1       SUBROUTINE ZSTEMR( JOBZ, RANGE, N, D, E, VL, VU, IL, IU,
  2      $                   M, W, Z, LDZ, NZC, ISUPPZ, TRYRAC, WORK, LWORK,
  3      $                   IWORK, LIWORK, INFO )
  4       IMPLICIT NONE
  5 *
  6 *  -- LAPACK computational routine (version 3.2.1)                    --
  7 *
  8 *  -- April 2009                                                      --
  9 *
 10 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
 11 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
 12 *
 13 *     .. Scalar Arguments ..
 14       CHARACTER          JOBZ, RANGE
 15       LOGICAL            TRYRAC
 16       INTEGER            IL, INFO, IU, LDZ, NZC, LIWORK, LWORK, M, N
 17       DOUBLE PRECISION VL, VU
 18 *     ..
 19 *     .. Array Arguments ..
 20       INTEGER            ISUPPZ( * ), IWORK( * )
 21       DOUBLE PRECISION   D( * ), E( * ), W( * ), WORK( * )
 22       COMPLEX*16         Z( LDZ, * )
 23 *     ..
 24 *
 25 *  Purpose
 26 *  =======
 27 *
 28 *  ZSTEMR computes selected eigenvalues and, optionally, eigenvectors
 29 *  of a real symmetric tridiagonal matrix T. Any such unreduced matrix has
 30 *  a well defined set of pairwise different real eigenvalues, the corresponding
 31 *  real eigenvectors are pairwise orthogonal.
 32 *
 33 *  The spectrum may be computed either completely or partially by specifying
 34 *  either an interval (VL,VU] or a range of indices IL:IU for the desired
 35 *  eigenvalues.
 36 *
 37 *  Depending on the number of desired eigenvalues, these are computed either
 38 *  by bisection or the dqds algorithm. Numerically orthogonal eigenvectors are
 39 *  computed by the use of various suitable L D L^T factorizations near clusters
 40 *  of close eigenvalues (referred to as RRRs, Relatively Robust
 41 *  Representations). An informal sketch of the algorithm follows.
 42 *
 43 *  For each unreduced block (submatrix) of T,
 44 *     (a) Compute T - sigma I  = L D L^T, so that L and D
 45 *         define all the wanted eigenvalues to high relative accuracy.
 46 *         This means that small relative changes in the entries of D and L
 47 *         cause only small relative changes in the eigenvalues and
 48 *         eigenvectors. The standard (unfactored) representation of the
 49 *         tridiagonal matrix T does not have this property in general.
 50 *     (b) Compute the eigenvalues to suitable accuracy.
 51 *         If the eigenvectors are desired, the algorithm attains full
 52 *         accuracy of the computed eigenvalues only right before
 53 *         the corresponding vectors have to be computed, see steps c) and d).
 54 *     (c) For each cluster of close eigenvalues, select a new
 55 *         shift close to the cluster, find a new factorization, and refine
 56 *         the shifted eigenvalues to suitable accuracy.
 57 *     (d) For each eigenvalue with a large enough relative separation compute
 58 *         the corresponding eigenvector by forming a rank revealing twisted
 59 *         factorization. Go back to (c) for any clusters that remain.
 60 *
 61 *  For more details, see:
 62 *  - Inderjit S. Dhillon and Beresford N. Parlett: "Multiple representations
 63 *    to compute orthogonal eigenvectors of symmetric tridiagonal matrices,"
 64 *    Linear Algebra and its Applications, 387(1), pp. 1-28, August 2004.
 65 *  - Inderjit Dhillon and Beresford Parlett: "Orthogonal Eigenvectors and
 66 *    Relative Gaps," SIAM Journal on Matrix Analysis and Applications, Vol. 25,
 67 *    2004.  Also LAPACK Working Note 154.
 68 *  - Inderjit Dhillon: "A new O(n^2) algorithm for the symmetric
 69 *    tridiagonal eigenvalue/eigenvector problem",
 70 *    Computer Science Division Technical Report No. UCB/CSD-97-971,
 71 *    UC Berkeley, May 1997.
 72 *
 73 *  Further Details
 74 *  1.ZSTEMR works only on machines which follow IEEE-754
 75 *  floating-point standard in their handling of infinities and NaNs.
 76 *  This permits the use of efficient inner loops avoiding a check for
 77 *  zero divisors.
 78 *
 79 *  2. LAPACK routines can be used to reduce a complex Hermitean matrix to
 80 *  real symmetric tridiagonal form.
 81 *
 82 *  (Any complex Hermitean tridiagonal matrix has real values on its diagonal
 83 *  and potentially complex numbers on its off-diagonals. By applying a
 84 *  similarity transform with an appropriate diagonal matrix
 85 *  diag(1,e^{i \phy_1}, ... , e^{i \phy_{n-1}}), the complex Hermitean
 86 *  matrix can be transformed into a real symmetric matrix and complex
 87 *  arithmetic can be entirely avoided.)
 88 *
 89 *  While the eigenvectors of the real symmetric tridiagonal matrix are real,
 90 *  the eigenvectors of original complex Hermitean matrix have complex entries
 91 *  in general.
 92 *  Since LAPACK drivers overwrite the matrix data with the eigenvectors,
 93 *  ZSTEMR accepts complex workspace to facilitate interoperability
 94 *  with ZUNMTR or ZUPMTR.
 95 *
 96 *  Arguments
 97 *  =========
 98 *
 99 *  JOBZ    (input) CHARACTER*1
100 *          = 'N':  Compute eigenvalues only;
101 *          = 'V':  Compute eigenvalues and eigenvectors.
102 *
103 *  RANGE   (input) CHARACTER*1
104 *          = 'A': all eigenvalues will be found.
105 *          = 'V': all eigenvalues in the half-open interval (VL,VU]
106 *                 will be found.
107 *          = 'I': the IL-th through IU-th eigenvalues will be found.
108 *
109 *  N       (input) INTEGER
110 *          The order of the matrix.  N >= 0.
111 *
112 *  D       (input/output) DOUBLE PRECISION array, dimension (N)
113 *          On entry, the N diagonal elements of the tridiagonal matrix
114 *          T. On exit, D is overwritten.
115 *
116 *  E       (input/output) DOUBLE PRECISION array, dimension (N)
117 *          On entry, the (N-1) subdiagonal elements of the tridiagonal
118 *          matrix T in elements 1 to N-1 of E. E(N) need not be set on
119 *          input, but is used internally as workspace.
120 *          On exit, E is overwritten.
121 *
122 *  VL      (input) DOUBLE PRECISION
123 *  VU      (input) DOUBLE PRECISION
124 *          If RANGE='V', the lower and upper bounds of the interval to
125 *          be searched for eigenvalues. VL < VU.
126 *          Not referenced if RANGE = 'A' or 'I'.
127 *
128 *  IL      (input) INTEGER
129 *  IU      (input) INTEGER
130 *          If RANGE='I', the indices (in ascending order) of the
131 *          smallest and largest eigenvalues to be returned.
132 *          1 <= IL <= IU <= N, if N > 0.
133 *          Not referenced if RANGE = 'A' or 'V'.
134 *
135 *  M       (output) INTEGER
136 *          The total number of eigenvalues found.  0 <= M <= N.
137 *          If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
138 *
139 *  W       (output) DOUBLE PRECISION array, dimension (N)
140 *          The first M elements contain the selected eigenvalues in
141 *          ascending order.
142 *
143 *  Z       (output) COMPLEX*16 array, dimension (LDZ, max(1,M) )
144 *          If JOBZ = 'V', and if INFO = 0, then the first M columns of Z
145 *          contain the orthonormal eigenvectors of the matrix T
146 *          corresponding to the selected eigenvalues, with the i-th
147 *          column of Z holding the eigenvector associated with W(i).
148 *          If JOBZ = 'N', then Z is not referenced.
149 *          Note: the user must ensure that at least max(1,M) columns are
150 *          supplied in the array Z; if RANGE = 'V', the exact value of M
151 *          is not known in advance and can be computed with a workspace
152 *          query by setting NZC = -1, see below.
153 *
154 *  LDZ     (input) INTEGER
155 *          The leading dimension of the array Z.  LDZ >= 1, and if
156 *          JOBZ = 'V', then LDZ >= max(1,N).
157 *
158 *  NZC     (input) INTEGER
159 *          The number of eigenvectors to be held in the array Z.
160 *          If RANGE = 'A', then NZC >= max(1,N).
161 *          If RANGE = 'V', then NZC >= the number of eigenvalues in (VL,VU].
162 *          If RANGE = 'I', then NZC >= IU-IL+1.
163 *          If NZC = -1, then a workspace query is assumed; the
164 *          routine calculates the number of columns of the array Z that
165 *          are needed to hold the eigenvectors.
166 *          This value is returned as the first entry of the Z array, and
167 *          no error message related to NZC is issued by XERBLA.
168 *
169 *  ISUPPZ  (output) INTEGER ARRAY, dimension ( 2*max(1,M) )
170 *          The support of the eigenvectors in Z, i.e., the indices
171 *          indicating the nonzero elements in Z. The i-th computed eigenvector
172 *          is nonzero only in elements ISUPPZ( 2*i-1 ) through
173 *          ISUPPZ( 2*i ). This is relevant in the case when the matrix
174 *          is split. ISUPPZ is only accessed when JOBZ is 'V' and N > 0.
175 *
176 *  TRYRAC  (input/output) LOGICAL
177 *          If TRYRAC.EQ..TRUE., indicates that the code should check whether
178 *          the tridiagonal matrix defines its eigenvalues to high relative
179 *          accuracy.  If so, the code uses relative-accuracy preserving
180 *          algorithms that might be (a bit) slower depending on the matrix.
181 *          If the matrix does not define its eigenvalues to high relative
182 *          accuracy, the code can uses possibly faster algorithms.
183 *          If TRYRAC.EQ..FALSE., the code is not required to guarantee
184 *          relatively accurate eigenvalues and can use the fastest possible
185 *          techniques.
186 *          On exit, a .TRUE. TRYRAC will be set to .FALSE. if the matrix
187 *          does not define its eigenvalues to high relative accuracy.
188 *
189 *  WORK    (workspace/output) DOUBLE PRECISION array, dimension (LWORK)
190 *          On exit, if INFO = 0, WORK(1) returns the optimal
191 *          (and minimal) LWORK.
192 *
193 *  LWORK   (input) INTEGER
194 *          The dimension of the array WORK. LWORK >= max(1,18*N)
195 *          if JOBZ = 'V', and LWORK >= max(1,12*N) if JOBZ = 'N'.
196 *          If LWORK = -1, then a workspace query is assumed; the routine
197 *          only calculates the optimal size of the WORK array, returns
198 *          this value as the first entry of the WORK array, and no error
199 *          message related to LWORK is issued by XERBLA.
200 *
201 *  IWORK   (workspace/output) INTEGER array, dimension (LIWORK)
202 *          On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
203 *
204 *  LIWORK  (input) INTEGER
205 *          The dimension of the array IWORK.  LIWORK >= max(1,10*N)
206 *          if the eigenvectors are desired, and LIWORK >= max(1,8*N)
207 *          if only the eigenvalues are to be computed.
208 *          If LIWORK = -1, then a workspace query is assumed; the
209 *          routine only calculates the optimal size of the IWORK array,
210 *          returns this value as the first entry of the IWORK array, and
211 *          no error message related to LIWORK is issued by XERBLA.
212 *
213 *  INFO    (output) INTEGER
214 *          On exit, INFO
215 *          = 0:  successful exit
216 *          < 0:  if INFO = -i, the i-th argument had an illegal value
217 *          > 0:  if INFO = 1X, internal error in DLARRE,
218 *                if INFO = 2X, internal error in ZLARRV.
219 *                Here, the digit X = ABS( IINFO ) < 10, where IINFO is
220 *                the nonzero error code returned by DLARRE or
221 *                ZLARRV, respectively.
222 *
223 *
224 *  Further Details
225 *  ===============
226 *
227 *  Based on contributions by
228 *     Beresford Parlett, University of California, Berkeley, USA
229 *     Jim Demmel, University of California, Berkeley, USA
230 *     Inderjit Dhillon, University of Texas, Austin, USA
231 *     Osni Marques, LBNL/NERSC, USA
232 *     Christof Voemel, University of California, Berkeley, USA
233 *
234 *  =====================================================================
235 *
236 *     .. Parameters ..
237       DOUBLE PRECISION   ZERO, ONE, FOUR, MINRGP
238       PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0,
239      $                     FOUR = 4.0D0,
240      $                     MINRGP = 1.0D-3 )
241 *     ..
242 *     .. Local Scalars ..
243       LOGICAL            ALLEIG, INDEIG, LQUERY, VALEIG, WANTZ, ZQUERY
244       INTEGER            I, IBEGIN, IEND, IFIRST, IIL, IINDBL, IINDW,
245      $                   IINDWK, IINFO, IINSPL, IIU, ILAST, IN, INDD,
246      $                   INDE2, INDERR, INDGP, INDGRS, INDWRK, ITMP,
247      $                   ITMP2, J, JBLK, JJ, LIWMIN, LWMIN, NSPLIT,
248      $                   NZCMIN, OFFSET, WBEGIN, WEND
249       DOUBLE PRECISION   BIGNUM, CS, EPS, PIVMIN, R1, R2, RMAX, RMIN,
250      $                   RTOL1, RTOL2, SAFMIN, SCALE, SMLNUM, SN,
251      $                   THRESH, TMP, TNRM, WL, WU
252 *     ..
253 *     ..
254 *     .. External Functions ..
255       LOGICAL            LSAME
256       DOUBLE PRECISION   DLAMCH, DLANST
257       EXTERNAL           LSAME, DLAMCH, DLANST
258 *     ..
259 *     .. External Subroutines ..
260       EXTERNAL           DCOPY, DLAE2, DLAEV2, DLARRC, DLARRE, DLARRJ,
261      $                   DLARRR, DLASRT, DSCAL, XERBLA, ZLARRV, ZSWAP
262 *     ..
263 *     .. Intrinsic Functions ..
264       INTRINSIC          MAXMINSQRT
265 
266 
267 *     ..
268 *     .. Executable Statements ..
269 *
270 *     Test the input parameters.
271 *
272       WANTZ = LSAME( JOBZ, 'V' )
273       ALLEIG = LSAME( RANGE'A' )
274       VALEIG = LSAME( RANGE'V' )
275       INDEIG = LSAME( RANGE'I' )
276 *
277       LQUERY = ( ( LWORK.EQ.-1 ).OR.( LIWORK.EQ.-1 ) )
278       ZQUERY = ( NZC.EQ.-1 )
279 
280 *     DSTEMR needs WORK of size 6*N, IWORK of size 3*N.
281 *     In addition, DLARRE needs WORK of size 6*N, IWORK of size 5*N.
282 *     Furthermore, ZLARRV needs WORK of size 12*N, IWORK of size 7*N.
283       IF( WANTZ ) THEN
284          LWMIN = 18*N
285          LIWMIN = 10*N
286       ELSE
287 *        need less workspace if only the eigenvalues are wanted
288          LWMIN = 12*N
289          LIWMIN = 8*N
290       ENDIF
291 
292       WL = ZERO
293       WU = ZERO
294       IIL = 0
295       IIU = 0
296 
297       IF( VALEIG ) THEN
298 *        We do not reference VL, VU in the cases RANGE = 'I','A'
299 *        The interval (WL, WU] contains all the wanted eigenvalues.
300 *        It is either given by the user or computed in DLARRE.
301          WL = VL
302          WU = VU
303       ELSEIF( INDEIG ) THEN
304 *        We do not reference IL, IU in the cases RANGE = 'V','A'
305          IIL = IL
306          IIU = IU
307       ENDIF
308 *
309       INFO = 0
310       IF.NOT.( WANTZ .OR. LSAME( JOBZ, 'N' ) ) ) THEN
311          INFO = -1
312       ELSE IF.NOT.( ALLEIG .OR. VALEIG .OR. INDEIG ) ) THEN
313          INFO = -2
314       ELSE IF( N.LT.0 ) THEN
315          INFO = -3
316       ELSE IF( VALEIG .AND. N.GT.0 .AND. WU.LE.WL ) THEN
317          INFO = -7
318       ELSE IF( INDEIG .AND. ( IIL.LT.1 .OR. IIL.GT.N ) ) THEN
319          INFO = -8
320       ELSE IF( INDEIG .AND. ( IIU.LT.IIL .OR. IIU.GT.N ) ) THEN
321          INFO = -9
322       ELSE IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
323          INFO = -13
324       ELSE IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
325          INFO = -17
326       ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
327          INFO = -19
328       END IF
329 *
330 *     Get machine constants.
331 *
332       SAFMIN = DLAMCH( 'Safe minimum' )
333       EPS = DLAMCH( 'Precision' )
334       SMLNUM = SAFMIN / EPS
335       BIGNUM = ONE / SMLNUM
336       RMIN = SQRT( SMLNUM )
337       RMAX = MINSQRT( BIGNUM ), ONE / SQRTSQRT( SAFMIN ) ) )
338 *
339       IF( INFO.EQ.0 ) THEN
340          WORK( 1 ) = LWMIN
341          IWORK( 1 ) = LIWMIN
342 *
343          IF( WANTZ .AND. ALLEIG ) THEN
344             NZCMIN = N
345          ELSE IF( WANTZ .AND. VALEIG ) THEN
346             CALL DLARRC( 'T', N, VL, VU, D, E, SAFMIN,
347      $                            NZCMIN, ITMP, ITMP2, INFO )
348          ELSE IF( WANTZ .AND. INDEIG ) THEN
349             NZCMIN = IIU-IIL+1
350          ELSE
351 *           WANTZ .EQ. FALSE.
352             NZCMIN = 0
353          ENDIF
354          IF( ZQUERY .AND. INFO.EQ.0 ) THEN
355             Z( 1,1 ) = NZCMIN
356          ELSE IF( NZC.LT.NZCMIN .AND. .NOT.ZQUERY ) THEN
357             INFO = -14
358          END IF
359       END IF
360 
361       IF( INFO.NE.0 ) THEN
362 *
363          CALL XERBLA( 'ZSTEMR'-INFO )
364 *
365          RETURN
366       ELSE IF( LQUERY .OR. ZQUERY ) THEN
367          RETURN
368       END IF
369 *
370 *     Handle N = 0, 1, and 2 cases immediately
371 *
372       M = 0
373       IF( N.EQ.0 )
374      $   RETURN
375 *
376       IF( N.EQ.1 ) THEN
377          IF( ALLEIG .OR. INDEIG ) THEN
378             M = 1
379             W( 1 ) = D( 1 )
380          ELSE
381             IF( WL.LT.D( 1 ) .AND. WU.GE.D( 1 ) ) THEN
382                M = 1
383                W( 1 ) = D( 1 )
384             END IF
385          END IF
386          IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
387             Z( 11 ) = ONE
388             ISUPPZ(1= 1
389             ISUPPZ(2= 1
390          END IF
391          RETURN
392       END IF
393 *
394       IF( N.EQ.2 ) THEN
395          IF.NOT.WANTZ ) THEN
396             CALL DLAE2( D(1), E(1), D(2), R1, R2 )
397          ELSE IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
398             CALL DLAEV2( D(1), E(1), D(2), R1, R2, CS, SN )
399          END IF
400          IF( ALLEIG.OR.
401      $      (VALEIG.AND.(R2.GT.WL).AND.
402      $                  (R2.LE.WU)).OR.
403      $      (INDEIG.AND.(IIL.EQ.1)) ) THEN
404             M = M+1
405             W( M ) = R2
406             IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
407                Z( 1, M ) = -SN
408                Z( 2, M ) = CS
409 *              Note: At most one of SN and CS can be zero.
410                IF (SN.NE.ZERO) THEN
411                   IF (CS.NE.ZERO) THEN
412                      ISUPPZ(2*M-1= 1
413                      ISUPPZ(2*M-1= 2
414                   ELSE
415                      ISUPPZ(2*M-1= 1
416                      ISUPPZ(2*M-1= 1
417                   END IF
418                ELSE
419                   ISUPPZ(2*M-1= 2
420                   ISUPPZ(2*M) = 2
421                END IF
422             ENDIF
423          ENDIF
424          IF( ALLEIG.OR.
425      $      (VALEIG.AND.(R1.GT.WL).AND.
426      $                  (R1.LE.WU)).OR.
427      $      (INDEIG.AND.(IIU.EQ.2)) ) THEN
428             M = M+1
429             W( M ) = R1
430             IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
431                Z( 1, M ) = CS
432                Z( 2, M ) = SN
433 *              Note: At most one of SN and CS can be zero.
434                IF (SN.NE.ZERO) THEN
435                   IF (CS.NE.ZERO) THEN
436                      ISUPPZ(2*M-1= 1
437                      ISUPPZ(2*M-1= 2
438                   ELSE
439                      ISUPPZ(2*M-1= 1
440                      ISUPPZ(2*M-1= 1
441                   END IF
442                ELSE
443                   ISUPPZ(2*M-1= 2
444                   ISUPPZ(2*M) = 2
445                END IF
446             ENDIF
447          ENDIF
448          RETURN
449       END IF
450 
451 *     Continue with general N
452 
453       INDGRS = 1
454       INDERR = 2*+ 1
455       INDGP = 3*+ 1
456       INDD = 4*+ 1
457       INDE2 = 5*+ 1
458       INDWRK = 6*+ 1
459 *
460       IINSPL = 1
461       IINDBL = N + 1
462       IINDW = 2*+ 1
463       IINDWK = 3*+ 1
464 *
465 *     Scale matrix to allowable range, if necessary.
466 *     The allowable range is related to the PIVMIN parameter; see the
467 *     comments in DLARRD.  The preference for scaling small values
468 *     up is heuristic; we expect users' matrices not to be close to the
469 *     RMAX threshold.
470 *
471       SCALE = ONE
472       TNRM = DLANST( 'M', N, D, E )
473       IF( TNRM.GT.ZERO .AND. TNRM.LT.RMIN ) THEN
474          SCALE = RMIN / TNRM
475       ELSE IF( TNRM.GT.RMAX ) THEN
476          SCALE = RMAX / TNRM
477       END IF
478       IFSCALE.NE.ONE ) THEN
479          CALL DSCAL( N, SCALE, D, 1 )
480          CALL DSCAL( N-1SCALE, E, 1 )
481          TNRM = TNRM*SCALE
482          IF( VALEIG ) THEN
483 *           If eigenvalues in interval have to be found,
484 *           scale (WL, WU] accordingly
485             WL = WL*SCALE
486             WU = WU*SCALE
487          ENDIF
488       END IF
489 *
490 *     Compute the desired eigenvalues of the tridiagonal after splitting
491 *     into smaller subblocks if the corresponding off-diagonal elements
492 *     are small
493 *     THRESH is the splitting parameter for DLARRE
494 *     A negative THRESH forces the old splitting criterion based on the
495 *     size of the off-diagonal. A positive THRESH switches to splitting
496 *     which preserves relative accuracy.
497 *
498       IF( TRYRAC ) THEN
499 *        Test whether the matrix warrants the more expensive relative approach.
500          CALL DLARRR( N, D, E, IINFO )
501       ELSE
502 *        The user does not care about relative accurately eigenvalues
503          IINFO = -1
504       ENDIF
505 *     Set the splitting criterion
506       IF (IINFO.EQ.0THEN
507          THRESH = EPS
508       ELSE
509          THRESH = -EPS
510 *        relative accuracy is desired but T does not guarantee it
511          TRYRAC = .FALSE.
512       ENDIF
513 *
514       IF( TRYRAC ) THEN
515 *        Copy original diagonal, needed to guarantee relative accuracy
516          CALL DCOPY(N,D,1,WORK(INDD),1)
517       ENDIF
518 *     Store the squares of the offdiagonal values of T
519       DO 5 J = 1, N-1
520          WORK( INDE2+J-1 ) = E(J)**2
521  5    CONTINUE
522 
523 *     Set the tolerance parameters for bisection
524       IF.NOT.WANTZ ) THEN
525 *        DLARRE computes the eigenvalues to full precision.
526          RTOL1 = FOUR * EPS
527          RTOL2 = FOUR * EPS
528       ELSE
529 *        DLARRE computes the eigenvalues to less than full precision.
530 *        ZLARRV will refine the eigenvalue approximations, and we only
531 *        need less accurate initial bisection in DLARRE.
532 *        Note: these settings do only affect the subset case and DLARRE
533          RTOL1 = SQRT(EPS)
534          RTOL2 = MAXSQRT(EPS)*5.0D-3, FOUR * EPS )
535       ENDIF
536       CALL DLARRE( RANGE, N, WL, WU, IIL, IIU, D, E,
537      $             WORK(INDE2), RTOL1, RTOL2, THRESH, NSPLIT,
538      $             IWORK( IINSPL ), M, W, WORK( INDERR ),
539      $             WORK( INDGP ), IWORK( IINDBL ),
540      $             IWORK( IINDW ), WORK( INDGRS ), PIVMIN,
541      $             WORK( INDWRK ), IWORK( IINDWK ), IINFO )
542       IF( IINFO.NE.0 ) THEN
543          INFO = 10 + ABS( IINFO )
544          RETURN
545       END IF
546 *     Note that if RANGE .NE. 'V', DLARRE computes bounds on the desired
547 *     part of the spectrum. All desired eigenvalues are contained in
548 *     (WL,WU]
549 
550 
551       IF( WANTZ ) THEN
552 *
553 *        Compute the desired eigenvectors corresponding to the computed
554 *        eigenvalues
555 *
556          CALL ZLARRV( N, WL, WU, D, E,
557      $                PIVMIN, IWORK( IINSPL ), M,
558      $                1, M, MINRGP, RTOL1, RTOL2,
559      $                W, WORK( INDERR ), WORK( INDGP ), IWORK( IINDBL ),
560      $                IWORK( IINDW ), WORK( INDGRS ), Z, LDZ,
561      $                ISUPPZ, WORK( INDWRK ), IWORK( IINDWK ), IINFO )
562          IF( IINFO.NE.0 ) THEN
563             INFO = 20 + ABS( IINFO )
564             RETURN
565          END IF
566       ELSE
567 *        DLARRE computes eigenvalues of the (shifted) root representation
568 *        ZLARRV returns the eigenvalues of the unshifted matrix.
569 *        However, if the eigenvectors are not desired by the user, we need
570 *        to apply the corresponding shifts from DLARRE to obtain the
571 *        eigenvalues of the original matrix.
572          DO 20 J = 1, M
573             ITMP = IWORK( IINDBL+J-1 )
574             W( J ) = W( J ) + E( IWORK( IINSPL+ITMP-1 ) )
575  20      CONTINUE
576       END IF
577 *
578 
579       IF ( TRYRAC ) THEN
580 *        Refine computed eigenvalues so that they are relatively accurate
581 *        with respect to the original matrix T.
582          IBEGIN = 1
583          WBEGIN = 1
584          DO 39  JBLK = 1, IWORK( IINDBL+M-1 )
585             IEND = IWORK( IINSPL+JBLK-1 )
586             IN = IEND - IBEGIN + 1
587             WEND = WBEGIN - 1
588 *           check if any eigenvalues have to be refined in this block
589  36         CONTINUE
590             IF( WEND.LT.M ) THEN
591                IF( IWORK( IINDBL+WEND ).EQ.JBLK ) THEN
592                   WEND = WEND + 1
593                   GO TO 36
594                END IF
595             END IF
596             IF( WEND.LT.WBEGIN ) THEN
597                IBEGIN = IEND + 1
598                GO TO 39
599             END IF
600 
601             OFFSET = IWORK(IINDW+WBEGIN-1)-1
602             IFIRST = IWORK(IINDW+WBEGIN-1)
603             ILAST = IWORK(IINDW+WEND-1)
604             RTOL2 = FOUR * EPS
605             CALL DLARRJ( IN,
606      $                   WORK(INDD+IBEGIN-1), WORK(INDE2+IBEGIN-1),
607      $                   IFIRST, ILAST, RTOL2, OFFSET, W(WBEGIN),
608      $                   WORK( INDERR+WBEGIN-1 ),
609      $                   WORK( INDWRK ), IWORK( IINDWK ), PIVMIN,
610      $                   TNRM, IINFO )
611             IBEGIN = IEND + 1
612             WBEGIN = WEND + 1
613  39      CONTINUE
614       ENDIF
615 *
616 *     If matrix was scaled, then rescale eigenvalues appropriately.
617 *
618       IFSCALE.NE.ONE ) THEN
619          CALL DSCAL( M, ONE / SCALE, W, 1 )
620       END IF
621 *
622 *     If eigenvalues are not in increasing order, then sort them,
623 *     possibly along with eigenvectors.
624 *
625       IF( NSPLIT.GT.1 ) THEN
626          IF.NOT. WANTZ ) THEN
627             CALL DLASRT( 'I', M, W, IINFO )
628             IF( IINFO.NE.0 ) THEN
629                INFO = 3
630                RETURN
631             END IF
632          ELSE
633             DO 60 J = 1, M - 1
634                I = 0
635                TMP = W( J )
636                DO 50 JJ = J + 1, M
637                   IF( W( JJ ).LT.TMP ) THEN
638                      I = JJ
639                      TMP = W( JJ )
640                   END IF
641  50            CONTINUE
642                IF( I.NE.0 ) THEN
643                   W( I ) = W( J )
644                   W( J ) = TMP
645                   IF( WANTZ ) THEN
646                      CALL ZSWAP( N, Z( 1, I ), 1, Z( 1, J ), 1 )
647                      ITMP = ISUPPZ( 2*I-1 )
648                      ISUPPZ( 2*I-1 ) = ISUPPZ( 2*J-1 )
649                      ISUPPZ( 2*J-1 ) = ITMP
650                      ITMP = ISUPPZ( 2*I )
651                      ISUPPZ( 2*I ) = ISUPPZ( 2*J )
652                      ISUPPZ( 2*J ) = ITMP
653                   END IF
654                END IF
655  60         CONTINUE
656          END IF
657       ENDIF
658 *
659 *
660       WORK( 1 ) = LWMIN
661       IWORK( 1 ) = LIWMIN
662       RETURN
663 *
664 *     End of ZSTEMR
665 *
666       END