1       SUBROUTINE DGGSVD( JOBU, JOBV, JOBQ, M, N, P, K, L, A, LDA, B,
  2      $                   LDB, ALPHA, BETA, U, LDU, V, LDV, Q, LDQ, WORK,
  3      $                   IWORK, INFO )
  4 *
  5 *  -- LAPACK driver routine (version 3.3.1) --
  6 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  7 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  8 *  -- April 2011                                                      --
  9 *
 10 *     .. Scalar Arguments ..
 11       CHARACTER          JOBQ, JOBU, JOBV
 12       INTEGER            INFO, K, L, LDA, LDB, LDQ, LDU, LDV, M, N, P
 13 *     ..
 14 *     .. Array Arguments ..
 15       INTEGER            IWORK( * )
 16       DOUBLE PRECISION   A( LDA, * ), ALPHA( * ), B( LDB, * ),
 17      $                   BETA( * ), Q( LDQ, * ), U( LDU, * ),
 18      $                   V( LDV, * ), WORK( * )
 19 *     ..
 20 *
 21 *  Purpose
 22 *  =======
 23 *
 24 *  DGGSVD computes the generalized singular value decomposition (GSVD)
 25 *  of an M-by-N real matrix A and P-by-N real matrix B:
 26 *
 27 *        U**T*A*Q = D1*( 0 R ),    V**T*B*Q = D2*( 0 R )
 28 *
 29 *  where U, V and Q are orthogonal matrices.
 30 *  Let K+L = the effective numerical rank of the matrix (A**T,B**T)**T,
 31 *  then R is a K+L-by-K+L nonsingular upper triangular matrix, D1 and
 32 *  D2 are M-by-(K+L) and P-by-(K+L) "diagonal" matrices and of the
 33 *  following structures, respectively:
 34 *
 35 *  If M-K-L >= 0,
 36 *
 37 *                      K  L
 38 *         D1 =     K ( I  0 )
 39 *                  L ( 0  C )
 40 *              M-K-L ( 0  0 )
 41 *
 42 *                    K  L
 43 *         D2 =   L ( 0  S )
 44 *              P-L ( 0  0 )
 45 *
 46 *                  N-K-L  K    L
 47 *    ( 0 R ) = K (  0   R11  R12 )
 48 *              L (  0    0   R22 )
 49 *
 50 *  where
 51 *
 52 *    C = diag( ALPHA(K+1), ... , ALPHA(K+L) ),
 53 *    S = diag( BETA(K+1),  ... , BETA(K+L) ),
 54 *    C**2 + S**2 = I.
 55 *
 56 *    R is stored in A(1:K+L,N-K-L+1:N) on exit.
 57 *
 58 *  If M-K-L < 0,
 59 *
 60 *                    K M-K K+L-M
 61 *         D1 =   K ( I  0    0   )
 62 *              M-K ( 0  C    0   )
 63 *
 64 *                      K M-K K+L-M
 65 *         D2 =   M-K ( 0  S    0  )
 66 *              K+L-M ( 0  0    I  )
 67 *                P-L ( 0  0    0  )
 68 *
 69 *                     N-K-L  K   M-K  K+L-M
 70 *    ( 0 R ) =     K ( 0    R11  R12  R13  )
 71 *                M-K ( 0     0   R22  R23  )
 72 *              K+L-M ( 0     0    0   R33  )
 73 *
 74 *  where
 75 *
 76 *    C = diag( ALPHA(K+1), ... , ALPHA(M) ),
 77 *    S = diag( BETA(K+1),  ... , BETA(M) ),
 78 *    C**2 + S**2 = I.
 79 *
 80 *    (R11 R12 R13 ) is stored in A(1:M, N-K-L+1:N), and R33 is stored
 81 *    ( 0  R22 R23 )
 82 *    in B(M-K+1:L,N+M-K-L+1:N) on exit.
 83 *
 84 *  The routine computes C, S, R, and optionally the orthogonal
 85 *  transformation matrices U, V and Q.
 86 *
 87 *  In particular, if B is an N-by-N nonsingular matrix, then the GSVD of
 88 *  A and B implicitly gives the SVD of A*inv(B):
 89 *                       A*inv(B) = U*(D1*inv(D2))*V**T.
 90 *  If ( A**T,B**T)**T  has orthonormal columns, then the GSVD of A and B is
 91 *  also equal to the CS decomposition of A and B. Furthermore, the GSVD
 92 *  can be used to derive the solution of the eigenvalue problem:
 93 *                       A**T*A x = lambda* B**T*B x.
 94 *  In some literature, the GSVD of A and B is presented in the form
 95 *                   U**T*A*X = ( 0 D1 ),   V**T*B*X = ( 0 D2 )
 96 *  where U and V are orthogonal and X is nonsingular, D1 and D2 are
 97 *  ``diagonal''.  The former GSVD form can be converted to the latter
 98 *  form by taking the nonsingular matrix X as
 99 *
100 *                       X = Q*( I   0    )
101 *                             ( 0 inv(R) ).
102 *
103 *  Arguments
104 *  =========
105 *
106 *  JOBU    (input) CHARACTER*1
107 *          = 'U':  Orthogonal matrix U is computed;
108 *          = 'N':  U is not computed.
109 *
110 *  JOBV    (input) CHARACTER*1
111 *          = 'V':  Orthogonal matrix V is computed;
112 *          = 'N':  V is not computed.
113 *
114 *  JOBQ    (input) CHARACTER*1
115 *          = 'Q':  Orthogonal matrix Q is computed;
116 *          = 'N':  Q is not computed.
117 *
118 *  M       (input) INTEGER
119 *          The number of rows of the matrix A.  M >= 0.
120 *
121 *  N       (input) INTEGER
122 *          The number of columns of the matrices A and B.  N >= 0.
123 *
124 *  P       (input) INTEGER
125 *          The number of rows of the matrix B.  P >= 0.
126 *
127 *  K       (output) INTEGER
128 *  L       (output) INTEGER
129 *          On exit, K and L specify the dimension of the subblocks
130 *          described in the Purpose section.
131 *          K + L = effective numerical rank of (A**T,B**T)**T.
132 *
133 *  A       (input/output) DOUBLE PRECISION array, dimension (LDA,N)
134 *          On entry, the M-by-N matrix A.
135 *          On exit, A contains the triangular matrix R, or part of R.
136 *          See Purpose for details.
137 *
138 *  LDA     (input) INTEGER
139 *          The leading dimension of the array A. LDA >= max(1,M).
140 *
141 *  B       (input/output) DOUBLE PRECISION array, dimension (LDB,N)
142 *          On entry, the P-by-N matrix B.
143 *          On exit, B contains the triangular matrix R if M-K-L < 0.
144 *          See Purpose for details.
145 *
146 *  LDB     (input) INTEGER
147 *          The leading dimension of the array B. LDB >= max(1,P).
148 *
149 *  ALPHA   (output) DOUBLE PRECISION array, dimension (N)
150 *  BETA    (output) DOUBLE PRECISION array, dimension (N)
151 *          On exit, ALPHA and BETA contain the generalized singular
152 *          value pairs of A and B;
153 *            ALPHA(1:K) = 1,
154 *            BETA(1:K)  = 0,
155 *          and if M-K-L >= 0,
156 *            ALPHA(K+1:K+L) = C,
157 *            BETA(K+1:K+L)  = S,
158 *          or if M-K-L < 0,
159 *            ALPHA(K+1:M)=C, ALPHA(M+1:K+L)=0
160 *            BETA(K+1:M) =S, BETA(M+1:K+L) =1
161 *          and
162 *            ALPHA(K+L+1:N) = 0
163 *            BETA(K+L+1:N)  = 0
164 *
165 *  U       (output) DOUBLE PRECISION array, dimension (LDU,M)
166 *          If JOBU = 'U', U contains the M-by-M orthogonal matrix U.
167 *          If JOBU = 'N', U is not referenced.
168 *
169 *  LDU     (input) INTEGER
170 *          The leading dimension of the array U. LDU >= max(1,M) if
171 *          JOBU = 'U'; LDU >= 1 otherwise.
172 *
173 *  V       (output) DOUBLE PRECISION array, dimension (LDV,P)
174 *          If JOBV = 'V', V contains the P-by-P orthogonal matrix V.
175 *          If JOBV = 'N', V is not referenced.
176 *
177 *  LDV     (input) INTEGER
178 *          The leading dimension of the array V. LDV >= max(1,P) if
179 *          JOBV = 'V'; LDV >= 1 otherwise.
180 *
181 *  Q       (output) DOUBLE PRECISION array, dimension (LDQ,N)
182 *          If JOBQ = 'Q', Q contains the N-by-N orthogonal matrix Q.
183 *          If JOBQ = 'N', Q is not referenced.
184 *
185 *  LDQ     (input) INTEGER
186 *          The leading dimension of the array Q. LDQ >= max(1,N) if
187 *          JOBQ = 'Q'; LDQ >= 1 otherwise.
188 *
189 *  WORK    (workspace) DOUBLE PRECISION array,
190 *                      dimension (max(3*N,M,P)+N)
191 *
192 *  IWORK   (workspace/output) INTEGER array, dimension (N)
193 *          On exit, IWORK stores the sorting information. More
194 *          precisely, the following loop will sort ALPHA
195 *             for I = K+1, min(M,K+L)
196 *                 swap ALPHA(I) and ALPHA(IWORK(I))
197 *             endfor
198 *          such that ALPHA(1) >= ALPHA(2) >= ... >= ALPHA(N).
199 *
200 *  INFO    (output) INTEGER
201 *          = 0:  successful exit
202 *          < 0:  if INFO = -i, the i-th argument had an illegal value.
203 *          > 0:  if INFO = 1, the Jacobi-type procedure failed to
204 *                converge.  For further details, see subroutine DTGSJA.
205 *
206 *  Internal Parameters
207 *  ===================
208 *
209 *  TOLA    DOUBLE PRECISION
210 *  TOLB    DOUBLE PRECISION
211 *          TOLA and TOLB are the thresholds to determine the effective
212 *          rank of (A',B')**T. Generally, they are set to
213 *                   TOLA = MAX(M,N)*norm(A)*MAZHEPS,
214 *                   TOLB = MAX(P,N)*norm(B)*MAZHEPS.
215 *          The size of TOLA and TOLB may affect the size of backward
216 *          errors of the decomposition.
217 *
218 *  Further Details
219 *  ===============
220 *
221 *  2-96 Based on modifications by
222 *     Ming Gu and Huan Ren, Computer Science Division, University of
223 *     California at Berkeley, USA
224 *
225 *  =====================================================================
226 *
227 *     .. Local Scalars ..
228       LOGICAL            WANTQ, WANTU, WANTV
229       INTEGER            I, IBND, ISUB, J, NCYCLE
230       DOUBLE PRECISION   ANORM, BNORM, SMAX, TEMP, TOLA, TOLB, ULP, UNFL
231 *     ..
232 *     .. External Functions ..
233       LOGICAL            LSAME
234       DOUBLE PRECISION   DLAMCH, DLANGE
235       EXTERNAL           LSAME, DLAMCH, DLANGE
236 *     ..
237 *     .. External Subroutines ..
238       EXTERNAL           DCOPY, DGGSVP, DTGSJA, XERBLA
239 *     ..
240 *     .. Intrinsic Functions ..
241       INTRINSIC          MAXMIN
242 *     ..
243 *     .. Executable Statements ..
244 *
245 *     Test the input parameters
246 *
247       WANTU = LSAME( JOBU, 'U' )
248       WANTV = LSAME( JOBV, 'V' )
249       WANTQ = LSAME( JOBQ, 'Q' )
250 *
251       INFO = 0
252       IF.NOT.( WANTU .OR. LSAME( JOBU, 'N' ) ) ) THEN
253          INFO = -1
254       ELSE IF.NOT.( WANTV .OR. LSAME( JOBV, 'N' ) ) ) THEN
255          INFO = -2
256       ELSE IF.NOT.( WANTQ .OR. LSAME( JOBQ, 'N' ) ) ) THEN
257          INFO = -3
258       ELSE IF( M.LT.0 ) THEN
259          INFO = -4
260       ELSE IF( N.LT.0 ) THEN
261          INFO = -5
262       ELSE IF( P.LT.0 ) THEN
263          INFO = -6
264       ELSE IF( LDA.LT.MAX1, M ) ) THEN
265          INFO = -10
266       ELSE IF( LDB.LT.MAX1, P ) ) THEN
267          INFO = -12
268       ELSE IF( LDU.LT.1 .OR. ( WANTU .AND. LDU.LT.M ) ) THEN
269          INFO = -16
270       ELSE IF( LDV.LT.1 .OR. ( WANTV .AND. LDV.LT.P ) ) THEN
271          INFO = -18
272       ELSE IF( LDQ.LT.1 .OR. ( WANTQ .AND. LDQ.LT.N ) ) THEN
273          INFO = -20
274       END IF
275       IF( INFO.NE.0 ) THEN
276          CALL XERBLA( 'DGGSVD'-INFO )
277          RETURN
278       END IF
279 *
280 *     Compute the Frobenius norm of matrices A and B
281 *
282       ANORM = DLANGE( '1', M, N, A, LDA, WORK )
283       BNORM = DLANGE( '1', P, N, B, LDB, WORK )
284 *
285 *     Get machine precision and set up threshold for determining
286 *     the effective numerical rank of the matrices A and B.
287 *
288       ULP = DLAMCH( 'Precision' )
289       UNFL = DLAMCH( 'Safe Minimum' )
290       TOLA = MAX( M, N )*MAX( ANORM, UNFL )*ULP
291       TOLB = MAX( P, N )*MAX( BNORM, UNFL )*ULP
292 *
293 *     Preprocessing
294 *
295       CALL DGGSVP( JOBU, JOBV, JOBQ, M, P, N, A, LDA, B, LDB, TOLA,
296      $             TOLB, K, L, U, LDU, V, LDV, Q, LDQ, IWORK, WORK,
297      $             WORK( N+1 ), INFO )
298 *
299 *     Compute the GSVD of two upper "triangular" matrices
300 *
301       CALL DTGSJA( JOBU, JOBV, JOBQ, M, P, N, K, L, A, LDA, B, LDB,
302      $             TOLA, TOLB, ALPHA, BETA, U, LDU, V, LDV, Q, LDQ,
303      $             WORK, NCYCLE, INFO )
304 *
305 *     Sort the singular values and store the pivot indices in IWORK
306 *     Copy ALPHA to WORK, then sort ALPHA in WORK
307 *
308       CALL DCOPY( N, ALPHA, 1, WORK, 1 )
309       IBND = MIN( L, M-K )
310       DO 20 I = 1, IBND
311 *
312 *        Scan for largest ALPHA(K+I)
313 *
314          ISUB = I
315          SMAX = WORK( K+I )
316          DO 10 J = I + 1, IBND
317             TEMP = WORK( K+J )
318             IF( TEMP.GT.SMAX ) THEN
319                ISUB = J
320                SMAX = TEMP
321             END IF
322    10    CONTINUE
323          IF( ISUB.NE.I ) THEN
324             WORK( K+ISUB ) = WORK( K+I )
325             WORK( K+I ) = SMAX
326             IWORK( K+I ) = K + ISUB
327          ELSE
328             IWORK( K+I ) = K + I
329          END IF
330    20 CONTINUE
331 *
332       RETURN
333 *
334 *     End of DGGSVD
335 *
336       END