| 1 |
SUBROUTINE DTGEX2( WANTQ, WANTZ, N, A, LDA, B, LDB, Q, LDQ, Z, |
| 2 |
$ LDZ, J1, N1, N2, WORK, LWORK, INFO ) |
| 3 |
* |
| 4 |
* -- LAPACK auxiliary routine (version 3.0) -- |
| 5 |
* Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., |
| 6 |
* Courant Institute, Argonne National Lab, and Rice University |
| 7 |
* June 30, 1999 |
| 8 |
* |
| 9 |
* .. Scalar Arguments .. |
| 10 |
LOGICAL WANTQ, WANTZ |
| 11 |
INTEGER INFO, J1, LDA, LDB, LDQ, LDZ, LWORK, N, N1, N2 |
| 12 |
* .. |
| 13 |
* .. Array Arguments .. |
| 14 |
DOUBLE PRECISION A( LDA, * ), B( LDB, * ), Q( LDQ, * ), |
| 15 |
$ WORK( * ), Z( LDZ, * ) |
| 16 |
* .. |
| 17 |
* |
| 18 |
* Purpose |
| 19 |
* ======= |
| 20 |
* |
| 21 |
* DTGEX2 swaps adjacent diagonal blocks (A11, B11) and (A22, B22) |
| 22 |
* of size 1-by-1 or 2-by-2 in an upper (quasi) triangular matrix pair |
| 23 |
* (A, B) by an orthogonal equivalence transformation. |
| 24 |
* |
| 25 |
* (A, B) must be in generalized real Schur canonical form (as returned |
| 26 |
* by DGGES), i.e. A is block upper triangular with 1-by-1 and 2-by-2 |
| 27 |
* diagonal blocks. B is upper triangular. |
| 28 |
* |
| 29 |
* Optionally, the matrices Q and Z of generalized Schur vectors are |
| 30 |
* updated. |
| 31 |
* |
| 32 |
* Q(in) * A(in) * Z(in)' = Q(out) * A(out) * Z(out)' |
| 33 |
* Q(in) * B(in) * Z(in)' = Q(out) * B(out) * Z(out)' |
| 34 |
* |
| 35 |
* |
| 36 |
* Arguments |
| 37 |
* ========= |
| 38 |
* |
| 39 |
* WANTQ (input) LOGICAL |
| 40 |
* .TRUE. : update the left transformation matrix Q; |
| 41 |
* .FALSE.: do not update Q. |
| 42 |
* |
| 43 |
* WANTZ (input) LOGICAL |
| 44 |
* .TRUE. : update the right transformation matrix Z; |
| 45 |
* .FALSE.: do not update Z. |
| 46 |
* |
| 47 |
* N (input) INTEGER |
| 48 |
* The order of the matrices A and B. N >= 0. |
| 49 |
* |
| 50 |
* A (input/output) DOUBLE PRECISION arrays, dimensions (LDA,N) |
| 51 |
* On entry, the matrix A in the pair (A, B). |
| 52 |
* On exit, the updated matrix A. |
| 53 |
* |
| 54 |
* LDA (input) INTEGER |
| 55 |
* The leading dimension of the array A. LDA >= max(1,N). |
| 56 |
* |
| 57 |
* B (input/output) DOUBLE PRECISION arrays, dimensions (LDB,N) |
| 58 |
* On entry, the matrix B in the pair (A, B). |
| 59 |
* On exit, the updated matrix B. |
| 60 |
* |
| 61 |
* LDB (input) INTEGER |
| 62 |
* The leading dimension of the array B. LDB >= max(1,N). |
| 63 |
* |
| 64 |
* Q (input/output) DOUBLE PRECISION array, dimension (LDZ,N) |
| 65 |
* On entry, if WANTQ = .TRUE., the orthogonal matrix Q. |
| 66 |
* On exit, the updated matrix Q. |
| 67 |
* Not referenced if WANTQ = .FALSE.. |
| 68 |
* |
| 69 |
* LDQ (input) INTEGER |
| 70 |
* The leading dimension of the array Q. LDQ >= 1. |
| 71 |
* If WANTQ = .TRUE., LDQ >= N. |
| 72 |
* |
| 73 |
* Z (input/output) DOUBLE PRECISION array, dimension (LDZ,N) |
| 74 |
* On entry, if WANTZ =.TRUE., the orthogonal matrix Z. |
| 75 |
* On exit, the updated matrix Z. |
| 76 |
* Not referenced if WANTZ = .FALSE.. |
| 77 |
* |
| 78 |
* LDZ (input) INTEGER |
| 79 |
* The leading dimension of the array Z. LDZ >= 1. |
| 80 |
* If WANTZ = .TRUE., LDZ >= N. |
| 81 |
* |
| 82 |
* J1 (input) INTEGER |
| 83 |
* The index to the first block (A11, B11). 1 <= J1 <= N. |
| 84 |
* |
| 85 |
* N1 (input) INTEGER |
| 86 |
* The order of the first block (A11, B11). N1 = 0, 1 or 2. |
| 87 |
* |
| 88 |
* N2 (input) INTEGER |
| 89 |
* The order of the second block (A22, B22). N2 = 0, 1 or 2. |
| 90 |
* |
| 91 |
* WORK (workspace) DOUBLE PRECISION array, dimension (LWORK). |
| 92 |
* |
| 93 |
* LWORK (input) INTEGER |
| 94 |
* The dimension of the array WORK. |
| 95 |
* LWORK >= MAX( N*(N2+N1), (N2+N1)*(N2+N1)*2 ) |
| 96 |
* |
| 97 |
* INFO (output) INTEGER |
| 98 |
* =0: Successful exit |
| 99 |
* >0: If INFO = 1, the transformed matrix (A, B) would be |
| 100 |
* too far from generalized Schur form; the blocks are |
| 101 |
* not swapped and (A, B) and (Q, Z) are unchanged. |
| 102 |
* The problem of swapping is too ill-conditioned. |
| 103 |
* <0: If INFO = -16: LWORK is too small. Appropriate value |
| 104 |
* for LWORK is returned in WORK(1). |
| 105 |
* |
| 106 |
* Further Details |
| 107 |
* =============== |
| 108 |
* |
| 109 |
* Based on contributions by |
| 110 |
* Bo Kagstrom and Peter Poromaa, Department of Computing Science, |
| 111 |
* Umea University, S-901 87 Umea, Sweden. |
| 112 |
* |
| 113 |
* In the current code both weak and strong stability tests are |
| 114 |
* performed. The user can omit the strong stability test by changing |
| 115 |
* the internal logical parameter WANDS to .FALSE.. See ref. [2] for |
| 116 |
* details. |
| 117 |
* |
| 118 |
* [1] B. Kagstrom; A Direct Method for Reordering Eigenvalues in the |
| 119 |
* Generalized Real Schur Form of a Regular Matrix Pair (A, B), in |
| 120 |
* M.S. Moonen et al (eds), Linear Algebra for Large Scale and |
| 121 |
* Real-Time Applications, Kluwer Academic Publ. 1993, pp 195-218. |
| 122 |
* |
| 123 |
* [2] B. Kagstrom and P. Poromaa; Computing Eigenspaces with Specified |
| 124 |
* Eigenvalues of a Regular Matrix Pair (A, B) and Condition |
| 125 |
* Estimation: Theory, Algorithms and Software, |
| 126 |
* Report UMINF - 94.04, Department of Computing Science, Umea |
| 127 |
* University, S-901 87 Umea, Sweden, 1994. Also as LAPACK Working |
| 128 |
* Note 87. To appear in Numerical Algorithms, 1996. |
| 129 |
* |
| 130 |
* ===================================================================== |
| 131 |
* |
| 132 |
* .. Parameters .. |
| 133 |
DOUBLE PRECISION ZERO, ONE |
| 134 |
PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 ) |
| 135 |
DOUBLE PRECISION TEN |
| 136 |
PARAMETER ( TEN = 1.0D+01 ) |
| 137 |
INTEGER LDST |
| 138 |
PARAMETER ( LDST = 4 ) |
| 139 |
LOGICAL WANDS |
| 140 |
PARAMETER ( WANDS = .TRUE. ) |
| 141 |
* .. |
| 142 |
* .. Local Scalars .. |
| 143 |
LOGICAL DTRONG, WEAK |
| 144 |
INTEGER I, IDUM, LINFO, M |
| 145 |
DOUBLE PRECISION BQRA21, BRQA21, DDUM, DNORM, DSCALE, DSUM, EPS, |
| 146 |
$ F, G, SA, SB, SCALE, SMLNUM, SS, THRESH, WS |
| 147 |
* .. |
| 148 |
* .. Local Arrays .. |
| 149 |
INTEGER IWORK( LDST ) |
| 150 |
DOUBLE PRECISION AI( 2 ), AR( 2 ), BE( 2 ), IR( LDST, LDST ), |
| 151 |
$ IRCOP( LDST, LDST ), LI( LDST, LDST ), |
| 152 |
$ LICOP( LDST, LDST ), S( LDST, LDST ), |
| 153 |
$ SCPY( LDST, LDST ), T( LDST, LDST ), |
| 154 |
$ TAUL( LDST ), TAUR( LDST ), TCPY( LDST, LDST ) |
| 155 |
* .. |
| 156 |
* .. External Functions .. |
| 157 |
DOUBLE PRECISION DLAMCH |
| 158 |
EXTERNAL DLAMCH |
| 159 |
* .. |
| 160 |
* .. External Subroutines .. |
| 161 |
EXTERNAL DCOPY, DGEMM, DGEQR2, DGERQ2, DLACPY, DLAGV2, |
| 162 |
$ DLARTG, DLASSQ, DORG2R, DORGR2, DORM2R, DORMR2, |
| 163 |
$ DROT, DSCAL, DTGSY2 |
| 164 |
* .. |
| 165 |
* .. Intrinsic Functions .. |
| 166 |
INTRINSIC ABS, MAX, SQRT |
| 167 |
* .. |
| 168 |
* .. Executable Statements .. |
| 169 |
* |
| 170 |
INFO = 0 |
| 171 |
* |
| 172 |
* Quick return if possible |
| 173 |
* |
| 174 |
IF( N.LE.1 .OR. N1.LE.0 .OR. N2.LE.0 ) |
| 175 |
$ RETURN |
| 176 |
IF( N1.GT.N .OR. ( J1+N1 ).GT.N ) |
| 177 |
$ RETURN |
| 178 |
M = N1 + N2 |
| 179 |
IF( LWORK.LT.MAX( N*M, M*M*2 ) ) THEN |
| 180 |
INFO = -16 |
| 181 |
WORK( 1 ) = MAX( N*M, M*M*2 ) |
| 182 |
RETURN |
| 183 |
END IF |
| 184 |
* |
| 185 |
WEAK = .FALSE. |
| 186 |
DTRONG = .FALSE. |
| 187 |
* |
| 188 |
* Make a local copy of selected block |
| 189 |
* |
| 190 |
CALL DCOPY( LDST*LDST, ZERO, 0, LI, 1 ) |
| 191 |
CALL DCOPY( LDST*LDST, ZERO, 0, IR, 1 ) |
| 192 |
CALL DLACPY( 'Full', M, M, A( J1, J1 ), LDA, S, LDST ) |
| 193 |
CALL DLACPY( 'Full', M, M, B( J1, J1 ), LDB, T, LDST ) |
| 194 |
* |
| 195 |
* Compute threshold for testing acceptance of swapping. |
| 196 |
* |
| 197 |
EPS = DLAMCH( 'P' ) |
| 198 |
SMLNUM = DLAMCH( 'S' ) / EPS |
| 199 |
DSCALE = ZERO |
| 200 |
DSUM = ONE |
| 201 |
CALL DLACPY( 'Full', M, M, S, LDST, WORK, M ) |
| 202 |
CALL DLASSQ( M*M, WORK, 1, DSCALE, DSUM ) |
| 203 |
CALL DLACPY( 'Full', M, M, T, LDST, WORK, M ) |
| 204 |
CALL DLASSQ( M*M, WORK, 1, DSCALE, DSUM ) |
| 205 |
DNORM = DSCALE*SQRT( DSUM ) |
| 206 |
THRESH = MAX( TEN*EPS*DNORM, SMLNUM ) |
| 207 |
* |
| 208 |
IF( M.EQ.2 ) THEN |
| 209 |
* |
| 210 |
* CASE 1: Swap 1-by-1 and 1-by-1 blocks. |
| 211 |
* |
| 212 |
* Compute orthogonal QL and RQ that swap 1-by-1 and 1-by-1 blocks |
| 213 |
* using Givens rotations and perform the swap tentatively. |
| 214 |
* |
| 215 |
F = S( 2, 2 )*T( 1, 1 ) - T( 2, 2 )*S( 1, 1 ) |
| 216 |
G = S( 2, 2 )*T( 1, 2 ) - T( 2, 2 )*S( 1, 2 ) |
| 217 |
SB = ABS( T( 2, 2 ) ) |
| 218 |
SA = ABS( S( 2, 2 ) ) |
| 219 |
CALL DLARTG( F, G, IR( 1, 2 ), IR( 1, 1 ), DDUM ) |
| 220 |
IR( 2, 1 ) = -IR( 1, 2 ) |
| 221 |
IR( 2, 2 ) = IR( 1, 1 ) |
| 222 |
CALL DROT( 2, S( 1, 1 ), 1, S( 1, 2 ), 1, IR( 1, 1 ), |
| 223 |
$ IR( 2, 1 ) ) |
| 224 |
CALL DROT( 2, T( 1, 1 ), 1, T( 1, 2 ), 1, IR( 1, 1 ), |
| 225 |
$ IR( 2, 1 ) ) |
| 226 |
IF( SA.GE.SB ) THEN |
| 227 |
CALL DLARTG( S( 1, 1 ), S( 2, 1 ), LI( 1, 1 ), LI( 2, 1 ), |
| 228 |
$ DDUM ) |
| 229 |
ELSE |
| 230 |
CALL DLARTG( T( 1, 1 ), T( 2, 1 ), LI( 1, 1 ), LI( 2, 1 ), |
| 231 |
$ DDUM ) |
| 232 |
END IF |
| 233 |
CALL DROT( 2, S( 1, 1 ), LDST, S( 2, 1 ), LDST, LI( 1, 1 ), |
| 234 |
$ LI( 2, 1 ) ) |
| 235 |
CALL DROT( 2, T( 1, 1 ), LDST, T( 2, 1 ), LDST, LI( 1, 1 ), |
| 236 |
$ LI( 2, 1 ) ) |
| 237 |
LI( 2, 2 ) = LI( 1, 1 ) |
| 238 |
LI( 1, 2 ) = -LI( 2, 1 ) |
| 239 |
* |
| 240 |
* Weak stability test: |
| 241 |
* |S21| + |T21| <= O(EPS * F-norm((S, T))) |
| 242 |
* |
| 243 |
WS = ABS( S( 2, 1 ) ) + ABS( T( 2, 1 ) ) |
| 244 |
WEAK = WS.LE.THRESH |
| 245 |
IF( .NOT.WEAK ) |
| 246 |
$ GO TO 70 |
| 247 |
* |
| 248 |
IF( WANDS ) THEN |
| 249 |
* |
| 250 |
* Strong stability test: |
| 251 |
* F-norm((A-QL'*S*QR, B-QL'*T*QR)) <= O(EPS*F-norm((A,B))) |
| 252 |
* |
| 253 |
CALL DLACPY( 'Full', M, M, A( J1, J1 ), LDA, WORK( M*M+1 ), |
| 254 |
$ M ) |
| 255 |
CALL DGEMM( 'N', 'N', M, M, M, ONE, LI, LDST, S, LDST, ZERO, |
| 256 |
$ WORK, M ) |
| 257 |
CALL DGEMM( 'N', 'T', M, M, M, -ONE, WORK, M, IR, LDST, ONE, |
| 258 |
$ WORK( M*M+1 ), M ) |
| 259 |
DSCALE = ZERO |
| 260 |
DSUM = ONE |
| 261 |
CALL DLASSQ( M*M, WORK( M*M+1 ), 1, DSCALE, DSUM ) |
| 262 |
* |
| 263 |
CALL DLACPY( 'Full', M, M, B( J1, J1 ), LDB, WORK( M*M+1 ), |
| 264 |
$ M ) |
| 265 |
CALL DGEMM( 'N', 'N', M, M, M, ONE, LI, LDST, T, LDST, ZERO, |
| 266 |
$ WORK, M ) |
| 267 |
CALL DGEMM( 'N', 'T', M, M, M, -ONE, WORK, M, IR, LDST, ONE, |
| 268 |
$ WORK( M*M+1 ), M ) |
| 269 |
CALL DLASSQ( M*M, WORK( M*M+1 ), 1, DSCALE, DSUM ) |
| 270 |
SS = DSCALE*SQRT( DSUM ) |
| 271 |
DTRONG = SS.LE.THRESH |
| 272 |
IF( .NOT.DTRONG ) |
| 273 |
$ GO TO 70 |
| 274 |
END IF |
| 275 |
* |
| 276 |
* Update (A(J1:J1+M-1, M+J1:N), B(J1:J1+M-1, M+J1:N)) and |
| 277 |
* (A(1:J1-1, J1:J1+M), B(1:J1-1, J1:J1+M)). |
| 278 |
* |
| 279 |
CALL DROT( J1+1, A( 1, J1 ), 1, A( 1, J1+1 ), 1, IR( 1, 1 ), |
| 280 |
$ IR( 2, 1 ) ) |
| 281 |
CALL DROT( J1+1, B( 1, J1 ), 1, B( 1, J1+1 ), 1, IR( 1, 1 ), |
| 282 |
$ IR( 2, 1 ) ) |
| 283 |
CALL DROT( N-J1+1, A( J1, J1 ), LDA, A( J1+1, J1 ), LDA, |
| 284 |
$ LI( 1, 1 ), LI( 2, 1 ) ) |
| 285 |
CALL DROT( N-J1+1, B( J1, J1 ), LDB, B( J1+1, J1 ), LDB, |
| 286 |
$ LI( 1, 1 ), LI( 2, 1 ) ) |
| 287 |
* |
| 288 |
* Set N1-by-N2 (2,1) - blocks to ZERO. |
| 289 |
* |
| 290 |
A( J1+1, J1 ) = ZERO |
| 291 |
B( J1+1, J1 ) = ZERO |
| 292 |
* |
| 293 |
* Accumulate transformations into Q and Z if requested. |
| 294 |
* |
| 295 |
IF( WANTZ ) |
| 296 |
$ CALL DROT( N, Z( 1, J1 ), 1, Z( 1, J1+1 ), 1, IR( 1, 1 ), |
| 297 |
$ IR( 2, 1 ) ) |
| 298 |
IF( WANTQ ) |
| 299 |
$ CALL DROT( N, Q( 1, J1 ), 1, Q( 1, J1+1 ), 1, LI( 1, 1 ), |
| 300 |
$ LI( 2, 1 ) ) |
| 301 |
* |
| 302 |
* Exit with INFO = 0 if swap was successfully performed. |
| 303 |
* |
| 304 |
RETURN |
| 305 |
* |
| 306 |
ELSE |
| 307 |
* |
| 308 |
* CASE 2: Swap 1-by-1 and 2-by-2 blocks, or 2-by-2 |
| 309 |
* and 2-by-2 blocks. |
| 310 |
* |
| 311 |
* Solve the generalized Sylvester equation |
| 312 |
* S11 * R - L * S22 = SCALE * S12 |
| 313 |
* T11 * R - L * T22 = SCALE * T12 |
| 314 |
* for R and L. Solutions in LI and IR. |
| 315 |
* |
| 316 |
CALL DLACPY( 'Full', N1, N2, T( 1, N1+1 ), LDST, LI, LDST ) |
| 317 |
CALL DLACPY( 'Full', N1, N2, S( 1, N1+1 ), LDST, |
| 318 |
$ IR( N2+1, N1+1 ), LDST ) |
| 319 |
CALL DTGSY2( 'N', 0, N1, N2, S, LDST, S( N1+1, N1+1 ), LDST, |
| 320 |
$ IR( N2+1, N1+1 ), LDST, T, LDST, T( N1+1, N1+1 ), |
| 321 |
$ LDST, LI, LDST, SCALE, DSUM, DSCALE, IWORK, IDUM, |
| 322 |
$ LINFO ) |
| 323 |
* |
| 324 |
* Compute orthogonal matrix QL: |
| 325 |
* |
| 326 |
* QL' * LI = [ TL ] |
| 327 |
* [ 0 ] |
| 328 |
* where |
| 329 |
* LI = [ -L ] |
| 330 |
* [ SCALE * identity(N2) ] |
| 331 |
* |
| 332 |
DO 10 I = 1, N2 |
| 333 |
CALL DSCAL( N1, -ONE, LI( 1, I ), 1 ) |
| 334 |
LI( N1+I, I ) = SCALE |
| 335 |
10 CONTINUE |
| 336 |
CALL DGEQR2( M, N2, LI, LDST, TAUL, WORK, LINFO ) |
| 337 |
IF( LINFO.NE.0 ) |
| 338 |
$ GO TO 70 |
| 339 |
CALL DORG2R( M, M, N2, LI, LDST, TAUL, WORK, LINFO ) |
| 340 |
IF( LINFO.NE.0 ) |
| 341 |
$ GO TO 70 |
| 342 |
* |
| 343 |
* Compute orthogonal matrix RQ: |
| 344 |
* |
| 345 |
* IR * RQ' = [ 0 TR], |
| 346 |
* |
| 347 |
* where IR = [ SCALE * identity(N1), R ] |
| 348 |
* |
| 349 |
DO 20 I = 1, N1 |
| 350 |
IR( N2+I, I ) = SCALE |
| 351 |
20 CONTINUE |
| 352 |
CALL DGERQ2( N1, M, IR( N2+1, 1 ), LDST, TAUR, WORK, LINFO ) |
| 353 |
IF( LINFO.NE.0 ) |
| 354 |
$ GO TO 70 |
| 355 |
CALL DORGR2( M, M, N1, IR, LDST, TAUR, WORK, LINFO ) |
| 356 |
IF( LINFO.NE.0 ) |
| 357 |
$ GO TO 70 |
| 358 |
* |
| 359 |
* Perform the swapping tentatively: |
| 360 |
* |
| 361 |
CALL DGEMM( 'T', 'N', M, M, M, ONE, LI, LDST, S, LDST, ZERO, |
| 362 |
$ WORK, M ) |
| 363 |
CALL DGEMM( 'N', 'T', M, M, M, ONE, WORK, M, IR, LDST, ZERO, S, |
| 364 |
$ LDST ) |
| 365 |
CALL DGEMM( 'T', 'N', M, M, M, ONE, LI, LDST, T, LDST, ZERO, |
| 366 |
$ WORK, M ) |
| 367 |
CALL DGEMM( 'N', 'T', M, M, M, ONE, WORK, M, IR, LDST, ZERO, T, |
| 368 |
$ LDST ) |
| 369 |
CALL DLACPY( 'F', M, M, S, LDST, SCPY, LDST ) |
| 370 |
CALL DLACPY( 'F', M, M, T, LDST, TCPY, LDST ) |
| 371 |
CALL DLACPY( 'F', M, M, IR, LDST, IRCOP, LDST ) |
| 372 |
CALL DLACPY( 'F', M, M, LI, LDST, LICOP, LDST ) |
| 373 |
* |
| 374 |
* Triangularize the B-part by an RQ factorization. |
| 375 |
* Apply transformation (from left) to A-part, giving S. |
| 376 |
* |
| 377 |
CALL DGERQ2( M, M, T, LDST, TAUR, WORK, LINFO ) |
| 378 |
IF( LINFO.NE.0 ) |
| 379 |
$ GO TO 70 |
| 380 |
CALL DORMR2( 'R', 'T', M, M, M, T, LDST, TAUR, S, LDST, WORK, |
| 381 |
$ LINFO ) |
| 382 |
IF( LINFO.NE.0 ) |
| 383 |
$ GO TO 70 |
| 384 |
CALL DORMR2( 'L', 'N', M, M, M, T, LDST, TAUR, IR, LDST, WORK, |
| 385 |
$ LINFO ) |
| 386 |
IF( LINFO.NE.0 ) |
| 387 |
$ GO TO 70 |
| 388 |
* |
| 389 |
* Compute F-norm(S21) in BRQA21. (T21 is 0.) |
| 390 |
* |
| 391 |
DSCALE = ZERO |
| 392 |
DSUM = ONE |
| 393 |
DO 30 I = 1, N2 |
| 394 |
CALL DLASSQ( N1, S( N2+1, I ), 1, DSCALE, DSUM ) |
| 395 |
30 CONTINUE |
| 396 |
BRQA21 = DSCALE*SQRT( DSUM ) |
| 397 |
* |
| 398 |
* Triangularize the B-part by a QR factorization. |
| 399 |
* Apply transformation (from right) to A-part, giving S. |
| 400 |
* |
| 401 |
CALL DGEQR2( M, M, TCPY, LDST, TAUL, WORK, LINFO ) |
| 402 |
IF( LINFO.NE.0 ) |
| 403 |
$ GO TO 70 |
| 404 |
CALL DORM2R( 'L', 'T', M, M, M, TCPY, LDST, TAUL, SCPY, LDST, |
| 405 |
$ WORK, INFO ) |
| 406 |
CALL DORM2R( 'R', 'N', M, M, M, TCPY, LDST, TAUL, LICOP, LDST, |
| 407 |
$ WORK, INFO ) |
| 408 |
IF( LINFO.NE.0 ) |
| 409 |
$ GO TO 70 |
| 410 |
* |
| 411 |
* Compute F-norm(S21) in BQRA21. (T21 is 0.) |
| 412 |
* |
| 413 |
DSCALE = ZERO |
| 414 |
DSUM = ONE |
| 415 |
DO 40 I = 1, N2 |
| 416 |
CALL DLASSQ( N1, SCPY( N2+1, I ), 1, DSCALE, DSUM ) |
| 417 |
40 CONTINUE |
| 418 |
BQRA21 = DSCALE*SQRT( DSUM ) |
| 419 |
* |
| 420 |
* Decide which method to use. |
| 421 |
* Weak stability test: |
| 422 |
* F-norm(S21) <= O(EPS * F-norm((S, T))) |
| 423 |
* |
| 424 |
IF( BQRA21.LE.BRQA21 .AND. BQRA21.LE.THRESH ) THEN |
| 425 |
CALL DLACPY( 'F', M, M, SCPY, LDST, S, LDST ) |
| 426 |
CALL DLACPY( 'F', M, M, TCPY, LDST, T, LDST ) |
| 427 |
CALL DLACPY( 'F', M, M, IRCOP, LDST, IR, LDST ) |
| 428 |
CALL DLACPY( 'F', M, M, LICOP, LDST, LI, LDST ) |
| 429 |
ELSE IF( BRQA21.GE.THRESH ) THEN |
| 430 |
GO TO 70 |
| 431 |
END IF |
| 432 |
* |
| 433 |
* Set lower triangle of B-part to zero |
| 434 |
* |
| 435 |
DO 50 I = 2, M |
| 436 |
CALL DCOPY( M-I+1, ZERO, 0, T( I, I-1 ), 1 ) |
| 437 |
50 CONTINUE |
| 438 |
* |
| 439 |
IF( WANDS ) THEN |
| 440 |
* |
| 441 |
* Strong stability test: |
| 442 |
* F-norm((A-QL*S*QR', B-QL*T*QR')) <= O(EPS*F-norm((A,B))) |
| 443 |
* |
| 444 |
CALL DLACPY( 'Full', M, M, A( J1, J1 ), LDA, WORK( M*M+1 ), |
| 445 |
$ M ) |
| 446 |
CALL DGEMM( 'N', 'N', M, M, M, ONE, LI, LDST, S, LDST, ZERO, |
| 447 |
$ WORK, M ) |
| 448 |
CALL DGEMM( 'N', 'N', M, M, M, -ONE, WORK, M, IR, LDST, ONE, |
| 449 |
$ WORK( M*M+1 ), M ) |
| 450 |
DSCALE = ZERO |
| 451 |
DSUM = ONE |
| 452 |
CALL DLASSQ( M*M, WORK( M*M+1 ), 1, DSCALE, DSUM ) |
| 453 |
* |
| 454 |
CALL DLACPY( 'Full', M, M, B( J1, J1 ), LDB, WORK( M*M+1 ), |
| 455 |
$ M ) |
| 456 |
CALL DGEMM( 'N', 'N', M, M, M, ONE, LI, LDST, T, LDST, ZERO, |
| 457 |
$ WORK, M ) |
| 458 |
CALL DGEMM( 'N', 'N', M, M, M, -ONE, WORK, M, IR, LDST, ONE, |
| 459 |
$ WORK( M*M+1 ), M ) |
| 460 |
CALL DLASSQ( M*M, WORK( M*M+1 ), 1, DSCALE, DSUM ) |
| 461 |
SS = DSCALE*SQRT( DSUM ) |
| 462 |
DTRONG = ( SS.LE.THRESH ) |
| 463 |
IF( .NOT.DTRONG ) |
| 464 |
$ GO TO 70 |
| 465 |
* |
| 466 |
END IF |
| 467 |
* |
| 468 |
* If the swap is accepted ("weakly" and "strongly"), apply the |
| 469 |
* transformations and set N1-by-N2 (2,1)-block to zero. |
| 470 |
* |
| 471 |
DO 60 I = 1, N2 |
| 472 |
CALL DCOPY( N1, ZERO, 0, S( N2+1, I ), 1 ) |
| 473 |
60 CONTINUE |
| 474 |
* |
| 475 |
* copy back M-by-M diagonal block starting at index J1 of (A, B) |
| 476 |
* |
| 477 |
CALL DLACPY( 'F', M, M, S, LDST, A( J1, J1 ), LDA ) |
| 478 |
CALL DLACPY( 'F', M, M, T, LDST, B( J1, J1 ), LDB ) |
| 479 |
CALL DCOPY( LDST*LDST, ZERO, 0, T, 1 ) |
| 480 |
* |
| 481 |
* Standardize existing 2-by-2 blocks. |
| 482 |
* |
| 483 |
CALL DCOPY( M*M, ZERO, 0, WORK, 1 ) |
| 484 |
WORK( 1 ) = ONE |
| 485 |
T( 1, 1 ) = ONE |
| 486 |
IDUM = LWORK - M*M - 2 |
| 487 |
IF( N2.GT.1 ) THEN |
| 488 |
CALL DLAGV2( A( J1, J1 ), LDA, B( J1, J1 ), LDB, AR, AI, BE, |
| 489 |
$ WORK( 1 ), WORK( 2 ), T( 1, 1 ), T( 2, 1 ) ) |
| 490 |
WORK( M+1 ) = -WORK( 2 ) |
| 491 |
WORK( M+2 ) = WORK( 1 ) |
| 492 |
T( N2, N2 ) = T( 1, 1 ) |
| 493 |
T( 1, 2 ) = -T( 2, 1 ) |
| 494 |
END IF |
| 495 |
WORK( M*M ) = ONE |
| 496 |
T( M, M ) = ONE |
| 497 |
* |
| 498 |
IF( N1.GT.1 ) THEN |
| 499 |
CALL DLAGV2( A( J1+N2, J1+N2 ), LDA, B( J1+N2, J1+N2 ), LDB, |
| 500 |
$ TAUR, TAUL, WORK( M*M+1 ), WORK( N2*M+N2+1 ), |
| 501 |
$ WORK( N2*M+N2+2 ), T( N2+1, N2+1 ), |
| 502 |
$ T( M, M-1 ) ) |
| 503 |
WORK( M*M ) = WORK( N2*M+N2+1 ) |
| 504 |
WORK( M*M-1 ) = -WORK( N2*M+N2+2 ) |
| 505 |
T( M, M ) = T( N2+1, N2+1 ) |
| 506 |
T( M-1, M ) = -T( M, M-1 ) |
| 507 |
END IF |
| 508 |
CALL DGEMM( 'T', 'N', N2, N1, N2, ONE, WORK, M, A( J1, J1+N2 ), |
| 509 |
$ LDA, ZERO, WORK( M*M+1 ), N2 ) |
| 510 |
CALL DLACPY( 'Full', N2, N1, WORK( M*M+1 ), N2, A( J1, J1+N2 ), |
| 511 |
$ LDA ) |
| 512 |
CALL DGEMM( 'T', 'N', N2, N1, N2, ONE, WORK, M, B( J1, J1+N2 ), |
| 513 |
$ LDB, ZERO, WORK( M*M+1 ), N2 ) |
| 514 |
CALL DLACPY( 'Full', N2, N1, WORK( M*M+1 ), N2, B( J1, J1+N2 ), |
| 515 |
$ LDB ) |
| 516 |
CALL DGEMM( 'N', 'N', M, M, M, ONE, LI, LDST, WORK, M, ZERO, |
| 517 |
$ WORK( M*M+1 ), M ) |
| 518 |
CALL DLACPY( 'Full', M, M, WORK( M*M+1 ), M, LI, LDST ) |
| 519 |
CALL DGEMM( 'N', 'N', N2, N1, N1, ONE, A( J1, J1+N2 ), LDA, |
| 520 |
$ T( N2+1, N2+1 ), LDST, ZERO, WORK, N2 ) |
| 521 |
CALL DLACPY( 'Full', N2, N1, WORK, N2, A( J1, J1+N2 ), LDA ) |
| 522 |
CALL DGEMM( 'N', 'N', N2, N1, N1, ONE, B( J1, J1+N2 ), LDA, |
| 523 |
$ T( N2+1, N2+1 ), LDST, ZERO, WORK, N2 ) |
| 524 |
CALL DLACPY( 'Full', N2, N1, WORK, N2, B( J1, J1+N2 ), LDB ) |
| 525 |
CALL DGEMM( 'T', 'N', M, M, M, ONE, IR, LDST, T, LDST, ZERO, |
| 526 |
$ WORK, M ) |
| 527 |
CALL DLACPY( 'Full', M, M, WORK, M, IR, LDST ) |
| 528 |
* |
| 529 |
* Accumulate transformations into Q and Z if requested. |
| 530 |
* |
| 531 |
IF( WANTQ ) THEN |
| 532 |
CALL DGEMM( 'N', 'N', N, M, M, ONE, Q( 1, J1 ), LDQ, LI, |
| 533 |
$ LDST, ZERO, WORK, N ) |
| 534 |
CALL DLACPY( 'Full', N, M, WORK, N, Q( 1, J1 ), LDQ ) |
| 535 |
* |
| 536 |
END IF |
| 537 |
* |
| 538 |
IF( WANTZ ) THEN |
| 539 |
CALL DGEMM( 'N', 'N', N, M, M, ONE, Z( 1, J1 ), LDZ, IR, |
| 540 |
$ LDST, ZERO, WORK, N ) |
| 541 |
CALL DLACPY( 'Full', N, M, WORK, N, Z( 1, J1 ), LDZ ) |
| 542 |
* |
| 543 |
END IF |
| 544 |
* |
| 545 |
* Update (A(J1:J1+M-1, M+J1:N), B(J1:J1+M-1, M+J1:N)) and |
| 546 |
* (A(1:J1-1, J1:J1+M), B(1:J1-1, J1:J1+M)). |
| 547 |
* |
| 548 |
I = J1 + M |
| 549 |
IF( I.LE.N ) THEN |
| 550 |
CALL DGEMM( 'T', 'N', M, N-I+1, M, ONE, LI, LDST, |
| 551 |
$ A( J1, I ), LDA, ZERO, WORK, M ) |
| 552 |
CALL DLACPY( 'Full', M, N-I+1, WORK, M, A( J1, I ), LDA ) |
| 553 |
CALL DGEMM( 'T', 'N', M, N-I+1, M, ONE, LI, LDST, |
| 554 |
$ B( J1, I ), LDA, ZERO, WORK, M ) |
| 555 |
CALL DLACPY( 'Full', M, N-I+1, WORK, M, B( J1, I ), LDA ) |
| 556 |
END IF |
| 557 |
I = J1 - 1 |
| 558 |
IF( I.GT.0 ) THEN |
| 559 |
CALL DGEMM( 'N', 'N', I, M, M, ONE, A( 1, J1 ), LDA, IR, |
| 560 |
$ LDST, ZERO, WORK, I ) |
| 561 |
CALL DLACPY( 'Full', I, M, WORK, I, A( 1, J1 ), LDA ) |
| 562 |
CALL DGEMM( 'N', 'N', I, M, M, ONE, B( 1, J1 ), LDB, IR, |
| 563 |
$ LDST, ZERO, WORK, I ) |
| 564 |
CALL DLACPY( 'Full', I, M, WORK, I, B( 1, J1 ), LDB ) |
| 565 |
END IF |
| 566 |
* |
| 567 |
* Exit with INFO = 0 if swap was successfully performed. |
| 568 |
* |
| 569 |
RETURN |
| 570 |
* |
| 571 |
END IF |
| 572 |
* |
| 573 |
* Exit with INFO = 1 if swap was rejected. |
| 574 |
* |
| 575 |
70 CONTINUE |
| 576 |
* |
| 577 |
INFO = 1 |
| 578 |
RETURN |
| 579 |
* |
| 580 |
* End of DTGEX2 |
| 581 |
* |
| 582 |
END |