1       SUBROUTINE DLAED3( K, N, N1, D, Q, LDQ, RHO, DLAMDA, Q2, INDX,
  2      $                   CTOT, W, S, 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, K, LDQ, N, N1
 11       DOUBLE PRECISION   RHO
 12 *     ..
 13 *     .. Array Arguments ..
 14       INTEGER            CTOT( * ), INDX( * )
 15       DOUBLE PRECISION   D( * ), DLAMDA( * ), Q( LDQ, * ), Q2( * ),
 16      $                   S( * ), W( * )
 17 *     ..
 18 *
 19 *  Purpose
 20 *  =======
 21 *
 22 *  DLAED3 finds the roots of the secular equation, as defined by the
 23 *  values in D, W, and RHO, between 1 and K.  It makes the
 24 *  appropriate calls to DLAED4 and then updates the eigenvectors by
 25 *  multiplying the matrix of eigenvectors of the pair of eigensystems
 26 *  being combined by the matrix of eigenvectors of the K-by-K system
 27 *  which is solved here.
 28 *
 29 *  This code makes very mild assumptions about floating point
 30 *  arithmetic. It will work on machines with a guard digit in
 31 *  add/subtract, or on those binary machines without guard digits
 32 *  which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
 33 *  It could conceivably fail on hexadecimal or decimal machines
 34 *  without guard digits, but we know of none.
 35 *
 36 *  Arguments
 37 *  =========
 38 *
 39 *  K       (input) INTEGER
 40 *          The number of terms in the rational function to be solved by
 41 *          DLAED4.  K >= 0.
 42 *
 43 *  N       (input) INTEGER
 44 *          The number of rows and columns in the Q matrix.
 45 *          N >= K (deflation may result in N>K).
 46 *
 47 *  N1      (input) INTEGER
 48 *          The location of the last eigenvalue in the leading submatrix.
 49 *          min(1,N) <= N1 <= N/2.
 50 *
 51 *  D       (output) DOUBLE PRECISION array, dimension (N)
 52 *          D(I) contains the updated eigenvalues for
 53 *          1 <= I <= K.
 54 *
 55 *  Q       (output) DOUBLE PRECISION array, dimension (LDQ,N)
 56 *          Initially the first K columns are used as workspace.
 57 *          On output the columns 1 to K contain
 58 *          the updated eigenvectors.
 59 *
 60 *  LDQ     (input) INTEGER
 61 *          The leading dimension of the array Q.  LDQ >= max(1,N).
 62 *
 63 *  RHO     (input) DOUBLE PRECISION
 64 *          The value of the parameter in the rank one update equation.
 65 *          RHO >= 0 required.
 66 *
 67 *  DLAMDA  (input/output) DOUBLE PRECISION array, dimension (K)
 68 *          The first K elements of this array contain the old roots
 69 *          of the deflated updating problem.  These are the poles
 70 *          of the secular equation. May be changed on output by
 71 *          having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
 72 *          Cray-2, or Cray C-90, as described above.
 73 *
 74 *  Q2      (input) DOUBLE PRECISION array, dimension (LDQ2, N)
 75 *          The first K columns of this matrix contain the non-deflated
 76 *          eigenvectors for the split problem.
 77 *
 78 *  INDX    (input) INTEGER array, dimension (N)
 79 *          The permutation used to arrange the columns of the deflated
 80 *          Q matrix into three groups (see DLAED2).
 81 *          The rows of the eigenvectors found by DLAED4 must be likewise
 82 *          permuted before the matrix multiply can take place.
 83 *
 84 *  CTOT    (input) INTEGER array, dimension (4)
 85 *          A count of the total number of the various types of columns
 86 *          in Q, as described in INDX.  The fourth column type is any
 87 *          column which has been deflated.
 88 *
 89 *  W       (input/output) DOUBLE PRECISION array, dimension (K)
 90 *          The first K elements of this array contain the components
 91 *          of the deflation-adjusted updating vector. Destroyed on
 92 *          output.
 93 *
 94 *  S       (workspace) DOUBLE PRECISION array, dimension (N1 + 1)*K
 95 *          Will contain the eigenvectors of the repaired matrix which
 96 *          will be multiplied by the previously accumulated eigenvectors
 97 *          to update the system.
 98 *
 99 *  LDS     (input) INTEGER
100 *          The leading dimension of S.  LDS >= max(1,K).
101 *
102 *  INFO    (output) INTEGER
103 *          = 0:  successful exit.
104 *          < 0:  if INFO = -i, the i-th argument had an illegal value.
105 *          > 0:  if INFO = 1, an eigenvalue did not converge
106 *
107 *  Further Details
108 *  ===============
109 *
110 *  Based on contributions by
111 *     Jeff Rutter, Computer Science Division, University of California
112 *     at Berkeley, USA
113 *  Modified by Francoise Tisseur, University of Tennessee.
114 *
115 *  =====================================================================
116 *
117 *     .. Parameters ..
118       DOUBLE PRECISION   ONE, ZERO
119       PARAMETER          ( ONE = 1.0D0, ZERO = 0.0D0 )
120 *     ..
121 *     .. Local Scalars ..
122       INTEGER            I, II, IQ2, J, N12, N2, N23
123       DOUBLE PRECISION   TEMP
124 *     ..
125 *     .. External Functions ..
126       DOUBLE PRECISION   DLAMC3, DNRM2
127       EXTERNAL           DLAMC3, DNRM2
128 *     ..
129 *     .. External Subroutines ..
130       EXTERNAL           DCOPY, DGEMM, DLACPY, DLAED4, DLASET, XERBLA
131 *     ..
132 *     .. Intrinsic Functions ..
133       INTRINSIC          MAXSIGNSQRT
134 *     ..
135 *     .. Executable Statements ..
136 *
137 *     Test the input parameters.
138 *
139       INFO = 0
140 *
141       IF( K.LT.0 ) THEN
142          INFO = -1
143       ELSE IF( N.LT.K ) THEN
144          INFO = -2
145       ELSE IF( LDQ.LT.MAX1, N ) ) THEN
146          INFO = -6
147       END IF
148       IF( INFO.NE.0 ) THEN
149          CALL XERBLA( 'DLAED3'-INFO )
150          RETURN
151       END IF
152 *
153 *     Quick return if possible
154 *
155       IF( K.EQ.0 )
156      $   RETURN
157 *
158 *     Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
159 *     be computed with high relative accuracy (barring over/underflow).
160 *     This is a problem on machines without a guard digit in
161 *     add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
162 *     The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
163 *     which on any of these machines zeros out the bottommost
164 *     bit of DLAMDA(I) if it is 1; this makes the subsequent
165 *     subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
166 *     occurs. On binary machines with a guard digit (almost all
167 *     machines) it does not change DLAMDA(I) at all. On hexadecimal
168 *     and decimal machines with a guard digit, it slightly
169 *     changes the bottommost bits of DLAMDA(I). It does not account
170 *     for hexadecimal or decimal machines without guard digits
171 *     (we know of none). We use a subroutine call to compute
172 *     2*DLAMBDA(I) to prevent optimizing compilers from eliminating
173 *     this code.
174 *
175       DO 10 I = 1, K
176          DLAMDA( I ) = DLAMC3( DLAMDA( I ), DLAMDA( I ) ) - DLAMDA( I )
177    10 CONTINUE
178 *
179       DO 20 J = 1, K
180          CALL DLAED4( K, J, DLAMDA, W, Q( 1, J ), RHO, D( J ), INFO )
181 *
182 *        If the zero finder fails, the computation is terminated.
183 *
184          IF( INFO.NE.0 )
185      $      GO TO 120
186    20 CONTINUE
187 *
188       IF( K.EQ.1 )
189      $   GO TO 110
190       IF( K.EQ.2 ) THEN
191          DO 30 J = 1, K
192             W( 1 ) = Q( 1, J )
193             W( 2 ) = Q( 2, J )
194             II = INDX( 1 )
195             Q( 1, J ) = W( II )
196             II = INDX( 2 )
197             Q( 2, J ) = W( II )
198    30    CONTINUE
199          GO TO 110
200       END IF
201 *
202 *     Compute updated W.
203 *
204       CALL DCOPY( K, W, 1, S, 1 )
205 *
206 *     Initialize W(I) = Q(I,I)
207 *
208       CALL DCOPY( K, Q, LDQ+1, W, 1 )
209       DO 60 J = 1, K
210          DO 40 I = 1, J - 1
211             W( I ) = W( I )*( Q( I, J ) / ( DLAMDA( I )-DLAMDA( J ) ) )
212    40    CONTINUE
213          DO 50 I = J + 1, K
214             W( I ) = W( I )*( Q( I, J ) / ( DLAMDA( I )-DLAMDA( J ) ) )
215    50    CONTINUE
216    60 CONTINUE
217       DO 70 I = 1, K
218          W( I ) = SIGNSQRT-W( I ) ), S( I ) )
219    70 CONTINUE
220 *
221 *     Compute eigenvectors of the modified rank-1 modification.
222 *
223       DO 100 J = 1, K
224          DO 80 I = 1, K
225             S( I ) = W( I ) / Q( I, J )
226    80    CONTINUE
227          TEMP = DNRM2( K, S, 1 )
228          DO 90 I = 1, K
229             II = INDX( I )
230             Q( I, J ) = S( II ) / TEMP
231    90    CONTINUE
232   100 CONTINUE
233 *
234 *     Compute the updated eigenvectors.
235 *
236   110 CONTINUE
237 *
238       N2 = N - N1
239       N12 = CTOT( 1 ) + CTOT( 2 )
240       N23 = CTOT( 2 ) + CTOT( 3 )
241 *
242       CALL DLACPY( 'A', N23, K, Q( CTOT( 1 )+11 ), LDQ, S, N23 )
243       IQ2 = N1*N12 + 1
244       IF( N23.NE.0 ) THEN
245          CALL DGEMM( 'N''N', N2, K, N23, ONE, Q2( IQ2 ), N2, S, N23,
246      $               ZERO, Q( N1+11 ), LDQ )
247       ELSE
248          CALL DLASET( 'A', N2, K, ZERO, ZERO, Q( N1+11 ), LDQ )
249       END IF
250 *
251       CALL DLACPY( 'A', N12, K, Q, LDQ, S, N12 )
252       IF( N12.NE.0 ) THEN
253          CALL DGEMM( 'N''N', N1, K, N12, ONE, Q2, N1, S, N12, ZERO, Q,
254      $               LDQ )
255       ELSE
256          CALL DLASET( 'A', N1, K, ZERO, ZERO, Q( 11 ), LDQ )
257       END IF
258 *
259 *
260   120 CONTINUE
261       RETURN
262 *
263 *     End of DLAED3
264 *
265       END