1       SUBROUTINE DSYTRD( UPLO, N, A, LDA, D, E, TAU, WORK, LWORK, INFO )
  2 *
  3 *  -- LAPACK routine (version 3.3.1) --
  4 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  5 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  6 *  -- April 2011                                                      --
  7 *
  8 *     .. Scalar Arguments ..
  9       CHARACTER          UPLO
 10       INTEGER            INFO, LDA, LWORK, N
 11 *     ..
 12 *     .. Array Arguments ..
 13       DOUBLE PRECISION   A( LDA, * ), D( * ), E( * ), TAU( * ),
 14      $                   WORK( * )
 15 *     ..
 16 *
 17 *  Purpose
 18 *  =======
 19 *
 20 *  DSYTRD reduces a real symmetric matrix A to real symmetric
 21 *  tridiagonal form T by an orthogonal similarity transformation:
 22 *  Q**T * A * Q = T.
 23 *
 24 *  Arguments
 25 *  =========
 26 *
 27 *  UPLO    (input) CHARACTER*1
 28 *          = 'U':  Upper triangle of A is stored;
 29 *          = 'L':  Lower triangle of A is stored.
 30 *
 31 *  N       (input) INTEGER
 32 *          The order of the matrix A.  N >= 0.
 33 *
 34 *  A       (input/output) DOUBLE PRECISION array, dimension (LDA,N)
 35 *          On entry, the symmetric matrix A.  If UPLO = 'U', the leading
 36 *          N-by-N upper triangular part of A contains the upper
 37 *          triangular part of the matrix A, and the strictly lower
 38 *          triangular part of A is not referenced.  If UPLO = 'L', the
 39 *          leading N-by-N lower triangular part of A contains the lower
 40 *          triangular part of the matrix A, and the strictly upper
 41 *          triangular part of A is not referenced.
 42 *          On exit, if UPLO = 'U', the diagonal and first superdiagonal
 43 *          of A are overwritten by the corresponding elements of the
 44 *          tridiagonal matrix T, and the elements above the first
 45 *          superdiagonal, with the array TAU, represent the orthogonal
 46 *          matrix Q as a product of elementary reflectors; if UPLO
 47 *          = 'L', the diagonal and first subdiagonal of A are over-
 48 *          written by the corresponding elements of the tridiagonal
 49 *          matrix T, and the elements below the first subdiagonal, with
 50 *          the array TAU, represent the orthogonal matrix Q as a product
 51 *          of elementary reflectors. See Further Details.
 52 *
 53 *  LDA     (input) INTEGER
 54 *          The leading dimension of the array A.  LDA >= max(1,N).
 55 *
 56 *  D       (output) DOUBLE PRECISION array, dimension (N)
 57 *          The diagonal elements of the tridiagonal matrix T:
 58 *          D(i) = A(i,i).
 59 *
 60 *  E       (output) DOUBLE PRECISION array, dimension (N-1)
 61 *          The off-diagonal elements of the tridiagonal matrix T:
 62 *          E(i) = A(i,i+1) if UPLO = 'U', E(i) = A(i+1,i) if UPLO = 'L'.
 63 *
 64 *  TAU     (output) DOUBLE PRECISION array, dimension (N-1)
 65 *          The scalar factors of the elementary reflectors (see Further
 66 *          Details).
 67 *
 68 *  WORK    (workspace/output) DOUBLE PRECISION array, dimension (MAX(1,LWORK))
 69 *          On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
 70 *
 71 *  LWORK   (input) INTEGER
 72 *          The dimension of the array WORK.  LWORK >= 1.
 73 *          For optimum performance LWORK >= N*NB, where NB is the
 74 *          optimal blocksize.
 75 *
 76 *          If LWORK = -1, then a workspace query is assumed; the routine
 77 *          only calculates the optimal size of the WORK array, returns
 78 *          this value as the first entry of the WORK array, and no error
 79 *          message related to LWORK is issued by XERBLA.
 80 *
 81 *  INFO    (output) INTEGER
 82 *          = 0:  successful exit
 83 *          < 0:  if INFO = -i, the i-th argument had an illegal value
 84 *
 85 *  Further Details
 86 *  ===============
 87 *
 88 *  If UPLO = 'U', the matrix Q is represented as a product of elementary
 89 *  reflectors
 90 *
 91 *     Q = H(n-1) . . . H(2) H(1).
 92 *
 93 *  Each H(i) has the form
 94 *
 95 *     H(i) = I - tau * v * v**T
 96 *
 97 *  where tau is a real scalar, and v is a real vector with
 98 *  v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in
 99 *  A(1:i-1,i+1), and tau in TAU(i).
100 *
101 *  If UPLO = 'L', the matrix Q is represented as a product of elementary
102 *  reflectors
103 *
104 *     Q = H(1) H(2) . . . H(n-1).
105 *
106 *  Each H(i) has the form
107 *
108 *     H(i) = I - tau * v * v**T
109 *
110 *  where tau is a real scalar, and v is a real vector with
111 *  v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i),
112 *  and tau in TAU(i).
113 *
114 *  The contents of A on exit are illustrated by the following examples
115 *  with n = 5:
116 *
117 *  if UPLO = 'U':                       if UPLO = 'L':
118 *
119 *    (  d   e   v2  v3  v4 )              (  d                  )
120 *    (      d   e   v3  v4 )              (  e   d              )
121 *    (          d   e   v4 )              (  v1  e   d          )
122 *    (              d   e  )              (  v1  v2  e   d      )
123 *    (                  d  )              (  v1  v2  v3  e   d  )
124 *
125 *  where d and e denote diagonal and off-diagonal elements of T, and vi
126 *  denotes an element of the vector defining H(i).
127 *
128 *  =====================================================================
129 *
130 *     .. Parameters ..
131       DOUBLE PRECISION   ONE
132       PARAMETER          ( ONE = 1.0D+0 )
133 *     ..
134 *     .. Local Scalars ..
135       LOGICAL            LQUERY, UPPER
136       INTEGER            I, IINFO, IWS, J, KK, LDWORK, LWKOPT, NB,
137      $                   NBMIN, NX
138 *     ..
139 *     .. External Subroutines ..
140       EXTERNAL           DLATRD, DSYR2K, DSYTD2, XERBLA
141 *     ..
142 *     .. Intrinsic Functions ..
143       INTRINSIC          MAX
144 *     ..
145 *     .. External Functions ..
146       LOGICAL            LSAME
147       INTEGER            ILAENV
148       EXTERNAL           LSAME, ILAENV
149 *     ..
150 *     .. Executable Statements ..
151 *
152 *     Test the input parameters
153 *
154       INFO = 0
155       UPPER = LSAME( UPLO, 'U' )
156       LQUERY = ( LWORK.EQ.-1 )
157       IF.NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
158          INFO = -1
159       ELSE IF( N.LT.0 ) THEN
160          INFO = -2
161       ELSE IF( LDA.LT.MAX1, N ) ) THEN
162          INFO = -4
163       ELSE IF( LWORK.LT.1 .AND. .NOT.LQUERY ) THEN
164          INFO = -9
165       END IF
166 *
167       IF( INFO.EQ.0 ) THEN
168 *
169 *        Determine the block size.
170 *
171          NB = ILAENV( 1'DSYTRD', UPLO, N, -1-1-1 )
172          LWKOPT = N*NB
173          WORK( 1 ) = LWKOPT
174       END IF
175 *
176       IF( INFO.NE.0 ) THEN
177          CALL XERBLA( 'DSYTRD'-INFO )
178          RETURN
179       ELSE IF( LQUERY ) THEN
180          RETURN
181       END IF
182 *
183 *     Quick return if possible
184 *
185       IF( N.EQ.0 ) THEN
186          WORK( 1 ) = 1
187          RETURN
188       END IF
189 *
190       NX = N
191       IWS = 1
192       IF( NB.GT.1 .AND. NB.LT.N ) THEN
193 *
194 *        Determine when to cross over from blocked to unblocked code
195 *        (last block is always handled by unblocked code).
196 *
197          NX = MAX( NB, ILAENV( 3'DSYTRD', UPLO, N, -1-1-1 ) )
198          IF( NX.LT.N ) THEN
199 *
200 *           Determine if workspace is large enough for blocked code.
201 *
202             LDWORK = N
203             IWS = LDWORK*NB
204             IF( LWORK.LT.IWS ) THEN
205 *
206 *              Not enough workspace to use optimal NB:  determine the
207 *              minimum value of NB, and reduce NB or force use of
208 *              unblocked code by setting NX = N.
209 *
210                NB = MAX( LWORK / LDWORK, 1 )
211                NBMIN = ILAENV( 2'DSYTRD', UPLO, N, -1-1-1 )
212                IF( NB.LT.NBMIN )
213      $            NX = N
214             END IF
215          ELSE
216             NX = N
217          END IF
218       ELSE
219          NB = 1
220       END IF
221 *
222       IF( UPPER ) THEN
223 *
224 *        Reduce the upper triangle of A.
225 *        Columns 1:kk are handled by the unblocked method.
226 *
227          KK = N - ( ( N-NX+NB-1 ) / NB )*NB
228          DO 20 I = N - NB + 1, KK + 1-NB
229 *
230 *           Reduce columns i:i+nb-1 to tridiagonal form and form the
231 *           matrix W which is needed to update the unreduced part of
232 *           the matrix
233 *
234             CALL DLATRD( UPLO, I+NB-1, NB, A, LDA, E, TAU, WORK,
235      $                   LDWORK )
236 *
237 *           Update the unreduced submatrix A(1:i-1,1:i-1), using an
238 *           update of the form:  A := A - V*W**T - W*V**T
239 *
240             CALL DSYR2K( UPLO, 'No transpose', I-1, NB, -ONE, A( 1, I ),
241      $                   LDA, WORK, LDWORK, ONE, A, LDA )
242 *
243 *           Copy superdiagonal elements back into A, and diagonal
244 *           elements into D
245 *
246             DO 10 J = I, I + NB - 1
247                A( J-1, J ) = E( J-1 )
248                D( J ) = A( J, J )
249    10       CONTINUE
250    20    CONTINUE
251 *
252 *        Use unblocked code to reduce the last or only block
253 *
254          CALL DSYTD2( UPLO, KK, A, LDA, D, E, TAU, IINFO )
255       ELSE
256 *
257 *        Reduce the lower triangle of A
258 *
259          DO 40 I = 1, N - NX, NB
260 *
261 *           Reduce columns i:i+nb-1 to tridiagonal form and form the
262 *           matrix W which is needed to update the unreduced part of
263 *           the matrix
264 *
265             CALL DLATRD( UPLO, N-I+1, NB, A( I, I ), LDA, E( I ),
266      $                   TAU( I ), WORK, LDWORK )
267 *
268 *           Update the unreduced submatrix A(i+ib:n,i+ib:n), using
269 *           an update of the form:  A := A - V*W**T - W*V**T
270 *
271             CALL DSYR2K( UPLO, 'No transpose', N-I-NB+1, NB, -ONE,
272      $                   A( I+NB, I ), LDA, WORK( NB+1 ), LDWORK, ONE,
273      $                   A( I+NB, I+NB ), LDA )
274 *
275 *           Copy subdiagonal elements back into A, and diagonal
276 *           elements into D
277 *
278             DO 30 J = I, I + NB - 1
279                A( J+1, J ) = E( J )
280                D( J ) = A( J, J )
281    30       CONTINUE
282    40    CONTINUE
283 *
284 *        Use unblocked code to reduce the last or only block
285 *
286          CALL DSYTD2( UPLO, N-I+1, A( I, I ), LDA, D( I ), E( I ),
287      $                TAU( I ), IINFO )
288       END IF
289 *
290       WORK( 1 ) = LWKOPT
291       RETURN
292 *
293 *     End of DSYTRD
294 *
295       END