1       SUBROUTINE DSTEMR( 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 3.2.2)                                  --
  7 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  8 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  9 *  -- June 2010                                                       --
 10 *
 11 *     .. Scalar Arguments ..
 12       CHARACTER          JOBZ, RANGE
 13       LOGICAL            TRYRAC
 14       INTEGER            IL, INFO, IU, LDZ, NZC, LIWORK, LWORK, M, N
 15       DOUBLE PRECISION VL, VU
 16 *     ..
 17 *     .. Array Arguments ..
 18       INTEGER            ISUPPZ( * ), IWORK( * )
 19       DOUBLE PRECISION   D( * ), E( * ), W( * ), WORK( * )
 20       DOUBLE PRECISION   Z( LDZ, * )
 21 *     ..
 22 *
 23 *  Purpose
 24 *  =======
 25 *
 26 *  DSTEMR computes selected eigenvalues and, optionally, eigenvectors
 27 *  of a real symmetric tridiagonal matrix T. Any such unreduced matrix has
 28 *  a well defined set of pairwise different real eigenvalues, the corresponding
 29 *  real eigenvectors are pairwise orthogonal.
 30 *
 31 *  The spectrum may be computed either completely or partially by specifying
 32 *  either an interval (VL,VU] or a range of indices IL:IU for the desired
 33 *  eigenvalues.
 34 *
 35 *  Depending on the number of desired eigenvalues, these are computed either
 36 *  by bisection or the dqds algorithm. Numerically orthogonal eigenvectors are
 37 *  computed by the use of various suitable L D L^T factorizations near clusters
 38 *  of close eigenvalues (referred to as RRRs, Relatively Robust
 39 *  Representations). An informal sketch of the algorithm follows.
 40 *
 41 *  For each unreduced block (submatrix) of T,
 42 *     (a) Compute T - sigma I  = L D L^T, so that L and D
 43 *         define all the wanted eigenvalues to high relative accuracy.
 44 *         This means that small relative changes in the entries of D and L
 45 *         cause only small relative changes in the eigenvalues and
 46 *         eigenvectors. The standard (unfactored) representation of the
 47 *         tridiagonal matrix T does not have this property in general.
 48 *     (b) Compute the eigenvalues to suitable accuracy.
 49 *         If the eigenvectors are desired, the algorithm attains full
 50 *         accuracy of the computed eigenvalues only right before
 51 *         the corresponding vectors have to be computed, see steps c) and d).
 52 *     (c) For each cluster of close eigenvalues, select a new
 53 *         shift close to the cluster, find a new factorization, and refine
 54 *         the shifted eigenvalues to suitable accuracy.
 55 *     (d) For each eigenvalue with a large enough relative separation compute
 56 *         the corresponding eigenvector by forming a rank revealing twisted
 57 *         factorization. Go back to (c) for any clusters that remain.
 58 *
 59 *  For more details, see:
 60 *  - Inderjit S. Dhillon and Beresford N. Parlett: "Multiple representations
 61 *    to compute orthogonal eigenvectors of symmetric tridiagonal matrices,"
 62 *    Linear Algebra and its Applications, 387(1), pp. 1-28, August 2004.
 63 *  - Inderjit Dhillon and Beresford Parlett: "Orthogonal Eigenvectors and
 64 *    Relative Gaps," SIAM Journal on Matrix Analysis and Applications, Vol. 25,
 65 *    2004.  Also LAPACK Working Note 154.
 66 *  - Inderjit Dhillon: "A new O(n^2) algorithm for the symmetric
 67 *    tridiagonal eigenvalue/eigenvector problem",
 68 *    Computer Science Division Technical Report No. UCB/CSD-97-971,
 69 *    UC Berkeley, May 1997.
 70 *
 71 *  Further Details
 72 *  1.DSTEMR works only on machines which follow IEEE-754
 73 *  floating-point standard in their handling of infinities and NaNs.
 74 *  This permits the use of efficient inner loops avoiding a check for
 75 *  zero divisors.
 76 *
 77 *  Arguments
 78 *  =========
 79 *
 80 *  JOBZ    (input) CHARACTER*1
 81 *          = 'N':  Compute eigenvalues only;
 82 *          = 'V':  Compute eigenvalues and eigenvectors.
 83 *
 84 *  RANGE   (input) CHARACTER*1
 85 *          = 'A': all eigenvalues will be found.
 86 *          = 'V': all eigenvalues in the half-open interval (VL,VU]
 87 *                 will be found.
 88 *          = 'I': the IL-th through IU-th eigenvalues will be found.
 89 *
 90 *  N       (input) INTEGER
 91 *          The order of the matrix.  N >= 0.
 92 *
 93 *  D       (input/output) DOUBLE PRECISION array, dimension (N)
 94 *          On entry, the N diagonal elements of the tridiagonal matrix
 95 *          T. On exit, D is overwritten.
 96 *
 97 *  E       (input/output) DOUBLE PRECISION array, dimension (N)
 98 *          On entry, the (N-1) subdiagonal elements of the tridiagonal
 99 *          matrix T in elements 1 to N-1 of E. E(N) need not be set on
100 *          input, but is used internally as workspace.
101 *          On exit, E is overwritten.
102 *
103 *  VL      (input) DOUBLE PRECISION
104 *  VU      (input) DOUBLE PRECISION
105 *          If RANGE='V', the lower and upper bounds of the interval to
106 *          be searched for eigenvalues. VL < VU.
107 *          Not referenced if RANGE = 'A' or 'I'.
108 *
109 *  IL      (input) INTEGER
110 *  IU      (input) INTEGER
111 *          If RANGE='I', the indices (in ascending order) of the
112 *          smallest and largest eigenvalues to be returned.
113 *          1 <= IL <= IU <= N, if N > 0.
114 *          Not referenced if RANGE = 'A' or 'V'.
115 *
116 *  M       (output) INTEGER
117 *          The total number of eigenvalues found.  0 <= M <= N.
118 *          If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
119 *
120 *  W       (output) DOUBLE PRECISION array, dimension (N)
121 *          The first M elements contain the selected eigenvalues in
122 *          ascending order.
123 *
124 *  Z       (output) DOUBLE PRECISION array, dimension (LDZ, max(1,M) )
125 *          If JOBZ = 'V', and if INFO = 0, then the first M columns of Z
126 *          contain the orthonormal eigenvectors of the matrix T
127 *          corresponding to the selected eigenvalues, with the i-th
128 *          column of Z holding the eigenvector associated with W(i).
129 *          If JOBZ = 'N', then Z is not referenced.
130 *          Note: the user must ensure that at least max(1,M) columns are
131 *          supplied in the array Z; if RANGE = 'V', the exact value of M
132 *          is not known in advance and can be computed with a workspace
133 *          query by setting NZC = -1, see below.
134 *
135 *  LDZ     (input) INTEGER
136 *          The leading dimension of the array Z.  LDZ >= 1, and if
137 *          JOBZ = 'V', then LDZ >= max(1,N).
138 *
139 *  NZC     (input) INTEGER
140 *          The number of eigenvectors to be held in the array Z.
141 *          If RANGE = 'A', then NZC >= max(1,N).
142 *          If RANGE = 'V', then NZC >= the number of eigenvalues in (VL,VU].
143 *          If RANGE = 'I', then NZC >= IU-IL+1.
144 *          If NZC = -1, then a workspace query is assumed; the
145 *          routine calculates the number of columns of the array Z that
146 *          are needed to hold the eigenvectors.
147 *          This value is returned as the first entry of the Z array, and
148 *          no error message related to NZC is issued by XERBLA.
149 *
150 *  ISUPPZ  (output) INTEGER ARRAY, dimension ( 2*max(1,M) )
151 *          The support of the eigenvectors in Z, i.e., the indices
152 *          indicating the nonzero elements in Z. The i-th computed eigenvector
153 *          is nonzero only in elements ISUPPZ( 2*i-1 ) through
154 *          ISUPPZ( 2*i ). This is relevant in the case when the matrix
155 *          is split. ISUPPZ is only accessed when JOBZ is 'V' and N > 0.
156 *
157 *  TRYRAC  (input/output) LOGICAL
158 *          If TRYRAC.EQ..TRUE., indicates that the code should check whether
159 *          the tridiagonal matrix defines its eigenvalues to high relative
160 *          accuracy.  If so, the code uses relative-accuracy preserving
161 *          algorithms that might be (a bit) slower depending on the matrix.
162 *          If the matrix does not define its eigenvalues to high relative
163 *          accuracy, the code can uses possibly faster algorithms.
164 *          If TRYRAC.EQ..FALSE., the code is not required to guarantee
165 *          relatively accurate eigenvalues and can use the fastest possible
166 *          techniques.
167 *          On exit, a .TRUE. TRYRAC will be set to .FALSE. if the matrix
168 *          does not define its eigenvalues to high relative accuracy.
169 *
170 *  WORK    (workspace/output) DOUBLE PRECISION array, dimension (LWORK)
171 *          On exit, if INFO = 0, WORK(1) returns the optimal
172 *          (and minimal) LWORK.
173 *
174 *  LWORK   (input) INTEGER
175 *          The dimension of the array WORK. LWORK >= max(1,18*N)
176 *          if JOBZ = 'V', and LWORK >= max(1,12*N) if JOBZ = 'N'.
177 *          If LWORK = -1, then a workspace query is assumed; the routine
178 *          only calculates the optimal size of the WORK array, returns
179 *          this value as the first entry of the WORK array, and no error
180 *          message related to LWORK is issued by XERBLA.
181 *
182 *  IWORK   (workspace/output) INTEGER array, dimension (LIWORK)
183 *          On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
184 *
185 *  LIWORK  (input) INTEGER
186 *          The dimension of the array IWORK.  LIWORK >= max(1,10*N)
187 *          if the eigenvectors are desired, and LIWORK >= max(1,8*N)
188 *          if only the eigenvalues are to be computed.
189 *          If LIWORK = -1, then a workspace query is assumed; the
190 *          routine only calculates the optimal size of the IWORK array,
191 *          returns this value as the first entry of the IWORK array, and
192 *          no error message related to LIWORK is issued by XERBLA.
193 *
194 *  INFO    (output) INTEGER
195 *          On exit, INFO
196 *          = 0:  successful exit
197 *          < 0:  if INFO = -i, the i-th argument had an illegal value
198 *          > 0:  if INFO = 1X, internal error in DLARRE,
199 *                if INFO = 2X, internal error in DLARRV.
200 *                Here, the digit X = ABS( IINFO ) < 10, where IINFO is
201 *                the nonzero error code returned by DLARRE or
202 *                DLARRV, respectively.
203 *
204 *
205 *  Further Details
206 *  ===============
207 *
208 *  Based on contributions by
209 *     Beresford Parlett, University of California, Berkeley, USA
210 *     Jim Demmel, University of California, Berkeley, USA
211 *     Inderjit Dhillon, University of Texas, Austin, USA
212 *     Osni Marques, LBNL/NERSC, USA
213 *     Christof Voemel, University of California, Berkeley, USA
214 *
215 *  =====================================================================
216 *
217 *     .. Parameters ..
218       DOUBLE PRECISION   ZERO, ONE, FOUR, MINRGP
219       PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0,
220      $                     FOUR = 4.0D0,
221      $                     MINRGP = 1.0D-3 )
222 *     ..
223 *     .. Local Scalars ..
224       LOGICAL            ALLEIG, INDEIG, LQUERY, VALEIG, WANTZ, ZQUERY
225       INTEGER            I, IBEGIN, IEND, IFIRST, IIL, IINDBL, IINDW,
226      $                   IINDWK, IINFO, IINSPL, IIU, ILAST, IN, INDD,
227      $                   INDE2, INDERR, INDGP, INDGRS, INDWRK, ITMP,
228      $                   ITMP2, J, JBLK, JJ, LIWMIN, LWMIN, NSPLIT,
229      $                   NZCMIN, OFFSET, WBEGIN, WEND
230       DOUBLE PRECISION   BIGNUM, CS, EPS, PIVMIN, R1, R2, RMAX, RMIN,
231      $                   RTOL1, RTOL2, SAFMIN, SCALE, SMLNUM, SN,
232      $                   THRESH, TMP, TNRM, WL, WU
233 *     ..
234 *     ..
235 *     .. External Functions ..
236       LOGICAL            LSAME
237       DOUBLE PRECISION   DLAMCH, DLANST
238       EXTERNAL           LSAME, DLAMCH, DLANST
239 *     ..
240 *     .. External Subroutines ..
241       EXTERNAL           DCOPY, DLAE2, DLAEV2, DLARRC, DLARRE, DLARRJ,
242      $                   DLARRR, DLARRV, DLASRT, DSCAL, DSWAP, XERBLA
243 *     ..
244 *     .. Intrinsic Functions ..
245       INTRINSIC          MAXMINSQRT
246 
247 
248 *     ..
249 *     .. Executable Statements ..
250 *
251 *     Test the input parameters.
252 *
253       WANTZ = LSAME( JOBZ, 'V' )
254       ALLEIG = LSAME( RANGE'A' )
255       VALEIG = LSAME( RANGE'V' )
256       INDEIG = LSAME( RANGE'I' )
257 *
258       LQUERY = ( ( LWORK.EQ.-1 ).OR.( LIWORK.EQ.-1 ) )
259       ZQUERY = ( NZC.EQ.-1 )
260 
261 *     DSTEMR needs WORK of size 6*N, IWORK of size 3*N.
262 *     In addition, DLARRE needs WORK of size 6*N, IWORK of size 5*N.
263 *     Furthermore, DLARRV needs WORK of size 12*N, IWORK of size 7*N.
264       IF( WANTZ ) THEN
265          LWMIN = 18*N
266          LIWMIN = 10*N
267       ELSE
268 *        need less workspace if only the eigenvalues are wanted
269          LWMIN = 12*N
270          LIWMIN = 8*N
271       ENDIF
272 
273       WL = ZERO
274       WU = ZERO
275       IIL = 0
276       IIU = 0
277 
278       IF( VALEIG ) THEN
279 *        We do not reference VL, VU in the cases RANGE = 'I','A'
280 *        The interval (WL, WU] contains all the wanted eigenvalues.
281 *        It is either given by the user or computed in DLARRE.
282          WL = VL
283          WU = VU
284       ELSEIF( INDEIG ) THEN
285 *        We do not reference IL, IU in the cases RANGE = 'V','A'
286          IIL = IL
287          IIU = IU
288       ENDIF
289 *
290       INFO = 0
291       IF.NOT.( WANTZ .OR. LSAME( JOBZ, 'N' ) ) ) THEN
292          INFO = -1
293       ELSE IF.NOT.( ALLEIG .OR. VALEIG .OR. INDEIG ) ) THEN
294          INFO = -2
295       ELSE IF( N.LT.0 ) THEN
296          INFO = -3
297       ELSE IF( VALEIG .AND. N.GT.0 .AND. WU.LE.WL ) THEN
298          INFO = -7
299       ELSE IF( INDEIG .AND. ( IIL.LT.1 .OR. IIL.GT.N ) ) THEN
300          INFO = -8
301       ELSE IF( INDEIG .AND. ( IIU.LT.IIL .OR. IIU.GT.N ) ) THEN
302          INFO = -9
303       ELSE IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
304          INFO = -13
305       ELSE IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
306          INFO = -17
307       ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
308          INFO = -19
309       END IF
310 *
311 *     Get machine constants.
312 *
313       SAFMIN = DLAMCH( 'Safe minimum' )
314       EPS = DLAMCH( 'Precision' )
315       SMLNUM = SAFMIN / EPS
316       BIGNUM = ONE / SMLNUM
317       RMIN = SQRT( SMLNUM )
318       RMAX = MINSQRT( BIGNUM ), ONE / SQRTSQRT( SAFMIN ) ) )
319 *
320       IF( INFO.EQ.0 ) THEN
321          WORK( 1 ) = LWMIN
322          IWORK( 1 ) = LIWMIN
323 *
324          IF( WANTZ .AND. ALLEIG ) THEN
325             NZCMIN = N
326          ELSE IF( WANTZ .AND. VALEIG ) THEN
327             CALL DLARRC( 'T', N, VL, VU, D, E, SAFMIN,
328      $                            NZCMIN, ITMP, ITMP2, INFO )
329          ELSE IF( WANTZ .AND. INDEIG ) THEN
330             NZCMIN = IIU-IIL+1
331          ELSE
332 *           WANTZ .EQ. FALSE.
333             NZCMIN = 0
334          ENDIF
335          IF( ZQUERY .AND. INFO.EQ.0 ) THEN
336             Z( 1,1 ) = NZCMIN
337          ELSE IF( NZC.LT.NZCMIN .AND. .NOT.ZQUERY ) THEN
338             INFO = -14
339          END IF
340       END IF
341 
342       IF( INFO.NE.0 ) THEN
343 *
344          CALL XERBLA( 'DSTEMR'-INFO )
345 *
346          RETURN
347       ELSE IF( LQUERY .OR. ZQUERY ) THEN
348          RETURN
349       END IF
350 *
351 *     Handle N = 0, 1, and 2 cases immediately
352 *
353       M = 0
354       IF( N.EQ.0 )
355      $   RETURN
356 *
357       IF( N.EQ.1 ) THEN
358          IF( ALLEIG .OR. INDEIG ) THEN
359             M = 1
360             W( 1 ) = D( 1 )
361          ELSE
362             IF( WL.LT.D( 1 ) .AND. WU.GE.D( 1 ) ) THEN
363                M = 1
364                W( 1 ) = D( 1 )
365             END IF
366          END IF
367          IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
368             Z( 11 ) = ONE
369             ISUPPZ(1= 1
370             ISUPPZ(2= 1
371          END IF
372          RETURN
373       END IF
374 *
375       IF( N.EQ.2 ) THEN
376          IF.NOT.WANTZ ) THEN
377             CALL DLAE2( D(1), E(1), D(2), R1, R2 )
378          ELSE IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
379             CALL DLAEV2( D(1), E(1), D(2), R1, R2, CS, SN )
380          END IF
381          IF( ALLEIG.OR.
382      $      (VALEIG.AND.(R2.GT.WL).AND.
383      $                  (R2.LE.WU)).OR.
384      $      (INDEIG.AND.(IIL.EQ.1)) ) THEN
385             M = M+1
386             W( M ) = R2
387             IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
388                Z( 1, M ) = -SN
389                Z( 2, M ) = CS
390 *              Note: At most one of SN and CS can be zero.
391                IF (SN.NE.ZERO) THEN
392                   IF (CS.NE.ZERO) THEN
393                      ISUPPZ(2*M-1= 1
394                      ISUPPZ(2*M) = 2
395                   ELSE
396                      ISUPPZ(2*M-1= 1
397                      ISUPPZ(2*M) = 1
398                   END IF
399                ELSE
400                   ISUPPZ(2*M-1= 2
401                   ISUPPZ(2*M) = 2
402                END IF
403             ENDIF
404          ENDIF
405          IF( ALLEIG.OR.
406      $      (VALEIG.AND.(R1.GT.WL).AND.
407      $                  (R1.LE.WU)).OR.
408      $      (INDEIG.AND.(IIU.EQ.2)) ) THEN
409             M = M+1
410             W( M ) = R1
411             IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
412                Z( 1, M ) = CS
413                Z( 2, M ) = SN
414 *              Note: At most one of SN and CS can be zero.
415                IF (SN.NE.ZERO) THEN
416                   IF (CS.NE.ZERO) THEN
417                      ISUPPZ(2*M-1= 1
418                      ISUPPZ(2*M) = 2
419                   ELSE
420                      ISUPPZ(2*M-1= 1
421                      ISUPPZ(2*M) = 1
422                   END IF
423                ELSE
424                   ISUPPZ(2*M-1= 2
425                   ISUPPZ(2*M) = 2
426                END IF
427             ENDIF
428          ENDIF
429          RETURN
430       END IF
431 
432 *     Continue with general N
433 
434       INDGRS = 1
435       INDERR = 2*+ 1
436       INDGP = 3*+ 1
437       INDD = 4*+ 1
438       INDE2 = 5*+ 1
439       INDWRK = 6*+ 1
440 *
441       IINSPL = 1
442       IINDBL = N + 1
443       IINDW = 2*+ 1
444       IINDWK = 3*+ 1
445 *
446 *     Scale matrix to allowable range, if necessary.
447 *     The allowable range is related to the PIVMIN parameter; see the
448 *     comments in DLARRD.  The preference for scaling small values
449 *     up is heuristic; we expect users' matrices not to be close to the
450 *     RMAX threshold.
451 *
452       SCALE = ONE
453       TNRM = DLANST( 'M', N, D, E )
454       IF( TNRM.GT.ZERO .AND. TNRM.LT.RMIN ) THEN
455          SCALE = RMIN / TNRM
456       ELSE IF( TNRM.GT.RMAX ) THEN
457          SCALE = RMAX / TNRM
458       END IF
459       IFSCALE.NE.ONE ) THEN
460          CALL DSCAL( N, SCALE, D, 1 )
461          CALL DSCAL( N-1SCALE, E, 1 )
462          TNRM = TNRM*SCALE
463          IF( VALEIG ) THEN
464 *           If eigenvalues in interval have to be found,
465 *           scale (WL, WU] accordingly
466             WL = WL*SCALE
467             WU = WU*SCALE
468          ENDIF
469       END IF
470 *
471 *     Compute the desired eigenvalues of the tridiagonal after splitting
472 *     into smaller subblocks if the corresponding off-diagonal elements
473 *     are small
474 *     THRESH is the splitting parameter for DLARRE
475 *     A negative THRESH forces the old splitting criterion based on the
476 *     size of the off-diagonal. A positive THRESH switches to splitting
477 *     which preserves relative accuracy.
478 *
479       IF( TRYRAC ) THEN
480 *        Test whether the matrix warrants the more expensive relative approach.
481          CALL DLARRR( N, D, E, IINFO )
482       ELSE
483 *        The user does not care about relative accurately eigenvalues
484          IINFO = -1
485       ENDIF
486 *     Set the splitting criterion
487       IF (IINFO.EQ.0THEN
488          THRESH = EPS
489       ELSE
490          THRESH = -EPS
491 *        relative accuracy is desired but T does not guarantee it
492          TRYRAC = .FALSE.
493       ENDIF
494 *
495       IF( TRYRAC ) THEN
496 *        Copy original diagonal, needed to guarantee relative accuracy
497          CALL DCOPY(N,D,1,WORK(INDD),1)
498       ENDIF
499 *     Store the squares of the offdiagonal values of T
500       DO 5 J = 1, N-1
501          WORK( INDE2+J-1 ) = E(J)**2
502  5    CONTINUE
503 
504 *     Set the tolerance parameters for bisection
505       IF.NOT.WANTZ ) THEN
506 *        DLARRE computes the eigenvalues to full precision.
507          RTOL1 = FOUR * EPS
508          RTOL2 = FOUR * EPS
509       ELSE
510 *        DLARRE computes the eigenvalues to less than full precision.
511 *        DLARRV will refine the eigenvalue approximations, and we can
512 *        need less accurate initial bisection in DLARRE.
513 *        Note: these settings do only affect the subset case and DLARRE
514          RTOL1 = SQRT(EPS)
515          RTOL2 = MAXSQRT(EPS)*5.0D-3, FOUR * EPS )
516       ENDIF
517       CALL DLARRE( RANGE, N, WL, WU, IIL, IIU, D, E,
518      $             WORK(INDE2), RTOL1, RTOL2, THRESH, NSPLIT,
519      $             IWORK( IINSPL ), M, W, WORK( INDERR ),
520      $             WORK( INDGP ), IWORK( IINDBL ),
521      $             IWORK( IINDW ), WORK( INDGRS ), PIVMIN,
522      $             WORK( INDWRK ), IWORK( IINDWK ), IINFO )
523       IF( IINFO.NE.0 ) THEN
524          INFO = 10 + ABS( IINFO )
525          RETURN
526       END IF
527 *     Note that if RANGE .NE. 'V', DLARRE computes bounds on the desired
528 *     part of the spectrum. All desired eigenvalues are contained in
529 *     (WL,WU]
530 
531 
532       IF( WANTZ ) THEN
533 *
534 *        Compute the desired eigenvectors corresponding to the computed
535 *        eigenvalues
536 *
537          CALL DLARRV( N, WL, WU, D, E,
538      $                PIVMIN, IWORK( IINSPL ), M,
539      $                1, M, MINRGP, RTOL1, RTOL2,
540      $                W, WORK( INDERR ), WORK( INDGP ), IWORK( IINDBL ),
541      $                IWORK( IINDW ), WORK( INDGRS ), Z, LDZ,
542      $                ISUPPZ, WORK( INDWRK ), IWORK( IINDWK ), IINFO )
543          IF( IINFO.NE.0 ) THEN
544             INFO = 20 + ABS( IINFO )
545             RETURN
546          END IF
547       ELSE
548 *        DLARRE computes eigenvalues of the (shifted) root representation
549 *        DLARRV returns the eigenvalues of the unshifted matrix.
550 *        However, if the eigenvectors are not desired by the user, we need
551 *        to apply the corresponding shifts from DLARRE to obtain the
552 *        eigenvalues of the original matrix.
553          DO 20 J = 1, M
554             ITMP = IWORK( IINDBL+J-1 )
555             W( J ) = W( J ) + E( IWORK( IINSPL+ITMP-1 ) )
556  20      CONTINUE
557       END IF
558 *
559 
560       IF ( TRYRAC ) THEN
561 *        Refine computed eigenvalues so that they are relatively accurate
562 *        with respect to the original matrix T.
563          IBEGIN = 1
564          WBEGIN = 1
565          DO 39  JBLK = 1, IWORK( IINDBL+M-1 )
566             IEND = IWORK( IINSPL+JBLK-1 )
567             IN = IEND - IBEGIN + 1
568             WEND = WBEGIN - 1
569 *           check if any eigenvalues have to be refined in this block
570  36         CONTINUE
571             IF( WEND.LT.M ) THEN
572                IF( IWORK( IINDBL+WEND ).EQ.JBLK ) THEN
573                   WEND = WEND + 1
574                   GO TO 36
575                END IF
576             END IF
577             IF( WEND.LT.WBEGIN ) THEN
578                IBEGIN = IEND + 1
579                GO TO 39
580             END IF
581 
582             OFFSET = IWORK(IINDW+WBEGIN-1)-1
583             IFIRST = IWORK(IINDW+WBEGIN-1)
584             ILAST = IWORK(IINDW+WEND-1)
585             RTOL2 = FOUR * EPS
586             CALL DLARRJ( IN,
587      $                   WORK(INDD+IBEGIN-1), WORK(INDE2+IBEGIN-1),
588      $                   IFIRST, ILAST, RTOL2, OFFSET, W(WBEGIN),
589      $                   WORK( INDERR+WBEGIN-1 ),
590      $                   WORK( INDWRK ), IWORK( IINDWK ), PIVMIN,
591      $                   TNRM, IINFO )
592             IBEGIN = IEND + 1
593             WBEGIN = WEND + 1
594  39      CONTINUE
595       ENDIF
596 *
597 *     If matrix was scaled, then rescale eigenvalues appropriately.
598 *
599       IFSCALE.NE.ONE ) THEN
600          CALL DSCAL( M, ONE / SCALE, W, 1 )
601       END IF
602 *
603 *     If eigenvalues are not in increasing order, then sort them,
604 *     possibly along with eigenvectors.
605 *
606       IF( NSPLIT.GT.1 ) THEN
607          IF.NOT. WANTZ ) THEN
608             CALL DLASRT( 'I', M, W, IINFO )
609             IF( IINFO.NE.0 ) THEN
610                INFO = 3
611                RETURN
612             END IF
613          ELSE
614             DO 60 J = 1, M - 1
615                I = 0
616                TMP = W( J )
617                DO 50 JJ = J + 1, M
618                   IF( W( JJ ).LT.TMP ) THEN
619                      I = JJ
620                      TMP = W( JJ )
621                   END IF
622  50            CONTINUE
623                IF( I.NE.0 ) THEN
624                   W( I ) = W( J )
625                   W( J ) = TMP
626                   IF( WANTZ ) THEN
627                      CALL DSWAP( N, Z( 1, I ), 1, Z( 1, J ), 1 )
628                      ITMP = ISUPPZ( 2*I-1 )
629                      ISUPPZ( 2*I-1 ) = ISUPPZ( 2*J-1 )
630                      ISUPPZ( 2*J-1 ) = ITMP
631                      ITMP = ISUPPZ( 2*I )
632                      ISUPPZ( 2*I ) = ISUPPZ( 2*J )
633                      ISUPPZ( 2*J ) = ITMP
634                   END IF
635                END IF
636  60         CONTINUE
637          END IF
638       ENDIF
639 *
640 *
641       WORK( 1 ) = LWMIN
642       IWORK( 1 ) = LIWMIN
643       RETURN
644 *
645 *     End of DSTEMR
646 *
647       END