 0) {
slasr_("L", "V", "F", &np1, ncvt, &work[1], &work[np1], &vt[
vt_offset], ldvt);
}
}
/* If matrix lower bidiagonal, rotate to be upper bidiagonal */
/* by applying Givens rotations on the left. */
if (iuplo == 2) {
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
if (rotate) {
work[i__] = cs;
work[*n + i__] = sn;
}
/* L20: */
}
/* If matrix (N+1)-by-N lower bidiagonal, one additional */
/* rotation is needed. */
if (sqre1 == 1) {
slartg_(&d__[*n], &e[*n], &cs, &sn, &r__);
d__[*n] = r__;
if (rotate) {
work[*n] = cs;
work[*n + *n] = sn;
}
}
/* Update singular vectors if desired. */
if (*nru > 0) {
if (sqre1 == 0) {
slasr_("R", "V", "F", nru, n, &work[1], &work[np1], &u[
u_offset], ldu);
} else {
slasr_("R", "V", "F", nru, &np1, &work[1], &work[np1], &u[
u_offset], ldu);
}
}
if (*ncc > 0) {
if (sqre1 == 0) {
slasr_("L", "V", "F", n, ncc, &work[1], &work[np1], &c__[
c_offset], ldc);
} else {
slasr_("L", "V", "F", &np1, ncc, &work[1], &work[np1], &c__[
c_offset], ldc);
}
}
}
/* Call SBDSQR to compute the SVD of the reduced float */
/* N-by-N upper bidiagonal matrix. */
sbdsqr_("U", n, ncvt, nru, ncc, &d__[1], &e[1], &vt[vt_offset], ldvt, &u[
u_offset], ldu, &c__[c_offset], ldc, &work[1], info);
/* Sort the singular values into ascending order (insertion sort on */
/* singular values, but only one transposition per singular vector) */
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Scan for smallest D(I). */
isub = i__;
smin = d__[i__];
i__2 = *n;
for (j = i__ + 1; j <= i__2; ++j) {
if (d__[j] < smin) {
isub = j;
smin = d__[j];
}
/* L30: */
}
if (isub != i__) {
/* Swap singular values and vectors. */
d__[isub] = d__[i__];
d__[i__] = smin;
if (*ncvt > 0) {
sswap_(ncvt, &vt[isub + vt_dim1], ldvt, &vt[i__ + vt_dim1],
ldvt);
}
if (*nru > 0) {
sswap_(nru, &u[isub * u_dim1 + 1], &c__1, &u[i__ * u_dim1 + 1]
, &c__1);
}
if (*ncc > 0) {
sswap_(ncc, &c__[isub + c_dim1], ldc, &c__[i__ + c_dim1], ldc)
;
}
}
/* L40: */
}
return 0;
/* End of SLASDQ */
} /* slasdq_ */
/* Subroutine */ int sbdsqr_(char *uplo, integer *n, integer *ncvt, integer *
nru, integer *ncc, real *d__, real *e, real *vt, integer *ldvt, real *
u, integer *ldu, real *c__, integer *ldc, real *work, integer *info)
{
/* System generated locals */
integer c_dim1, c_offset, u_dim1, u_offset, vt_dim1, vt_offset, i__1,
i__2;
real r__1, r__2, r__3, r__4;
doublereal d__1;
/* Builtin functions */
double pow_dd(doublereal *, doublereal *), sqrt(doublereal), r_sign(real *
, real *);
/* Local variables */
static real abse;
static integer idir;
static real abss;
static integer oldm;
static real cosl;
static integer isub, iter;
static real unfl, sinl, cosr, smin, smax, sinr;
extern /* Subroutine */ int srot_(integer *, real *, integer *, real *,
integer *, real *, real *), slas2_(real *, real *, real *, real *,
real *);
static real f, g, h__;
static integer i__, j, m;
static real r__;
extern logical lsame_(char *, char *);
static real oldcs;
extern /* Subroutine */ int sscal_(integer *, real *, real *, integer *);
static integer oldll;
static real shift, sigmn, oldsn;
static integer maxit;
static real sminl;
extern /* Subroutine */ int slasr_(char *, char *, char *, integer *,
integer *, real *, real *, real *, integer *);
static real sigmx;
static logical lower;
extern /* Subroutine */ int sswap_(integer *, real *, integer *, real *,
integer *), slasq1_(integer *, real *, real *, real *, integer *),
slasv2_(real *, real *, real *, real *, real *, real *, real *,
real *, real *);
static real cs;
static integer ll;
static real sn, mu;
extern doublereal slamch_(char *);
extern /* Subroutine */ int xerbla_(char *, integer *);
static real sminoa;
extern /* Subroutine */ int slartg_(real *, real *, real *, real *, real *
);
static real thresh;
static logical rotate;
static real sminlo;
static integer nm1;
static real tolmul;
static integer nm12, nm13, lll;
static real eps, sll, tol;
#define c___ref(a_1,a_2) c__[(a_2)*c_dim1 + a_1]
#define u_ref(a_1,a_2) u[(a_2)*u_dim1 + a_1]
#define vt_ref(a_1,a_2) vt[(a_2)*vt_dim1 + a_1]
/* -- LAPACK routine (version 3.0) --
Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd.,
Courant Institute, Argonne National Lab, and Rice University
October 31, 1999
Purpose
=======
SBDSQR computes the singular value decomposition (SVD) of a real
N-by-N (upper or lower) bidiagonal matrix B: B = Q * S * P' (P'
denotes the transpose of P), where S is a diagonal matrix with
non-negative diagonal elements (the singular values of B), and Q
and P are orthogonal matrices.
The routine computes S, and optionally computes U * Q, P' * VT,
or Q' * C, for given real input matrices U, VT, and C.
See "Computing Small Singular Values of Bidiagonal Matrices With
Guaranteed High Relative Accuracy," by J. Demmel and W. Kahan,
LAPACK Working Note #3 (or SIAM J. Sci. Statist. Comput. vol. 11,
no. 5, pp. 873-912, Sept 1990) and
"Accurate singular values and differential qd algorithms," by
B. Parlett and V. Fernando, Technical Report CPAM-554, Mathematics
Department, University of California at Berkeley, July 1992
for a detailed description of the algorithm.
Arguments
=========
UPLO (input) CHARACTER*1
= 'U': B is upper bidiagonal;
= 'L': B is lower bidiagonal.
N (input) INTEGER
The order of the matrix B. N >= 0.
NCVT (input) INTEGER
The number of columns of the matrix VT. NCVT >= 0.
NRU (input) INTEGER
The number of rows of the matrix U. NRU >= 0.
NCC (input) INTEGER
The number of columns of the matrix C. NCC >= 0.
D (input/output) REAL array, dimension (N)
On entry, the n diagonal elements of the bidiagonal matrix B.
On exit, if INFO=0, the singular values of B in decreasing
order.
E (input/output) REAL array, dimension (N)
On entry, the elements of E contain the
offdiagonal elements of the bidiagonal matrix whose SVD
is desired. On normal exit (INFO = 0), E is destroyed.
If the algorithm does not converge (INFO > 0), D and E
will contain the diagonal and superdiagonal elements of a
bidiagonal matrix orthogonally equivalent to the one given
as input. E(N) is used for workspace.
VT (input/output) REAL array, dimension (LDVT, NCVT)
On entry, an N-by-NCVT matrix VT.
On exit, VT is overwritten by P' * VT.
VT is not referenced if NCVT = 0.
LDVT (input) INTEGER
The leading dimension of the array VT.
LDVT >= max(1,N) if NCVT > 0; LDVT >= 1 if NCVT = 0.
U (input/output) REAL array, dimension (LDU, N)
On entry, an NRU-by-N matrix U.
On exit, U is overwritten by U * Q.
U is not referenced if NRU = 0.
LDU (input) INTEGER
The leading dimension of the array U. LDU >= max(1,NRU).
C (input/output) REAL array, dimension (LDC, NCC)
On entry, an N-by-NCC matrix C.
On exit, C is overwritten by Q' * C.
C is not referenced if NCC = 0.
LDC (input) INTEGER
The leading dimension of the array C.
LDC >= max(1,N) if NCC > 0; LDC >=1 if NCC = 0.
WORK (workspace) REAL array, dimension (4*N)
INFO (output) INTEGER
= 0: successful exit
< 0: If INFO = -i, the i-th argument had an illegal value
> 0: the algorithm did not converge; D and E contain the
elements of a bidiagonal matrix which is orthogonally
similar to the input matrix B; if INFO = i, i
elements of E have not converged to zero.
Internal Parameters
===================
TOLMUL REAL, default = max(10,min(100,EPS**(-1/8)))
TOLMUL controls the convergence criterion of the QR loop.
If it is positive, TOLMUL*EPS is the desired relative
precision in the computed singular values.
If it is negative, abs(TOLMUL*EPS*sigma_max) is the
desired absolute accuracy in the computed singular
values (corresponds to relative accuracy
abs(TOLMUL*EPS) in the largest singular value.
abs(TOLMUL) should be between 1 and 1/EPS, and preferably
between 10 (for fast convergence) and .1/EPS
(for there to be some accuracy in the results).
Default is to lose at either one eighth or 2 of the
available decimal digits in each computed singular value
(whichever is smaller).
MAXITR INTEGER, default = 6
MAXITR controls the maximum number of passes of the
algorithm through its inner loop. The algorithms stops
(and so fails to converge) if the number of passes
through the inner loop exceeds MAXITR*N**2.
=====================================================================
Test the input parameters.
Parameter adjustments */
--d__;
--e;
vt_dim1 = *ldvt;
vt_offset = 1 + vt_dim1 * 1;
vt -= vt_offset;
u_dim1 = *ldu;
u_offset = 1 + u_dim1 * 1;
u -= u_offset;
c_dim1 = *ldc;
c_offset = 1 + c_dim1 * 1;
c__ -= c_offset;
--work;
/* Function Body */
*info = 0;
lower = lsame_(uplo, "L");
if (! lsame_(uplo, "U") && ! lower) {
*info = -1;
} else if (*n < 0) {
*info = -2;
} else if (*ncvt < 0) {
*info = -3;
} else if (*nru < 0) {
*info = -4;
} else if (*ncc < 0) {
*info = -5;
} else if (*ncvt == 0 && *ldvt < 1 || *ncvt > 0 && *ldvt < max(1,*n)) {
*info = -9;
} else if (*ldu < max(1,*nru)) {
*info = -11;
} else if (*ncc == 0 && *ldc < 1 || *ncc > 0 && *ldc < max(1,*n)) {
*info = -13;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SBDSQR", &i__1);
return 0;
}
if (*n == 0) {
return 0;
}
if (*n == 1) {
goto L160;
}
/* ROTATE is true if any singular vectors desired, false otherwise */
rotate = *ncvt > 0 || *nru > 0 || *ncc > 0;
/* If no singular vectors desired, use qd algorithm */
if (! rotate) {
slasq1_(n, &d__[1], &e[1], &work[1], info);
return 0;
}
nm1 = *n - 1;
nm12 = nm1 + nm1;
nm13 = nm12 + nm1;
idir = 0;
/* Get machine constants */
eps = slamch_("Epsilon");
unfl = slamch_("Safe minimum");
/* If matrix lower bidiagonal, rotate to be upper bidiagonal
by applying Givens rotations on the left */
if (lower) {
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
work[i__] = cs;
work[nm1 + i__] = sn;
/* L10: */
}
/* Update singular vectors if desired */
if (*nru > 0) {
slasr_("R", "V", "F", nru, n, &work[1], &work[*n], &u[u_offset],
ldu);
}
if (*ncc > 0) {
slasr_("L", "V", "F", n, ncc, &work[1], &work[*n], &c__[c_offset],
ldc);
}
}
/* Compute singular values to relative accuracy TOL
(By setting TOL to be negative, algorithm will compute
singular values to absolute accuracy ABS(TOL)*norm(input matrix))
Computing MAX
Computing MIN */
d__1 = (doublereal) eps;
r__3 = 100.f, r__4 = pow_dd(&d__1, &c_b15);
r__1 = 10.f, r__2 = dmin(r__3,r__4);
tolmul = dmax(r__1,r__2);
tol = tolmul * eps;
/* Compute approximate maximum, minimum singular values */
smax = 0.f;
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Computing MAX */
r__2 = smax, r__3 = (r__1 = d__[i__], dabs(r__1));
smax = dmax(r__2,r__3);
/* L20: */
}
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Computing MAX */
r__2 = smax, r__3 = (r__1 = e[i__], dabs(r__1));
smax = dmax(r__2,r__3);
/* L30: */
}
sminl = 0.f;
if (tol >= 0.f) {
/* Relative accuracy desired */
sminoa = dabs(d__[1]);
if (sminoa == 0.f) {
goto L50;
}
mu = sminoa;
i__1 = *n;
for (i__ = 2; i__ <= i__1; ++i__) {
mu = (r__2 = d__[i__], dabs(r__2)) * (mu / (mu + (r__1 = e[i__ -
1], dabs(r__1))));
sminoa = dmin(sminoa,mu);
if (sminoa == 0.f) {
goto L50;
}
/* L40: */
}
L50:
sminoa /= sqrt((real) (*n));
/* Computing MAX */
r__1 = tol * sminoa, r__2 = *n * 6 * *n * unfl;
thresh = dmax(r__1,r__2);
} else {
/* Absolute accuracy desired
Computing MAX */
r__1 = dabs(tol) * smax, r__2 = *n * 6 * *n * unfl;
thresh = dmax(r__1,r__2);
}
/* Prepare for main iteration loop for the singular values
(MAXIT is the maximum number of passes through the inner
loop permitted before nonconvergence signalled.) */
maxit = *n * 6 * *n;
iter = 0;
oldll = -1;
oldm = -1;
/* M points to last element of unconverged part of matrix */
m = *n;
/* Begin main iteration loop */
L60:
/* Check for convergence or exceeding iteration count */
if (m <= 1) {
goto L160;
}
if (iter > maxit) {
goto L200;
}
/* Find diagonal block of matrix to work on */
if (tol < 0.f && (r__1 = d__[m], dabs(r__1)) <= thresh) {
d__[m] = 0.f;
}
smax = (r__1 = d__[m], dabs(r__1));
smin = smax;
i__1 = m - 1;
for (lll = 1; lll <= i__1; ++lll) {
ll = m - lll;
abss = (r__1 = d__[ll], dabs(r__1));
abse = (r__1 = e[ll], dabs(r__1));
if (tol < 0.f && abss <= thresh) {
d__[ll] = 0.f;
}
if (abse <= thresh) {
goto L80;
}
smin = dmin(smin,abss);
/* Computing MAX */
r__1 = max(smax,abss);
smax = dmax(r__1,abse);
/* L70: */
}
ll = 0;
goto L90;
L80:
e[ll] = 0.f;
/* Matrix splits since E(LL) = 0 */
if (ll == m - 1) {
/* Convergence of bottom singular value, return to top of loop */
--m;
goto L60;
}
L90:
++ll;
/* E(LL) through E(M-1) are nonzero, E(LL-1) is zero */
if (ll == m - 1) {
/* 2 by 2 block, handle separately */
slasv2_(&d__[m - 1], &e[m - 1], &d__[m], &sigmn, &sigmx, &sinr, &cosr,
&sinl, &cosl);
d__[m - 1] = sigmx;
e[m - 1] = 0.f;
d__[m] = sigmn;
/* Compute singular vectors, if desired */
if (*ncvt > 0) {
srot_(ncvt, &vt_ref(m - 1, 1), ldvt, &vt_ref(m, 1), ldvt, &cosr, &
sinr);
}
if (*nru > 0) {
srot_(nru, &u_ref(1, m - 1), &c__1, &u_ref(1, m), &c__1, &cosl, &
sinl);
}
if (*ncc > 0) {
srot_(ncc, &c___ref(m - 1, 1), ldc, &c___ref(m, 1), ldc, &cosl, &
sinl);
}
m += -2;
goto L60;
}
/* If working on new submatrix, choose shift direction
(from larger end diagonal element towards smaller) */
if (ll > oldm || m < oldll) {
if ((r__1 = d__[ll], dabs(r__1)) >= (r__2 = d__[m], dabs(r__2))) {
/* Chase bulge from top (big end) to bottom (small end) */
idir = 1;
} else {
/* Chase bulge from bottom (big end) to top (small end) */
idir = 2;
}
}
/* Apply convergence tests */
if (idir == 1) {
/* Run convergence test in forward direction
First apply standard test to bottom of matrix */
if ((r__2 = e[m - 1], dabs(r__2)) <= dabs(tol) * (r__1 = d__[m], dabs(
r__1)) || tol < 0.f && (r__3 = e[m - 1], dabs(r__3)) <=
thresh) {
e[m - 1] = 0.f;
goto L60;
}
if (tol >= 0.f) {
/* If relative accuracy desired,
apply convergence criterion forward */
mu = (r__1 = d__[ll], dabs(r__1));
sminl = mu;
i__1 = m - 1;
for (lll = ll; lll <= i__1; ++lll) {
if ((r__1 = e[lll], dabs(r__1)) <= tol * mu) {
e[lll] = 0.f;
goto L60;
}
sminlo = sminl;
mu = (r__2 = d__[lll + 1], dabs(r__2)) * (mu / (mu + (r__1 =
e[lll], dabs(r__1))));
sminl = dmin(sminl,mu);
/* L100: */
}
}
} else {
/* Run convergence test in backward direction
First apply standard test to top of matrix */
if ((r__2 = e[ll], dabs(r__2)) <= dabs(tol) * (r__1 = d__[ll], dabs(
r__1)) || tol < 0.f && (r__3 = e[ll], dabs(r__3)) <= thresh) {
e[ll] = 0.f;
goto L60;
}
if (tol >= 0.f) {
/* If relative accuracy desired,
apply convergence criterion backward */
mu = (r__1 = d__[m], dabs(r__1));
sminl = mu;
i__1 = ll;
for (lll = m - 1; lll >= i__1; --lll) {
if ((r__1 = e[lll], dabs(r__1)) <= tol * mu) {
e[lll] = 0.f;
goto L60;
}
sminlo = sminl;
mu = (r__2 = d__[lll], dabs(r__2)) * (mu / (mu + (r__1 = e[
lll], dabs(r__1))));
sminl = dmin(sminl,mu);
/* L110: */
}
}
}
oldll = ll;
oldm = m;
/* Compute shift. First, test if shifting would ruin relative
accuracy, and if so set the shift to zero.
Computing MAX */
r__1 = eps, r__2 = tol * .01f;
if (tol >= 0.f && *n * tol * (sminl / smax) <= dmax(r__1,r__2)) {
/* Use a zero shift to avoid loss of relative accuracy */
shift = 0.f;
} else {
/* Compute the shift from 2-by-2 block at end of matrix */
if (idir == 1) {
sll = (r__1 = d__[ll], dabs(r__1));
slas2_(&d__[m - 1], &e[m - 1], &d__[m], &shift, &r__);
} else {
sll = (r__1 = d__[m], dabs(r__1));
slas2_(&d__[ll], &e[ll], &d__[ll + 1], &shift, &r__);
}
/* Test if shift negligible, and if so set to zero */
if (sll > 0.f) {
/* Computing 2nd power */
r__1 = shift / sll;
if (r__1 * r__1 < eps) {
shift = 0.f;
}
}
}
/* Increment iteration count */
iter = iter + m - ll;
/* If SHIFT = 0, do simplified QR iteration */
if (shift == 0.f) {
if (idir == 1) {
/* Chase bulge from top to bottom
Save cosines and sines for later singular vector updates */
cs = 1.f;
oldcs = 1.f;
i__1 = m - 1;
for (i__ = ll; i__ <= i__1; ++i__) {
r__1 = d__[i__] * cs;
slartg_(&r__1, &e[i__], &cs, &sn, &r__);
if (i__ > ll) {
e[i__ - 1] = oldsn * r__;
}
r__1 = oldcs * r__;
r__2 = d__[i__ + 1] * sn;
slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
work[i__ - ll + 1] = cs;
work[i__ - ll + 1 + nm1] = sn;
work[i__ - ll + 1 + nm12] = oldcs;
work[i__ - ll + 1 + nm13] = oldsn;
/* L120: */
}
h__ = d__[m] * cs;
d__[m] = h__ * oldcs;
e[m - 1] = h__ * oldsn;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &
vt_ref(ll, 1), ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
+ 1], &u_ref(1, ll), ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
+ 1], &c___ref(ll, 1), ldc);
}
/* Test convergence */
if ((r__1 = e[m - 1], dabs(r__1)) <= thresh) {
e[m - 1] = 0.f;
}
} else {
/* Chase bulge from bottom to top
Save cosines and sines for later singular vector updates */
cs = 1.f;
oldcs = 1.f;
i__1 = ll + 1;
for (i__ = m; i__ >= i__1; --i__) {
r__1 = d__[i__] * cs;
slartg_(&r__1, &e[i__ - 1], &cs, &sn, &r__);
if (i__ < m) {
e[i__] = oldsn * r__;
}
r__1 = oldcs * r__;
r__2 = d__[i__ - 1] * sn;
slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
work[i__ - ll] = cs;
work[i__ - ll + nm1] = -sn;
work[i__ - ll + nm12] = oldcs;
work[i__ - ll + nm13] = -oldsn;
/* L130: */
}
h__ = d__[ll] * cs;
d__[ll] = h__ * oldcs;
e[ll] = h__ * oldsn;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
nm13 + 1], &vt_ref(ll, 1), ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u_ref(
1, ll), ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &
c___ref(ll, 1), ldc);
}
/* Test convergence */
if ((r__1 = e[ll], dabs(r__1)) <= thresh) {
e[ll] = 0.f;
}
}
} else {
/* Use nonzero shift */
if (idir == 1) {
/* Chase bulge from top to bottom
Save cosines and sines for later singular vector updates */
f = ((r__1 = d__[ll], dabs(r__1)) - shift) * (r_sign(&c_b49, &d__[
ll]) + shift / d__[ll]);
g = e[ll];
i__1 = m - 1;
for (i__ = ll; i__ <= i__1; ++i__) {
slartg_(&f, &g, &cosr, &sinr, &r__);
if (i__ > ll) {
e[i__ - 1] = r__;
}
f = cosr * d__[i__] + sinr * e[i__];
e[i__] = cosr * e[i__] - sinr * d__[i__];
g = sinr * d__[i__ + 1];
d__[i__ + 1] = cosr * d__[i__ + 1];
slartg_(&f, &g, &cosl, &sinl, &r__);
d__[i__] = r__;
f = cosl * e[i__] + sinl * d__[i__ + 1];
d__[i__ + 1] = cosl * d__[i__ + 1] - sinl * e[i__];
if (i__ < m - 1) {
g = sinl * e[i__ + 1];
e[i__ + 1] = cosl * e[i__ + 1];
}
work[i__ - ll + 1] = cosr;
work[i__ - ll + 1 + nm1] = sinr;
work[i__ - ll + 1 + nm12] = cosl;
work[i__ - ll + 1 + nm13] = sinl;
/* L140: */
}
e[m - 1] = f;
/* Update singular vectors */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &
vt_ref(ll, 1), ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
+ 1], &u_ref(1, ll), ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
+ 1], &c___ref(ll, 1), ldc);
}
/* Test convergence */
if ((r__1 = e[m - 1], dabs(r__1)) <= thresh) {
e[m - 1] = 0.f;
}
} else {
/* Chase bulge from bottom to top
Save cosines and sines for later singular vector updates */
f = ((r__1 = d__[m], dabs(r__1)) - shift) * (r_sign(&c_b49, &d__[
m]) + shift / d__[m]);
g = e[m - 1];
i__1 = ll + 1;
for (i__ = m; i__ >= i__1; --i__) {
slartg_(&f, &g, &cosr, &sinr, &r__);
if (i__ < m) {
e[i__] = r__;
}
f = cosr * d__[i__] + sinr * e[i__ - 1];
e[i__ - 1] = cosr * e[i__ - 1] - sinr * d__[i__];
g = sinr * d__[i__ - 1];
d__[i__ - 1] = cosr * d__[i__ - 1];
slartg_(&f, &g, &cosl, &sinl, &r__);
d__[i__] = r__;
f = cosl * e[i__ - 1] + sinl * d__[i__ - 1];
d__[i__ - 1] = cosl * d__[i__ - 1] - sinl * e[i__ - 1];
if (i__ > ll + 1) {
g = sinl * e[i__ - 2];
e[i__ - 2] = cosl * e[i__ - 2];
}
work[i__ - ll] = cosr;
work[i__ - ll + nm1] = -sinr;
work[i__ - ll + nm12] = cosl;
work[i__ - ll + nm13] = -sinl;
/* L150: */
}
e[ll] = f;
/* Test convergence */
if ((r__1 = e[ll], dabs(r__1)) <= thresh) {
e[ll] = 0.f;
}
/* Update singular vectors if desired */
if (*ncvt > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
nm13 + 1], &vt_ref(ll, 1), ldvt);
}
if (*nru > 0) {
i__1 = m - ll + 1;
slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u_ref(
1, ll), ldu);
}
if (*ncc > 0) {
i__1 = m - ll + 1;
slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &
c___ref(ll, 1), ldc);
}
}
}
/* QR iteration finished, go back and check convergence */
goto L60;
/* All singular values converged, so make them positive */
L160:
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
if (d__[i__] < 0.f) {
d__[i__] = -d__[i__];
/* Change sign of singular vectors, if desired */
if (*ncvt > 0) {
sscal_(ncvt, &c_b72, &vt_ref(i__, 1), ldvt);
}
}
/* L170: */
}
/* Sort the singular values into decreasing order (insertion sort on
singular values, but only one transposition per singular vector) */
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
/* Scan for smallest D(I) */
isub = 1;
smin = d__[1];
i__2 = *n + 1 - i__;
for (j = 2; j <= i__2; ++j) {
if (d__[j] <= smin) {
isub = j;
smin = d__[j];
}
/* L180: */
}
if (isub != *n + 1 - i__) {
/* Swap singular values and vectors */
d__[isub] = d__[*n + 1 - i__];
d__[*n + 1 - i__] = smin;
if (*ncvt > 0) {
sswap_(ncvt, &vt_ref(isub, 1), ldvt, &vt_ref(*n + 1 - i__, 1),
ldvt);
}
if (*nru > 0) {
sswap_(nru, &u_ref(1, isub), &c__1, &u_ref(1, *n + 1 - i__), &
c__1);
}
if (*ncc > 0) {
sswap_(ncc, &c___ref(isub, 1), ldc, &c___ref(*n + 1 - i__, 1),
ldc);
}
}
/* L190: */
}
goto L220;
/* Maximum number of iterations exceeded, failure to converge */
L200:
*info = 0;
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
if (e[i__] != 0.f) {
++(*info);
}
/* L210: */
}
L220:
return 0;
/* End of SBDSQR */
} /* sbdsqr_ */
/* Subroutine */ int ssteqr_(char *compz, int *n, real *d, real *e, real *
z, int *ldz, real *work, int *info)
{
/* -- LAPACK routine (version 2.0) --
Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd.,
Courant Institute, Argonne National Lab, and Rice University
September 30, 1994
Purpose
=======
SSTEQR computes all eigenvalues and, optionally, eigenvectors of a
symmetric tridiagonal matrix using the implicit QL or QR method.
The eigenvectors of a full or band symmetric matrix can also be found
if SSYTRD or SSPTRD or SSBTRD has been used to reduce this matrix to
tridiagonal form.
Arguments
=========
COMPZ (input) CHARACTER*1
= 'N': Compute eigenvalues only.
= 'V': Compute eigenvalues and eigenvectors of the original
symmetric matrix. On entry, Z must contain the
orthogonal matrix used to reduce the original matrix
to tridiagonal form.
= 'I': Compute eigenvalues and eigenvectors of the
tridiagonal matrix. Z is initialized to the identity
matrix.
N (input) INTEGER
The order of the matrix. N >= 0.
D (input/output) REAL array, dimension (N)
On entry, the diagonal elements of the tridiagonal matrix.
On exit, if INFO = 0, the eigenvalues in ascending order.
E (input/output) REAL array, dimension (N-1)
On entry, the (n-1) subdiagonal elements of the tridiagonal
matrix.
On exit, E has been destroyed.
Z (input/output) REAL array, dimension (LDZ, N)
On entry, if COMPZ = 'V', then Z contains the orthogonal
matrix used in the reduction to tridiagonal form.
On exit, if INFO = 0, then if COMPZ = 'V', Z contains the
orthonormal eigenvectors of the original symmetric matrix,
and if COMPZ = 'I', Z contains the orthonormal eigenvectors
of the symmetric tridiagonal matrix.
If COMPZ = 'N', then Z is not referenced.
LDZ (input) INTEGER
The leading dimension of the array Z. LDZ >= 1, and if
eigenvectors are desired, then LDZ >= max(1,N).
WORK (workspace) REAL array, dimension (max(1,2*N-2))
If COMPZ = 'N', then WORK is not referenced.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: the algorithm has failed to find all the eigenvalues in
a total of 30*N iterations; if INFO = i, then i
elements of E have not converged to zero; on exit, D
and E contain the elements of a symmetric tridiagonal
matrix which is orthogonally similar to the original
matrix.
=====================================================================
Test the input parameters.
Parameter adjustments
Function Body */
/* Table of constant values */
static real c_b9 = 0.f;
static real c_b10 = 1.f;
static int c__0 = 0;
static int c__1 = 1;
static int c__2 = 2;
/* System generated locals */
/* Unused variables commented out by MDG on 03-09-05
int z_dim1, z_offset;
*/
int i__1, i__2;
real r__1, r__2;
/* Builtin functions */
double sqrt(doublereal), r_sign(real *, real *);
/* Local variables */
static int lend, jtot;
extern /* Subroutine */ int slae2_(real *, real *, real *, real *, real *)
;
static real b, c, f, g;
static int i, j, k, l, m;
static real p, r, s;
extern logical lsame_(char *, char *);
static real anorm;
extern /* Subroutine */ int slasr_(char *, char *, char *, int *,
int *, real *, real *, real *, int *);
static int l1;
extern /* Subroutine */ int sswap_(int *, real *, int *, real *,
int *);
static int lendm1, lendp1;
extern /* Subroutine */ int slaev2_(real *, real *, real *, real *, real *
, real *, real *);
extern doublereal slapy2_(real *, real *);
static int ii, mm, iscale;
extern doublereal slamch_(char *);
static real safmin;
extern /* Subroutine */ int xerbla_(char *, int *);
static real safmax;
extern /* Subroutine */ int slascl_(char *, int *, int *, real *,
real *, int *, int *, real *, int *, int *);
static int lendsv;
extern /* Subroutine */ int slartg_(real *, real *, real *, real *, real *
), slaset_(char *, int *, int *, real *, real *, real *,
int *);
static real ssfmin;
static int nmaxit, icompz;
static real ssfmax;
extern doublereal slanst_(char *, int *, real *, real *);
extern /* Subroutine */ int slasrt_(char *, int *, real *, int *);
static int lm1, mm1, nm1;
static real rt1, rt2, eps;
static int lsv;
static real tst, eps2;
#define D(I) d[(I)-1]
#define E(I) e[(I)-1]
#define WORK(I) work[(I)-1]
#define Z(I,J) z[(I)-1 + ((J)-1)* ( *ldz)]
*info = 0;
if (lsame_(compz, "N")) {
icompz = 0;
} else if (lsame_(compz, "V")) {
icompz = 1;
} else if (lsame_(compz, "I")) {
icompz = 2;
} else {
icompz = -1;
}
if (icompz < 0) {
*info = -1;
} else if (*n < 0) {
*info = -2;
/*
} else if (*ldz < 1 || icompz > 0 && *ldz < max(1,*n)) {
*/
/* Paretheses added by MDG on 03-09-05 */
} else if ((*ldz < 1 || icompz > 0) && (*ldz < max(1,*n))) {
*info = -6;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SSTEQR", &i__1);
return 0;
}
/* Quick return if possible */
if (*n == 0) {
return 0;
}
if (*n == 1) {
if (icompz == 2) {
Z(1,1) = 1.f;
}
return 0;
}
/* Determine the unit roundoff and over/underflow thresholds. */
eps = slamch_("E");
/* Computing 2nd power */
r__1 = eps;
eps2 = r__1 * r__1;
safmin = slamch_("S");
safmax = 1.f / safmin;
ssfmax = sqrt(safmax) / 3.f;
ssfmin = sqrt(safmin) / eps2;
/* Compute the eigenvalues and eigenvectors of the tridiagonal
matrix. */
if (icompz == 2) {
slaset_("Full", n, n, &c_b9, &c_b10, &Z(1,1), ldz);
}
nmaxit = *n * 30;
jtot = 0;
/* Determine where the matrix splits and choose QL or QR iteration
for each block, according to whether top or bottom diagonal
element is smaller. */
l1 = 1;
nm1 = *n - 1;
L10:
if (l1 > *n) {
goto L160;
}
if (l1 > 1) {
E(l1 - 1) = 0.f;
}
if (l1 <= nm1) {
i__1 = nm1;
for (m = l1; m <= nm1; ++m) {
tst = (r__1 = E(m), dabs(r__1));
if (tst == 0.f) {
goto L30;
}
if (tst <= sqrt((r__1 = D(m), dabs(r__1))) * sqrt((r__2 = D(m + 1)
, dabs(r__2))) * eps) {
E(m) = 0.f;
goto L30;
}
/* L20: */
}
}
m = *n;
L30:
l = l1;
lsv = l;
lend = m;
lendsv = lend;
l1 = m + 1;
if (lend == l) {
goto L10;
}
/* Scale submatrix in rows and columns L to LEND */
i__1 = lend - l + 1;
anorm = slanst_("I", &i__1, &D(l), &E(l));
iscale = 0;
if (anorm == 0.f) {
goto L10;
}
if (anorm > ssfmax) {
iscale = 1;
i__1 = lend - l + 1;
slascl_("G", &c__0, &c__0, &anorm, &ssfmax, &i__1, &c__1, &D(l), n,
info);
i__1 = lend - l;
slascl_("G", &c__0, &c__0, &anorm, &ssfmax, &i__1, &c__1, &E(l), n,
info);
} else if (anorm < ssfmin) {
iscale = 2;
i__1 = lend - l + 1;
slascl_("G", &c__0, &c__0, &anorm, &ssfmin, &i__1, &c__1, &D(l), n,
info);
i__1 = lend - l;
slascl_("G", &c__0, &c__0, &anorm, &ssfmin, &i__1, &c__1, &E(l), n,
info);
}
/* Choose between QL and QR iteration */
if ((r__1 = D(lend), dabs(r__1)) < (r__2 = D(l), dabs(r__2))) {
lend = lsv;
l = lendsv;
}
if (lend > l) {
/* QL Iteration
Look for small subdiagonal element. */
L40:
if (l != lend) {
lendm1 = lend - 1;
i__1 = lendm1;
for (m = l; m <= lendm1; ++m) {
/* Computing 2nd power */
r__2 = (r__1 = E(m), dabs(r__1));
tst = r__2 * r__2;
if (tst <= eps2 * (r__1 = D(m), dabs(r__1)) * (r__2 = D(m + 1)
, dabs(r__2)) + safmin) {
goto L60;
}
/* L50: */
}
}
m = lend;
L60:
if (m < lend) {
E(m) = 0.f;
}
p = D(l);
if (m == l) {
goto L80;
}
/* If remaining matrix is 2-by-2, use SLAE2 or SLAEV2
to compute its eigensystem. */
if (m == l + 1) {
if (icompz > 0) {
slaev2_(&D(l), &E(l), &D(l + 1), &rt1, &rt2, &c, &s);
WORK(l) = c;
WORK(*n - 1 + l) = s;
slasr_("R", "V", "B", n, &c__2, &WORK(l), &WORK(*n - 1 + l), &
Z(1,l), ldz);
} else {
slae2_(&D(l), &E(l), &D(l + 1), &rt1, &rt2);
}
D(l) = rt1;
D(l + 1) = rt2;
E(l) = 0.f;
l += 2;
if (l <= lend) {
goto L40;
}
goto L140;
}
if (jtot == nmaxit) {
goto L140;
}
++jtot;
/* Form shift. */
g = (D(l + 1) - p) / (E(l) * 2.f);
r = slapy2_(&g, &c_b10);
g = D(m) - p + E(l) / (g + r_sign(&r, &g));
s = 1.f;
c = 1.f;
p = 0.f;
/* Inner loop */
mm1 = m - 1;
i__1 = l;
for (i = mm1; i >= l; --i) {
f = s * E(i);
b = c * E(i);
slartg_(&g, &f, &c, &s, &r);
if (i != m - 1) {
E(i + 1) = r;
}
g = D(i + 1) - p;
r = (D(i) - g) * s + c * 2.f * b;
p = s * r;
D(i + 1) = g + p;
g = c * r - b;
/* If eigenvectors are desired, then save rotations. */
if (icompz > 0) {
WORK(i) = c;
WORK(*n - 1 + i) = -(doublereal)s;
}
/* L70: */
}
/* If eigenvectors are desired, then apply saved rotations. */
if (icompz > 0) {
mm = m - l + 1;
slasr_("R", "V", "B", n, &mm, &WORK(l), &WORK(*n - 1 + l), &Z(1,l), ldz);
}
D(l) -= p;
E(l) = g;
goto L40;
/* Eigenvalue found. */
L80:
D(l) = p;
++l;
if (l <= lend) {
goto L40;
}
goto L140;
} else {
/* QR Iteration
Look for small superdiagonal element. */
L90:
if (l != lend) {
lendp1 = lend + 1;
i__1 = lendp1;
for (m = l; m >= lendp1; --m) {
/* Computing 2nd power */
r__2 = (r__1 = E(m - 1), dabs(r__1));
tst = r__2 * r__2;
if (tst <= eps2 * (r__1 = D(m), dabs(r__1)) * (r__2 = D(m - 1)
, dabs(r__2)) + safmin) {
goto L110;
}
/* L100: */
}
}
m = lend;
L110:
if (m > lend) {
E(m - 1) = 0.f;
}
p = D(l);
if (m == l) {
goto L130;
}
/* If remaining matrix is 2-by-2, use SLAE2 or SLAEV2
to compute its eigensystem. */
if (m == l - 1) {
if (icompz > 0) {
slaev2_(&D(l - 1), &E(l - 1), &D(l), &rt1, &rt2, &c, &s);
WORK(m) = c;
WORK(*n - 1 + m) = s;
slasr_("R", "V", "F", n, &c__2, &WORK(m), &WORK(*n - 1 + m), &
Z(1,l-1), ldz);
} else {
slae2_(&D(l - 1), &E(l - 1), &D(l), &rt1, &rt2);
}
D(l - 1) = rt1;
D(l) = rt2;
E(l - 1) = 0.f;
l += -2;
if (l >= lend) {
goto L90;
}
goto L140;
}
if (jtot == nmaxit) {
goto L140;
}
++jtot;
/* Form shift. */
g = (D(l - 1) - p) / (E(l - 1) * 2.f);
r = slapy2_(&g, &c_b10);
g = D(m) - p + E(l - 1) / (g + r_sign(&r, &g));
s = 1.f;
c = 1.f;
p = 0.f;
/* Inner loop */
lm1 = l - 1;
i__1 = lm1;
for (i = m; i <= lm1; ++i) {
f = s * E(i);
b = c * E(i);
slartg_(&g, &f, &c, &s, &r);
if (i != m) {
E(i - 1) = r;
}
g = D(i) - p;
r = (D(i + 1) - g) * s + c * 2.f * b;
p = s * r;
D(i) = g + p;
g = c * r - b;
/* If eigenvectors are desired, then save rotations. */
if (icompz > 0) {
WORK(i) = c;
WORK(*n - 1 + i) = s;
}
/* L120: */
}
/* If eigenvectors are desired, then apply saved rotations. */
if (icompz > 0) {
mm = l - m + 1;
slasr_("R", "V", "F", n, &mm, &WORK(m), &WORK(*n - 1 + m), &Z(1,m), ldz);
}
D(l) -= p;
E(lm1) = g;
goto L90;
/* Eigenvalue found. */
L130:
D(l) = p;
--l;
if (l >= lend) {
goto L90;
}
goto L140;
}
/* Undo scaling if necessary */
L140:
if (iscale == 1) {
i__1 = lendsv - lsv + 1;
slascl_("G", &c__0, &c__0, &ssfmax, &anorm, &i__1, &c__1, &D(lsv), n,
info);
i__1 = lendsv - lsv;
slascl_("G", &c__0, &c__0, &ssfmax, &anorm, &i__1, &c__1, &E(lsv), n,
info);
} else if (iscale == 2) {
i__1 = lendsv - lsv + 1;
slascl_("G", &c__0, &c__0, &ssfmin, &anorm, &i__1, &c__1, &D(lsv), n,
info);
i__1 = lendsv - lsv;
slascl_("G", &c__0, &c__0, &ssfmin, &anorm, &i__1, &c__1, &E(lsv), n,
info);
}
/* Check for no convergence to an eigenvalue after a total
of N*MAXIT iterations. */
if (jtot < nmaxit) {
goto L10;
}
i__1 = *n - 1;
for (i = 1; i <= *n-1; ++i) {
if (E(i) != 0.f) {
++(*info);
}
/* L150: */
}
goto L190;
/* Order eigenvalues and eigenvectors. */
L160:
if (icompz == 0) {
/* Use Quick Sort */
slasrt_("I", n, &D(1), info);
} else {
/* Use Selection Sort to minimize swaps of eigenvectors */
i__1 = *n;
for (ii = 2; ii <= *n; ++ii) {
i = ii - 1;
k = i;
p = D(i);
i__2 = *n;
for (j = ii; j <= *n; ++j) {
if (D(j) < p) {
k = j;
p = D(j);
}
/* L170: */
}
if (k != i) {
D(k) = D(i);
D(i) = p;
sswap_(n, &Z(1,i), &c__1, &Z(1,k), &
c__1);
}
/* L180: */
}
}
L190:
return 0;
/* End of SSTEQR */
} /* ssteqr_ */
/* Subroutine */ int sbdsdc_(char *uplo, char *compq, integer *n, real *d__,
real *e, real *u, integer *ldu, real *vt, integer *ldvt, real *q,
integer *iq, real *work, integer *iwork, integer *info)
{
/* System generated locals */
integer u_dim1, u_offset, vt_dim1, vt_offset, i__1, i__2;
real r__1;
/* Builtin functions */
double r_sign(real *, real *), log(doublereal);
/* Local variables */
static integer difl, difr, ierr, perm, mlvl, sqre, i__, j, k;
static real p, r__;
static integer z__;
extern logical lsame_(char *, char *);
static integer poles;
extern /* Subroutine */ int slasr_(char *, char *, char *, integer *,
integer *, real *, real *, real *, integer *);
static integer iuplo, nsize, start;
extern /* Subroutine */ int scopy_(integer *, real *, integer *, real *,
integer *), sswap_(integer *, real *, integer *, real *, integer *
), slasd0_(integer *, integer *, real *, real *, real *, integer *
, real *, integer *, integer *, integer *, real *, integer *);
static integer ic, ii, kk;
static real cs;
static integer is, iu;
static real sn;
extern doublereal slamch_(char *);
extern /* Subroutine */ int slasda_(integer *, integer *, integer *,
integer *, real *, real *, real *, integer *, real *, integer *,
real *, real *, real *, real *, integer *, integer *, integer *,
integer *, real *, real *, real *, real *, integer *, integer *),
xerbla_(char *, integer *);
extern integer ilaenv_(integer *, char *, char *, integer *, integer *,
integer *, integer *, ftnlen, ftnlen);
extern /* Subroutine */ int slascl_(char *, integer *, integer *, real *,
real *, integer *, integer *, real *, integer *, integer *);
static integer givcol;
extern /* Subroutine */ int slasdq_(char *, integer *, integer *, integer
*, integer *, integer *, real *, real *, real *, integer *, real *
, integer *, real *, integer *, real *, integer *);
static integer icompq;
extern /* Subroutine */ int slaset_(char *, integer *, integer *, real *,
real *, real *, integer *), slartg_(real *, real *, real *
, real *, real *);
static real orgnrm;
static integer givnum;
extern doublereal slanst_(char *, integer *, real *, real *);
static integer givptr, nm1, qstart, smlsiz, wstart, smlszp;
static real eps;
static integer ivt;
#define u_ref(a_1,a_2) u[(a_2)*u_dim1 + a_1]
#define vt_ref(a_1,a_2) vt[(a_2)*vt_dim1 + a_1]
/* -- LAPACK routine (instrumented to count ops, version 3.0) --
Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd.,
Courant Institute, Argonne National Lab, and Rice University
October 31, 1999
Purpose
=======
SBDSDC computes the singular value decomposition (SVD) of a real
N-by-N (upper or lower) bidiagonal matrix B: B = U * S * VT,
using a divide and conquer method, where S is a diagonal matrix
with non-negative diagonal elements (the singular values of B), and
U and VT are orthogonal matrices of left and right singular vectors,
respectively. SBDSDC can be used to compute all singular values,
and optionally, singular vectors or singular vectors in compact form.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none. See SLASD3 for details.
The code currently call SLASDQ if singular values only are desired.
However, it can be slightly modified to compute singular values
using the divide and conquer method.
Arguments
=========
UPLO (input) CHARACTER*1
= 'U': B is upper bidiagonal.
= 'L': B is lower bidiagonal.
COMPQ (input) CHARACTER*1
Specifies whether singular vectors are to be computed
as follows:
= 'N': Compute singular values only;
= 'P': Compute singular values and compute singular
vectors in compact form;
= 'I': Compute singular values and singular vectors.
N (input) INTEGER
The order of the matrix B. N >= 0.
D (input/output) REAL array, dimension (N)
On entry, the n diagonal elements of the bidiagonal matrix B.
On exit, if INFO=0, the singular values of B.
E (input/output) REAL array, dimension (N)
On entry, the elements of E contain the offdiagonal
elements of the bidiagonal matrix whose SVD is desired.
On exit, E has been destroyed.
U (output) REAL array, dimension (LDU,N)
If COMPQ = 'I', then:
On exit, if INFO = 0, U contains the left singular vectors
of the bidiagonal matrix.
For other values of COMPQ, U is not referenced.
LDU (input) INTEGER
The leading dimension of the array U. LDU >= 1.
If singular vectors are desired, then LDU >= max( 1, N ).
VT (output) REAL array, dimension (LDVT,N)
If COMPQ = 'I', then:
On exit, if INFO = 0, VT' contains the right singular
vectors of the bidiagonal matrix.
For other values of COMPQ, VT is not referenced.
LDVT (input) INTEGER
The leading dimension of the array VT. LDVT >= 1.
If singular vectors are desired, then LDVT >= max( 1, N ).
Q (output) REAL array, dimension (LDQ)
If COMPQ = 'P', then:
On exit, if INFO = 0, Q and IQ contain the left
and right singular vectors in a compact form,
requiring O(N log N) space instead of 2*N**2.
In particular, Q contains all the REAL data in
LDQ >= N*(11 + 2*SMLSIZ + 8*INT(LOG_2(N/(SMLSIZ+1))))
words of memory, where SMLSIZ is returned by ILAENV and
is equal to the maximum size of the subproblems at the
bottom of the computation tree (usually about 25).
For other values of COMPQ, Q is not referenced.
IQ (output) INTEGER array, dimension (LDIQ)
If COMPQ = 'P', then:
On exit, if INFO = 0, Q and IQ contain the left
and right singular vectors in a compact form,
requiring O(N log N) space instead of 2*N**2.
In particular, IQ contains all INTEGER data in
LDIQ >= N*(3 + 3*INT(LOG_2(N/(SMLSIZ+1))))
words of memory, where SMLSIZ is returned by ILAENV and
is equal to the maximum size of the subproblems at the
bottom of the computation tree (usually about 25).
For other values of COMPQ, IQ is not referenced.
WORK (workspace) REAL array, dimension (LWORK)
If COMPQ = 'N' then LWORK >= (4 * N).
If COMPQ = 'P' then LWORK >= (6 * N).
If COMPQ = 'I' then LWORK >= (3 * N**2 + 4 * N).
IWORK (workspace) INTEGER array, dimension (7*N)
INFO (output) INTEGER
= 0: successful exit.
< 0: if INFO = -i, the i-th argument had an illegal value.
> 0: The algorithm failed to compute an singular value.
The update process of divide and conquer failed.
Further Details
===============
Based on contributions by
Ming Gu and Huan Ren, Computer Science Division, University of
California at Berkeley, USA
=====================================================================
Test the input parameters.
Parameter adjustments */
--d__;
--e;
u_dim1 = *ldu;
u_offset = 1 + u_dim1 * 1;
u -= u_offset;
vt_dim1 = *ldvt;
vt_offset = 1 + vt_dim1 * 1;
vt -= vt_offset;
--q;
--iq;
--work;
--iwork;
/* Function Body */
*info = 0;
iuplo = 0;
if (lsame_(uplo, "U")) {
iuplo = 1;
}
if (lsame_(uplo, "L")) {
iuplo = 2;
}
if (lsame_(compq, "N")) {
icompq = 0;
} else if (lsame_(compq, "P")) {
icompq = 1;
} else if (lsame_(compq, "I")) {
icompq = 2;
} else {
icompq = -1;
}
if (iuplo == 0) {
*info = -1;
} else if (icompq < 0) {
*info = -2;
} else if (*n < 0) {
*info = -3;
} else if (*ldu < 1 || icompq == 2 && *ldu < *n) {
*info = -7;
} else if (*ldvt < 1 || icompq == 2 && *ldvt < *n) {
*info = -9;
}
if (*info != 0) {
i__1 = -(*info);
xerbla_("SBDSDC", &i__1);
return 0;
}
/* Quick return if possible */
if (*n == 0) {
return 0;
}
smlsiz = ilaenv_(&c__9, "SBDSDC", " ", &c__0, &c__0, &c__0, &c__0, (
ftnlen)6, (ftnlen)1);
if (*n == 1) {
if (icompq == 1) {
q[1] = r_sign(&c_b15, &d__[1]);
q[smlsiz * *n + 1] = 1.f;
} else if (icompq == 2) {
u_ref(1, 1) = r_sign(&c_b15, &d__[1]);
vt_ref(1, 1) = 1.f;
}
d__[1] = dabs(d__[1]);
return 0;
}
nm1 = *n - 1;
/* If matrix lower bidiagonal, rotate to be upper bidiagonal
by applying Givens rotations on the left */
wstart = 1;
qstart = 3;
if (icompq == 1) {
scopy_(n, &d__[1], &c__1, &q[1], &c__1);
i__1 = *n - 1;
scopy_(&i__1, &e[1], &c__1, &q[*n + 1], &c__1);
}
if (iuplo == 2) {
qstart = 5;
wstart = (*n << 1) - 1;
latime_1.ops += (real) (*n - 1 << 3);
i__1 = *n - 1;
for (i__ = 1; i__ <= i__1; ++i__) {
slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
d__[i__] = r__;
e[i__] = sn * d__[i__ + 1];
d__[i__ + 1] = cs * d__[i__ + 1];
if (icompq == 1) {
q[i__ + (*n << 1)] = cs;
q[i__ + *n * 3] = sn;
} else if (icompq == 2) {
work[i__] = cs;
work[nm1 + i__] = -sn;
}
/* L10: */
}
}
/* If ICOMPQ = 0, use SLASDQ to compute the singular values. */
if (icompq == 0) {
slasdq_("U", &c__0, n, &c__0, &c__0, &c__0, &d__[1], &e[1], &vt[
vt_offset], ldvt, &u[u_offset], ldu, &u[u_offset], ldu, &work[
wstart], info);
goto L40;
}
/* If N is smaller than the minimum divide size SMLSIZ, then solve
the problem with another solver. */
if (*n <= smlsiz) {
if (icompq == 2) {
slaset_("A", n, n, &c_b29, &c_b15, &u[u_offset], ldu);
slaset_("A", n, n, &c_b29, &c_b15, &vt[vt_offset], ldvt);
slasdq_("U", &c__0, n, n, n, &c__0, &d__[1], &e[1], &vt[vt_offset]
, ldvt, &u[u_offset], ldu, &u[u_offset], ldu, &work[
wstart], info);
} else if (icompq == 1) {
iu = 1;
ivt = iu + *n;
slaset_("A", n, n, &c_b29, &c_b15, &q[iu + (qstart - 1) * *n], n);
slaset_("A", n, n, &c_b29, &c_b15, &q[ivt + (qstart - 1) * *n], n);
slasdq_("U", &c__0, n, n, n, &c__0, &d__[1], &e[1], &q[ivt + (
qstart - 1) * *n], n, &q[iu + (qstart - 1) * *n], n, &q[
iu + (qstart - 1) * *n], n, &work[wstart], info);
}
goto L40;
}
if (icompq == 2) {
slaset_("A", n, n, &c_b29, &c_b15, &u[u_offset], ldu);
slaset_("A", n, n, &c_b29, &c_b15, &vt[vt_offset], ldvt);
}
/* Scale. */
orgnrm = slanst_("M", n, &d__[1], &e[1]);
if (orgnrm == 0.f) {
return 0;
}
latime_1.ops += (real) (*n + nm1);
slascl_("G", &c__0, &c__0, &orgnrm, &c_b15, n, &c__1, &d__[1], n, &ierr);
slascl_("G", &c__0, &c__0, &orgnrm, &c_b15, &nm1, &c__1, &e[1], &nm1, &
ierr);
eps = slamch_("Epsilon");
mlvl = (integer) (log((real) (*n) / (real) (smlsiz + 1)) / log(2.f)) + 1;
smlszp = smlsiz + 1;
if (icompq == 1) {
iu = 1;
ivt = smlsiz + 1;
difl = ivt + smlszp;
difr = difl + mlvl;
z__ = difr + (mlvl << 1);
ic = z__ + mlvl;
is = ic + 1;
poles = is + 1;
givnum = poles + (mlvl << 1);
k = 1;
givptr = 2;
perm = 3;
givcol = perm + mlvl;
}
i__1 = *n;
for (i__ = 1; i__ <= i__1; ++i__) {
if ((r__1 = d__[i__], dabs(r__1)) < eps) {
d__[i__] = r_sign(&eps, &d__[i__]);
}
/* L20: */
}
start = 1;
sqre = 0;
i__1 = nm1;
for (i__ = 1; i__ <= i__1; ++i__) {
if ((r__1 = e[i__], dabs(r__1)) < eps || i__ == nm1) {
/* Subproblem found. First determine its size and then
apply divide and conquer on it. */
if (i__ < nm1) {
/* A subproblem with E(I) small for I < NM1. */
nsize = i__ - start + 1;
} else if ((r__1 = e[i__], dabs(r__1)) >= eps) {
/* A subproblem with E(NM1) not too small but I = NM1. */
nsize = *n - start + 1;
} else {
/* A subproblem with E(NM1) small. This implies an
1-by-1 subproblem at D(N). Solve this 1-by-1 problem
first. */
nsize = i__ - start + 1;
if (icompq == 2) {
u_ref(*n, *n) = r_sign(&c_b15, &d__[*n]);
vt_ref(*n, *n) = 1.f;
} else if (icompq == 1) {
q[*n + (qstart - 1) * *n] = r_sign(&c_b15, &d__[*n]);
q[*n + (smlsiz + qstart - 1) * *n] = 1.f;
}
d__[*n] = (r__1 = d__[*n], dabs(r__1));
}
if (icompq == 2) {
slasd0_(&nsize, &sqre, &d__[start], &e[start], &u_ref(start,
start), ldu, &vt_ref(start, start), ldvt, &smlsiz, &
iwork[1], &work[wstart], info);
} else {
slasda_(&icompq, &smlsiz, &nsize, &sqre, &d__[start], &e[
start], &q[start + (iu + qstart - 2) * *n], n, &q[
start + (ivt + qstart - 2) * *n], &iq[start + k * *n],
&q[start + (difl + qstart - 2) * *n], &q[start + (
difr + qstart - 2) * *n], &q[start + (z__ + qstart -
2) * *n], &q[start + (poles + qstart - 2) * *n], &iq[
start + givptr * *n], &iq[start + givcol * *n], n, &
iq[start + perm * *n], &q[start + (givnum + qstart -
2) * *n], &q[start + (ic + qstart - 2) * *n], &q[
start + (is + qstart - 2) * *n], &work[wstart], &
iwork[1], info);
if (*info != 0) {
return 0;
}
}
start = i__ + 1;
}
/* L30: */
}
/* Unscale */
latime_1.ops += (real) (*n);
slascl_("G", &c__0, &c__0, &c_b15, &orgnrm, n, &c__1, &d__[1], n, &ierr);
L40:
/* Use Selection Sort to minimize swaps of singular vectors */
i__1 = *n;
for (ii = 2; ii <= i__1; ++ii) {
i__ = ii - 1;
kk = i__;
p = d__[i__];
i__2 = *n;
for (j = ii; j <= i__2; ++j) {
if (d__[j] > p) {
kk = j;
p = d__[j];
}
/* L50: */
}
if (kk != i__) {
d__[kk] = d__[i__];
d__[i__] = p;
if (icompq == 1) {
iq[i__] = kk;
} else if (icompq == 2) {
sswap_(n, &u_ref(1, i__), &c__1, &u_ref(1, kk), &c__1);
sswap_(n, &vt_ref(i__, 1), ldvt, &vt_ref(kk, 1), ldvt);
}
} else if (icompq == 1) {
iq[i__] = i__;
}
/* L60: */
}
/* If ICOMPQ = 1, use IQ(N,1) as the indicator for UPLO */
if (icompq == 1) {
if (iuplo == 1) {
iq[*n] = 1;
} else {
iq[*n] = 0;
}
}
/* If B is lower bidiagonal, update U by those Givens rotations
which rotated B to be upper bidiagonal */
if (iuplo == 2 && icompq == 2) {
latime_1.ops += (real) ((*n - 1) * 6 * *n);
slasr_("L", "V", "B", n, n, &work[1], &work[*n], &u[u_offset], ldu);
}
return 0;
/* End of SBDSDC */
} /* sbdsdc_ */