1       SUBROUTINE DSYTRF( UPLO, N, A, LDA, IPIV, 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       INTEGER            IPIV( * )
 14       DOUBLE PRECISION   A( LDA, * ), WORK( * )
 15 *     ..
 16 *
 17 *  Purpose
 18 *  =======
 19 *
 20 *  DSYTRF computes the factorization of a real symmetric matrix A using
 21 *  the Bunch-Kaufman diagonal pivoting method.  The form of the
 22 *  factorization is
 23 *
 24 *     A = U*D*U**T  or  A = L*D*L**T
 25 *
 26 *  where U (or L) is a product of permutation and unit upper (lower)
 27 *  triangular matrices, and D is symmetric and block diagonal with
 28 *  1-by-1 and 2-by-2 diagonal blocks.
 29 *
 30 *  This is the blocked version of the algorithm, calling Level 3 BLAS.
 31 *
 32 *  Arguments
 33 *  =========
 34 *
 35 *  UPLO    (input) CHARACTER*1
 36 *          = 'U':  Upper triangle of A is stored;
 37 *          = 'L':  Lower triangle of A is stored.
 38 *
 39 *  N       (input) INTEGER
 40 *          The order of the matrix A.  N >= 0.
 41 *
 42 *  A       (input/output) DOUBLE PRECISION array, dimension (LDA,N)
 43 *          On entry, the symmetric matrix A.  If UPLO = 'U', the leading
 44 *          N-by-N upper triangular part of A contains the upper
 45 *          triangular part of the matrix A, and the strictly lower
 46 *          triangular part of A is not referenced.  If UPLO = 'L', the
 47 *          leading N-by-N lower triangular part of A contains the lower
 48 *          triangular part of the matrix A, and the strictly upper
 49 *          triangular part of A is not referenced.
 50 *
 51 *          On exit, the block diagonal matrix D and the multipliers used
 52 *          to obtain the factor U or L (see below for further details).
 53 *
 54 *  LDA     (input) INTEGER
 55 *          The leading dimension of the array A.  LDA >= max(1,N).
 56 *
 57 *  IPIV    (output) INTEGER array, dimension (N)
 58 *          Details of the interchanges and the block structure of D.
 59 *          If IPIV(k) > 0, then rows and columns k and IPIV(k) were
 60 *          interchanged and D(k,k) is a 1-by-1 diagonal block.
 61 *          If UPLO = 'U' and IPIV(k) = IPIV(k-1) < 0, then rows and
 62 *          columns k-1 and -IPIV(k) were interchanged and D(k-1:k,k-1:k)
 63 *          is a 2-by-2 diagonal block.  If UPLO = 'L' and IPIV(k) =
 64 *          IPIV(k+1) < 0, then rows and columns k+1 and -IPIV(k) were
 65 *          interchanged and D(k:k+1,k:k+1) is a 2-by-2 diagonal block.
 66 *
 67 *  WORK    (workspace/output) DOUBLE PRECISION array, dimension (MAX(1,LWORK))
 68 *          On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
 69 *
 70 *  LWORK   (input) INTEGER
 71 *          The length of WORK.  LWORK >=1.  For best performance
 72 *          LWORK >= N*NB, where NB is the block size returned by ILAENV.
 73 *
 74 *          If LWORK = -1, then a workspace query is assumed; the routine
 75 *          only calculates the optimal size of the WORK array, returns
 76 *          this value as the first entry of the WORK array, and no error
 77 *          message related to LWORK is issued by XERBLA.
 78 *
 79 *  INFO    (output) INTEGER
 80 *          = 0:  successful exit
 81 *          < 0:  if INFO = -i, the i-th argument had an illegal value
 82 *          > 0:  if INFO = i, D(i,i) is exactly zero.  The factorization
 83 *                has been completed, but the block diagonal matrix D is
 84 *                exactly singular, and division by zero will occur if it
 85 *                is used to solve a system of equations.
 86 *
 87 *  Further Details
 88 *  ===============
 89 *
 90 *  If UPLO = 'U', then A = U*D*U**T, where
 91 *     U = P(n)*U(n)* ... *P(k)U(k)* ...,
 92 *  i.e., U is a product of terms P(k)*U(k), where k decreases from n to
 93 *  1 in steps of 1 or 2, and D is a block diagonal matrix with 1-by-1
 94 *  and 2-by-2 diagonal blocks D(k).  P(k) is a permutation matrix as
 95 *  defined by IPIV(k), and U(k) is a unit upper triangular matrix, such
 96 *  that if the diagonal block D(k) is of order s (s = 1 or 2), then
 97 *
 98 *             (   I    v    0   )   k-s
 99 *     U(k) =  (   0    I    0   )   s
100 *             (   0    0    I   )   n-k
101 *                k-s   s   n-k
102 *
103 *  If s = 1, D(k) overwrites A(k,k), and v overwrites A(1:k-1,k).
104 *  If s = 2, the upper triangle of D(k) overwrites A(k-1,k-1), A(k-1,k),
105 *  and A(k,k), and v overwrites A(1:k-2,k-1:k).
106 *
107 *  If UPLO = 'L', then A = L*D*L**T, where
108 *     L = P(1)*L(1)* ... *P(k)*L(k)* ...,
109 *  i.e., L is a product of terms P(k)*L(k), where k increases from 1 to
110 *  n in steps of 1 or 2, and D is a block diagonal matrix with 1-by-1
111 *  and 2-by-2 diagonal blocks D(k).  P(k) is a permutation matrix as
112 *  defined by IPIV(k), and L(k) is a unit lower triangular matrix, such
113 *  that if the diagonal block D(k) is of order s (s = 1 or 2), then
114 *
115 *             (   I    0     0   )  k-1
116 *     L(k) =  (   0    I     0   )  s
117 *             (   0    v     I   )  n-k-s+1
118 *                k-1   s  n-k-s+1
119 *
120 *  If s = 1, D(k) overwrites A(k,k), and v overwrites A(k+1:n,k).
121 *  If s = 2, the lower triangle of D(k) overwrites A(k,k), A(k+1,k),
122 *  and A(k+1,k+1), and v overwrites A(k+2:n,k:k+1).
123 *
124 *  =====================================================================
125 *
126 *     .. Local Scalars ..
127       LOGICAL            LQUERY, UPPER
128       INTEGER            IINFO, IWS, J, K, KB, LDWORK, LWKOPT, NB, NBMIN
129 *     ..
130 *     .. External Functions ..
131       LOGICAL            LSAME
132       INTEGER            ILAENV
133       EXTERNAL           LSAME, ILAENV
134 *     ..
135 *     .. External Subroutines ..
136       EXTERNAL           DLASYF, DSYTF2, XERBLA
137 *     ..
138 *     .. Intrinsic Functions ..
139       INTRINSIC          MAX
140 *     ..
141 *     .. Executable Statements ..
142 *
143 *     Test the input parameters.
144 *
145       INFO = 0
146       UPPER = LSAME( UPLO, 'U' )
147       LQUERY = ( LWORK.EQ.-1 )
148       IF.NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
149          INFO = -1
150       ELSE IF( N.LT.0 ) THEN
151          INFO = -2
152       ELSE IF( LDA.LT.MAX1, N ) ) THEN
153          INFO = -4
154       ELSE IF( LWORK.LT.1 .AND. .NOT.LQUERY ) THEN
155          INFO = -7
156       END IF
157 *
158       IF( INFO.EQ.0 ) THEN
159 *
160 *        Determine the block size
161 *
162          NB = ILAENV( 1'DSYTRF', UPLO, N, -1-1-1 )
163          LWKOPT = N*NB
164          WORK( 1 ) = LWKOPT
165       END IF
166 *
167       IF( INFO.NE.0 ) THEN
168          CALL XERBLA( 'DSYTRF'-INFO )
169          RETURN
170       ELSE IF( LQUERY ) THEN
171          RETURN
172       END IF
173 *
174       NBMIN = 2
175       LDWORK = N
176       IF( NB.GT.1 .AND. NB.LT.N ) THEN
177          IWS = LDWORK*NB
178          IF( LWORK.LT.IWS ) THEN
179             NB = MAX( LWORK / LDWORK, 1 )
180             NBMIN = MAX2, ILAENV( 2'DSYTRF', UPLO, N, -1-1-1 ) )
181          END IF
182       ELSE
183          IWS = 1
184       END IF
185       IF( NB.LT.NBMIN )
186      $   NB = N
187 *
188       IF( UPPER ) THEN
189 *
190 *        Factorize A as U*D*U**T using the upper triangle of A
191 *
192 *        K is the main loop index, decreasing from N to 1 in steps of
193 *        KB, where KB is the number of columns factorized by DLASYF;
194 *        KB is either NB or NB-1, or K for the last block
195 *
196          K = N
197    10    CONTINUE
198 *
199 *        If K < 1, exit from loop
200 *
201          IF( K.LT.1 )
202      $      GO TO 40
203 *
204          IF( K.GT.NB ) THEN
205 *
206 *           Factorize columns k-kb+1:k of A and use blocked code to
207 *           update columns 1:k-kb
208 *
209             CALL DLASYF( UPLO, K, NB, KB, A, LDA, IPIV, WORK, LDWORK,
210      $                   IINFO )
211          ELSE
212 *
213 *           Use unblocked code to factorize columns 1:k of A
214 *
215             CALL DSYTF2( UPLO, K, A, LDA, IPIV, IINFO )
216             KB = K
217          END IF
218 *
219 *        Set INFO on the first occurrence of a zero pivot
220 *
221          IF( INFO.EQ.0 .AND. IINFO.GT.0 )
222      $      INFO = IINFO
223 *
224 *        Decrease K and return to the start of the main loop
225 *
226          K = K - KB
227          GO TO 10
228 *
229       ELSE
230 *
231 *        Factorize A as L*D*L**T using the lower triangle of A
232 *
233 *        K is the main loop index, increasing from 1 to N in steps of
234 *        KB, where KB is the number of columns factorized by DLASYF;
235 *        KB is either NB or NB-1, or N-K+1 for the last block
236 *
237          K = 1
238    20    CONTINUE
239 *
240 *        If K > N, exit from loop
241 *
242          IF( K.GT.N )
243      $      GO TO 40
244 *
245          IF( K.LE.N-NB ) THEN
246 *
247 *           Factorize columns k:k+kb-1 of A and use blocked code to
248 *           update columns k+kb:n
249 *
250             CALL DLASYF( UPLO, N-K+1, NB, KB, A( K, K ), LDA, IPIV( K ),
251      $                   WORK, LDWORK, IINFO )
252          ELSE
253 *
254 *           Use unblocked code to factorize columns k:n of A
255 *
256             CALL DSYTF2( UPLO, N-K+1, A( K, K ), LDA, IPIV( K ), IINFO )
257             KB = N - K + 1
258          END IF
259 *
260 *        Set INFO on the first occurrence of a zero pivot
261 *
262          IF( INFO.EQ.0 .AND. IINFO.GT.0 )
263      $      INFO = IINFO + K - 1
264 *
265 *        Adjust IPIV
266 *
267          DO 30 J = K, K + KB - 1
268             IF( IPIV( J ).GT.0 ) THEN
269                IPIV( J ) = IPIV( J ) + K - 1
270             ELSE
271                IPIV( J ) = IPIV( J ) - K + 1
272             END IF
273    30    CONTINUE
274 *
275 *        Increase K and return to the start of the main loop
276 *
277          K = K + KB
278          GO TO 20
279 *
280       END IF
281 *
282    40 CONTINUE
283       WORK( 1 ) = LWKOPT
284       RETURN
285 *
286 *     End of DSYTRF
287 *
288       END