1       SUBROUTINE ZLAED0( QSIZ, N, D, E, Q, LDQ, QSTORE, LDQS, RWORK,
  2      $                   IWORK, INFO )
  3 *
  4 *  -- LAPACK routine (version 3.2) --
  5 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  6 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  7 *     November 2006
  8 *
  9 *     .. Scalar Arguments ..
 10       INTEGER            INFO, LDQ, LDQS, N, QSIZ
 11 *     ..
 12 *     .. Array Arguments ..
 13       INTEGER            IWORK( * )
 14       DOUBLE PRECISION   D( * ), E( * ), RWORK( * )
 15       COMPLEX*16         Q( LDQ, * ), QSTORE( LDQS, * )
 16 *     ..
 17 *
 18 *  Purpose
 19 *  =======
 20 *
 21 *  Using the divide and conquer method, ZLAED0 computes all eigenvalues
 22 *  of a symmetric tridiagonal matrix which is one diagonal block of
 23 *  those from reducing a dense or band Hermitian matrix and
 24 *  corresponding eigenvectors of the dense or band matrix.
 25 *
 26 *  Arguments
 27 *  =========
 28 *
 29 *  QSIZ   (input) INTEGER
 30 *         The dimension of the unitary matrix used to reduce
 31 *         the full matrix to tridiagonal form.  QSIZ >= N if ICOMPQ = 1.
 32 *
 33 *  N      (input) INTEGER
 34 *         The dimension of the symmetric tridiagonal matrix.  N >= 0.
 35 *
 36 *  D      (input/output) DOUBLE PRECISION array, dimension (N)
 37 *         On entry, the diagonal elements of the tridiagonal matrix.
 38 *         On exit, the eigenvalues in ascending order.
 39 *
 40 *  E      (input/output) DOUBLE PRECISION array, dimension (N-1)
 41 *         On entry, the off-diagonal elements of the tridiagonal matrix.
 42 *         On exit, E has been destroyed.
 43 *
 44 *  Q      (input/output) COMPLEX*16 array, dimension (LDQ,N)
 45 *         On entry, Q must contain an QSIZ x N matrix whose columns
 46 *         unitarily orthonormal. It is a part of the unitary matrix
 47 *         that reduces the full dense Hermitian matrix to a
 48 *         (reducible) symmetric tridiagonal matrix.
 49 *
 50 *  LDQ    (input) INTEGER
 51 *         The leading dimension of the array Q.  LDQ >= max(1,N).
 52 *
 53 *  IWORK  (workspace) INTEGER array,
 54 *         the dimension of IWORK must be at least
 55 *                      6 + 6*N + 5*N*lg N
 56 *                      ( lg( N ) = smallest integer k
 57 *                                  such that 2^k >= N )
 58 *
 59 *  RWORK  (workspace) DOUBLE PRECISION array,
 60 *                               dimension (1 + 3*N + 2*N*lg N + 3*N**2)
 61 *                        ( lg( N ) = smallest integer k
 62 *                                    such that 2^k >= N )
 63 *
 64 *  QSTORE (workspace) COMPLEX*16 array, dimension (LDQS, N)
 65 *         Used to store parts of
 66 *         the eigenvector matrix when the updating matrix multiplies
 67 *         take place.
 68 *
 69 *  LDQS   (input) INTEGER
 70 *         The leading dimension of the array QSTORE.
 71 *         LDQS >= max(1,N).
 72 *
 73 *  INFO   (output) INTEGER
 74 *          = 0:  successful exit.
 75 *          < 0:  if INFO = -i, the i-th argument had an illegal value.
 76 *          > 0:  The algorithm failed to compute an eigenvalue while
 77 *                working on the submatrix lying in rows and columns
 78 *                INFO/(N+1) through mod(INFO,N+1).
 79 *
 80 *  =====================================================================
 81 *
 82 *  Warning:      N could be as big as QSIZ!
 83 *
 84 *     .. Parameters ..
 85       DOUBLE PRECISION   TWO
 86       PARAMETER          ( TWO = 2.D+0 )
 87 *     ..
 88 *     .. Local Scalars ..
 89       INTEGER            CURLVL, CURPRB, CURR, I, IGIVCL, IGIVNM,
 90      $                   IGIVPT, INDXQ, IPERM, IPRMPT, IQ, IQPTR, IWREM,
 91      $                   J, K, LGN, LL, MATSIZ, MSD2, SMLSIZ, SMM1,
 92      $                   SPM1, SPM2, SUBMAT, SUBPBS, TLVLS
 93       DOUBLE PRECISION   TEMP
 94 *     ..
 95 *     .. External Subroutines ..
 96       EXTERNAL           DCOPY, DSTEQR, XERBLA, ZCOPY, ZLACRM, ZLAED7
 97 *     ..
 98 *     .. External Functions ..
 99       INTEGER            ILAENV
100       EXTERNAL           ILAENV
101 *     ..
102 *     .. Intrinsic Functions ..
103       INTRINSIC          ABSDBLEINTLOGMAX
104 *     ..
105 *     .. Executable Statements ..
106 *
107 *     Test the input parameters.
108 *
109       INFO = 0
110 *
111 *     IF( ICOMPQ .LT. 0 .OR. ICOMPQ .GT. 2 ) THEN
112 *        INFO = -1
113 *     ELSE IF( ( ICOMPQ .EQ. 1 ) .AND. ( QSIZ .LT. MAX( 0, N ) ) )
114 *    $        THEN
115       IF( QSIZ.LT.MAX0, N ) ) THEN
116          INFO = -1
117       ELSE IF( N.LT.0 ) THEN
118          INFO = -2
119       ELSE IF( LDQ.LT.MAX1, N ) ) THEN
120          INFO = -6
121       ELSE IF( LDQS.LT.MAX1, N ) ) THEN
122          INFO = -8
123       END IF
124       IF( INFO.NE.0 ) THEN
125          CALL XERBLA( 'ZLAED0'-INFO )
126          RETURN
127       END IF
128 *
129 *     Quick return if possible
130 *
131       IF( N.EQ.0 )
132      $   RETURN
133 *
134       SMLSIZ = ILAENV( 9'ZLAED0'' '0000 )
135 *
136 *     Determine the size and placement of the submatrices, and save in
137 *     the leading elements of IWORK.
138 *
139       IWORK( 1 ) = N
140       SUBPBS = 1
141       TLVLS = 0
142    10 CONTINUE
143       IF( IWORK( SUBPBS ).GT.SMLSIZ ) THEN
144          DO 20 J = SUBPBS, 1-1
145             IWORK( 2*J ) = ( IWORK( J )+1 ) / 2
146             IWORK( 2*J-1 ) = IWORK( J ) / 2
147    20    CONTINUE
148          TLVLS = TLVLS + 1
149          SUBPBS = 2*SUBPBS
150          GO TO 10
151       END IF
152       DO 30 J = 2, SUBPBS
153          IWORK( J ) = IWORK( J ) + IWORK( J-1 )
154    30 CONTINUE
155 *
156 *     Divide the matrix into SUBPBS submatrices of size at most SMLSIZ+1
157 *     using rank-1 modifications (cuts).
158 *
159       SPM1 = SUBPBS - 1
160       DO 40 I = 1, SPM1
161          SUBMAT = IWORK( I ) + 1
162          SMM1 = SUBMAT - 1
163          D( SMM1 ) = D( SMM1 ) - ABS( E( SMM1 ) )
164          D( SUBMAT ) = D( SUBMAT ) - ABS( E( SMM1 ) )
165    40 CONTINUE
166 *
167       INDXQ = 4*+ 3
168 *
169 *     Set up workspaces for eigenvalues only/accumulate new vectors
170 *     routine
171 *
172       TEMP = LOGDBLE( N ) ) / LOG( TWO )
173       LGN = INT( TEMP )
174       IF2**LGN.LT.N )
175      $   LGN = LGN + 1
176       IF2**LGN.LT.N )
177      $   LGN = LGN + 1
178       IPRMPT = INDXQ + N + 1
179       IPERM = IPRMPT + N*LGN
180       IQPTR = IPERM + N*LGN
181       IGIVPT = IQPTR + N + 2
182       IGIVCL = IGIVPT + N*LGN
183 *
184       IGIVNM = 1
185       IQ = IGIVNM + 2*N*LGN
186       IWREM = IQ + N**2 + 1
187 *     Initialize pointers
188       DO 50 I = 0, SUBPBS
189          IWORK( IPRMPT+I ) = 1
190          IWORK( IGIVPT+I ) = 1
191    50 CONTINUE
192       IWORK( IQPTR ) = 1
193 *
194 *     Solve each submatrix eigenproblem at the bottom of the divide and
195 *     conquer tree.
196 *
197       CURR = 0
198       DO 70 I = 0, SPM1
199          IF( I.EQ.0 ) THEN
200             SUBMAT = 1
201             MATSIZ = IWORK( 1 )
202          ELSE
203             SUBMAT = IWORK( I ) + 1
204             MATSIZ = IWORK( I+1 ) - IWORK( I )
205          END IF
206          LL = IQ - 1 + IWORK( IQPTR+CURR )
207          CALL DSTEQR( 'I', MATSIZ, D( SUBMAT ), E( SUBMAT ),
208      $                RWORK( LL ), MATSIZ, RWORK, INFO )
209          CALL ZLACRM( QSIZ, MATSIZ, Q( 1, SUBMAT ), LDQ, RWORK( LL ),
210      $                MATSIZ, QSTORE( 1, SUBMAT ), LDQS,
211      $                RWORK( IWREM ) )
212          IWORK( IQPTR+CURR+1 ) = IWORK( IQPTR+CURR ) + MATSIZ**2
213          CURR = CURR + 1
214          IF( INFO.GT.0 ) THEN
215             INFO = SUBMAT*( N+1 ) + SUBMAT + MATSIZ - 1
216             RETURN
217          END IF
218          K = 1
219          DO 60 J = SUBMAT, IWORK( I+1 )
220             IWORK( INDXQ+J ) = K
221             K = K + 1
222    60    CONTINUE
223    70 CONTINUE
224 *
225 *     Successively merge eigensystems of adjacent submatrices
226 *     into eigensystem for the corresponding larger matrix.
227 *
228 *     while ( SUBPBS > 1 )
229 *
230       CURLVL = 1
231    80 CONTINUE
232       IF( SUBPBS.GT.1 ) THEN
233          SPM2 = SUBPBS - 2
234          DO 90 I = 0, SPM2, 2
235             IF( I.EQ.0 ) THEN
236                SUBMAT = 1
237                MATSIZ = IWORK( 2 )
238                MSD2 = IWORK( 1 )
239                CURPRB = 0
240             ELSE
241                SUBMAT = IWORK( I ) + 1
242                MATSIZ = IWORK( I+2 ) - IWORK( I )
243                MSD2 = MATSIZ / 2
244                CURPRB = CURPRB + 1
245             END IF
246 *
247 *     Merge lower order eigensystems (of size MSD2 and MATSIZ - MSD2)
248 *     into an eigensystem of size MATSIZ.  ZLAED7 handles the case
249 *     when the eigenvectors of a full or band Hermitian matrix (which
250 *     was reduced to tridiagonal form) are desired.
251 *
252 *     I am free to use Q as a valuable working space until Loop 150.
253 *
254             CALL ZLAED7( MATSIZ, MSD2, QSIZ, TLVLS, CURLVL, CURPRB,
255      $                   D( SUBMAT ), QSTORE( 1, SUBMAT ), LDQS,
256      $                   E( SUBMAT+MSD2-1 ), IWORK( INDXQ+SUBMAT ),
257      $                   RWORK( IQ ), IWORK( IQPTR ), IWORK( IPRMPT ),
258      $                   IWORK( IPERM ), IWORK( IGIVPT ),
259      $                   IWORK( IGIVCL ), RWORK( IGIVNM ),
260      $                   Q( 1, SUBMAT ), RWORK( IWREM ),
261      $                   IWORK( SUBPBS+1 ), INFO )
262             IF( INFO.GT.0 ) THEN
263                INFO = SUBMAT*( N+1 ) + SUBMAT + MATSIZ - 1
264                RETURN
265             END IF
266             IWORK( I / 2+1 ) = IWORK( I+2 )
267    90    CONTINUE
268          SUBPBS = SUBPBS / 2
269          CURLVL = CURLVL + 1
270          GO TO 80
271       END IF
272 *
273 *     end while
274 *
275 *     Re-merge the eigenvalues/vectors which were deflated at the final
276 *     merge step.
277 *
278       DO 100 I = 1, N
279          J = IWORK( INDXQ+I )
280          RWORK( I ) = D( J )
281          CALL ZCOPY( QSIZ, QSTORE( 1, J ), 1, Q( 1, I ), 1 )
282   100 CONTINUE
283       CALL DCOPY( N, RWORK, 1, D, 1 )
284 *
285       RETURN
286 *
287 *     End of ZLAED0
288 *
289       END