| 1 |
c----------------------------------------------------------------------- |
| 2 |
c\BeginDoc |
| 3 |
c |
| 4 |
c\Name: dsaitr |
| 5 |
c |
| 6 |
c\Description: |
| 7 |
c Reverse communication interface for applying NP additional steps to |
| 8 |
c a K step symmetric Arnoldi factorization. |
| 9 |
c |
| 10 |
c Input: OP*V_{k} - V_{k}*H = r_{k}*e_{k}^T |
| 11 |
c |
| 12 |
c with (V_{k}^T)*B*V_{k} = I, (V_{k}^T)*B*r_{k} = 0. |
| 13 |
c |
| 14 |
c Output: OP*V_{k+p} - V_{k+p}*H = r_{k+p}*e_{k+p}^T |
| 15 |
c |
| 16 |
c with (V_{k+p}^T)*B*V_{k+p} = I, (V_{k+p}^T)*B*r_{k+p} = 0. |
| 17 |
c |
| 18 |
c where OP and B are as in dsaupd. The B-norm of r_{k+p} is also |
| 19 |
c computed and returned. |
| 20 |
c |
| 21 |
c\Usage: |
| 22 |
c call dsaitr |
| 23 |
c ( IDO, BMAT, N, K, NP, MODE, RESID, RNORM, V, LDV, H, LDH, |
| 24 |
c IPNTR, WORKD, INFO ) |
| 25 |
c |
| 26 |
c\Arguments |
| 27 |
c IDO Integer. (INPUT/OUTPUT) |
| 28 |
c Reverse communication flag. |
| 29 |
c ------------------------------------------------------------- |
| 30 |
c IDO = 0: first call to the reverse communication interface |
| 31 |
c IDO = -1: compute Y = OP * X where |
| 32 |
c IPNTR(1) is the pointer into WORK for X, |
| 33 |
c IPNTR(2) is the pointer into WORK for Y. |
| 34 |
c This is for the restart phase to force the new |
| 35 |
c starting vector into the range of OP. |
| 36 |
c IDO = 1: compute Y = OP * X where |
| 37 |
c IPNTR(1) is the pointer into WORK for X, |
| 38 |
c IPNTR(2) is the pointer into WORK for Y, |
| 39 |
c IPNTR(3) is the pointer into WORK for B * X. |
| 40 |
c IDO = 2: compute Y = B * X where |
| 41 |
c IPNTR(1) is the pointer into WORK for X, |
| 42 |
c IPNTR(2) is the pointer into WORK for Y. |
| 43 |
c IDO = 99: done |
| 44 |
c ------------------------------------------------------------- |
| 45 |
c When the routine is used in the "shift-and-invert" mode, the |
| 46 |
c vector B * Q is already available and does not need to be |
| 47 |
c recomputed in forming OP * Q. |
| 48 |
c |
| 49 |
c BMAT Character*1. (INPUT) |
| 50 |
c BMAT specifies the type of matrix B that defines the |
| 51 |
c semi-inner product for the operator OP. See dsaupd. |
| 52 |
c B = 'I' -> standard eigenvalue problem A*x = lambda*x |
| 53 |
c B = 'G' -> generalized eigenvalue problem A*x = lambda*M*x |
| 54 |
c |
| 55 |
c N Integer. (INPUT) |
| 56 |
c Dimension of the eigenproblem. |
| 57 |
c |
| 58 |
c K Integer. (INPUT) |
| 59 |
c Current order of H and the number of columns of V. |
| 60 |
c |
| 61 |
c NP Integer. (INPUT) |
| 62 |
c Number of additional Arnoldi steps to take. |
| 63 |
c |
| 64 |
c MODE Integer. (INPUT) |
| 65 |
c Signifies which form for "OP". If MODE=2 then |
| 66 |
c a reduction in the number of B matrix vector multiplies |
| 67 |
c is possible since the B-norm of OP*x is equivalent to |
| 68 |
c the inv(B)-norm of A*x. |
| 69 |
c |
| 70 |
c RESID Double precision array of length N. (INPUT/OUTPUT) |
| 71 |
c On INPUT: RESID contains the residual vector r_{k}. |
| 72 |
c On OUTPUT: RESID contains the residual vector r_{k+p}. |
| 73 |
c |
| 74 |
c RNORM Double precision scalar. (INPUT/OUTPUT) |
| 75 |
c On INPUT the B-norm of r_{k}. |
| 76 |
c On OUTPUT the B-norm of the updated residual r_{k+p}. |
| 77 |
c |
| 78 |
c V Double precision N by K+NP array. (INPUT/OUTPUT) |
| 79 |
c On INPUT: V contains the Arnoldi vectors in the first K |
| 80 |
c columns. |
| 81 |
c On OUTPUT: V contains the new NP Arnoldi vectors in the next |
| 82 |
c NP columns. The first K columns are unchanged. |
| 83 |
c |
| 84 |
c LDV Integer. (INPUT) |
| 85 |
c Leading dimension of V exactly as declared in the calling |
| 86 |
c program. |
| 87 |
c |
| 88 |
c H Double precision (K+NP) by 2 array. (INPUT/OUTPUT) |
| 89 |
c H is used to store the generated symmetric tridiagonal matrix |
| 90 |
c with the subdiagonal in the first column starting at H(2,1) |
| 91 |
c and the main diagonal in the second column. |
| 92 |
c |
| 93 |
c LDH Integer. (INPUT) |
| 94 |
c Leading dimension of H exactly as declared in the calling |
| 95 |
c program. |
| 96 |
c |
| 97 |
c IPNTR Integer array of length 3. (OUTPUT) |
| 98 |
c Pointer to mark the starting locations in the WORK for |
| 99 |
c vectors used by the Arnoldi iteration. |
| 100 |
c ------------------------------------------------------------- |
| 101 |
c IPNTR(1): pointer to the current operand vector X. |
| 102 |
c IPNTR(2): pointer to the current result vector Y. |
| 103 |
c IPNTR(3): pointer to the vector B * X when used in the |
| 104 |
c shift-and-invert mode. X is the current operand. |
| 105 |
c ------------------------------------------------------------- |
| 106 |
c |
| 107 |
c WORKD Double precision work array of length 3*N. (REVERSE COMMUNICATION) |
| 108 |
c Distributed array to be used in the basic Arnoldi iteration |
| 109 |
c for reverse communication. The calling program should not |
| 110 |
c use WORKD as temporary workspace during the iteration !!!!!! |
| 111 |
c On INPUT, WORKD(1:N) = B*RESID where RESID is associated |
| 112 |
c with the K step Arnoldi factorization. Used to save some |
| 113 |
c computation at the first step. |
| 114 |
c On OUTPUT, WORKD(1:N) = B*RESID where RESID is associated |
| 115 |
c with the K+NP step Arnoldi factorization. |
| 116 |
c |
| 117 |
c INFO Integer. (OUTPUT) |
| 118 |
c = 0: Normal exit. |
| 119 |
c > 0: Size of an invariant subspace of OP is found that is |
| 120 |
c less than K + NP. |
| 121 |
c |
| 122 |
c\EndDoc |
| 123 |
c |
| 124 |
c----------------------------------------------------------------------- |
| 125 |
c |
| 126 |
c\BeginLib |
| 127 |
c |
| 128 |
c\Local variables: |
| 129 |
c xxxxxx real |
| 130 |
c |
| 131 |
c\Routines called: |
| 132 |
c dgetv0 ARPACK routine to generate the initial vector. |
| 133 |
c ivout ARPACK utility routine that prints integers. |
| 134 |
c dmout ARPACK utility routine that prints matrices. |
| 135 |
c dlamch LAPACK routine that determines machine constants. |
| 136 |
c dlascl LAPACK routine for careful scaling of a matrix. |
| 137 |
c dgemv Level 2 BLAS routine for matrix vector multiplication. |
| 138 |
c daxpy Level 1 BLAS that computes a vector triad. |
| 139 |
c dscal Level 1 BLAS that scales a vector. |
| 140 |
c dcopy Level 1 BLAS that copies one vector to another . |
| 141 |
c ddot Level 1 BLAS that computes the scalar product of two vectors. |
| 142 |
c dnrm2 Level 1 BLAS that computes the norm of a vector. |
| 143 |
c |
| 144 |
c\Author |
| 145 |
c Danny Sorensen Phuong Vu |
| 146 |
c Richard Lehoucq CRPC / Rice University |
| 147 |
c Dept. of Computational & Houston, Texas |
| 148 |
c Applied Mathematics |
| 149 |
c Rice University |
| 150 |
c Houston, Texas |
| 151 |
c |
| 152 |
c\Revision history: |
| 153 |
c xx/xx/93: Version ' 2.4' |
| 154 |
c |
| 155 |
c\SCCS Information: @(#) |
| 156 |
c FILE: saitr.F SID: 2.6 DATE OF SID: 8/28/96 RELEASE: 2 |
| 157 |
c |
| 158 |
c\Remarks |
| 159 |
c The algorithm implemented is: |
| 160 |
c |
| 161 |
c restart = .false. |
| 162 |
c Given V_{k} = [v_{1}, ..., v_{k}], r_{k}; |
| 163 |
c r_{k} contains the initial residual vector even for k = 0; |
| 164 |
c Also assume that rnorm = || B*r_{k} || and B*r_{k} are already |
| 165 |
c computed by the calling program. |
| 166 |
c |
| 167 |
c betaj = rnorm ; p_{k+1} = B*r_{k} ; |
| 168 |
c For j = k+1, ..., k+np Do |
| 169 |
c 1) if ( betaj < tol ) stop or restart depending on j. |
| 170 |
c if ( restart ) generate a new starting vector. |
| 171 |
c 2) v_{j} = r(j-1)/betaj; V_{j} = [V_{j-1}, v_{j}]; |
| 172 |
c p_{j} = p_{j}/betaj |
| 173 |
c 3) r_{j} = OP*v_{j} where OP is defined as in dsaupd |
| 174 |
c For shift-invert mode p_{j} = B*v_{j} is already available. |
| 175 |
c wnorm = || OP*v_{j} || |
| 176 |
c 4) Compute the j-th step residual vector. |
| 177 |
c w_{j} = V_{j}^T * B * OP * v_{j} |
| 178 |
c r_{j} = OP*v_{j} - V_{j} * w_{j} |
| 179 |
c alphaj <- j-th component of w_{j} |
| 180 |
c rnorm = || r_{j} || |
| 181 |
c betaj+1 = rnorm |
| 182 |
c If (rnorm > 0.717*wnorm) accept step and go back to 1) |
| 183 |
c 5) Re-orthogonalization step: |
| 184 |
c s = V_{j}'*B*r_{j} |
| 185 |
c r_{j} = r_{j} - V_{j}*s; rnorm1 = || r_{j} || |
| 186 |
c alphaj = alphaj + s_{j}; |
| 187 |
c 6) Iterative refinement step: |
| 188 |
c If (rnorm1 > 0.717*rnorm) then |
| 189 |
c rnorm = rnorm1 |
| 190 |
c accept step and go back to 1) |
| 191 |
c Else |
| 192 |
c rnorm = rnorm1 |
| 193 |
c If this is the first time in step 6), go to 5) |
| 194 |
c Else r_{j} lies in the span of V_{j} numerically. |
| 195 |
c Set r_{j} = 0 and rnorm = 0; go to 1) |
| 196 |
c EndIf |
| 197 |
c End Do |
| 198 |
c |
| 199 |
c\EndLib |
| 200 |
c |
| 201 |
c----------------------------------------------------------------------- |
| 202 |
c |
| 203 |
subroutine dsaitr |
| 204 |
& (ido, bmat, n, k, np, mode, resid, rnorm, v, ldv, h, ldh, |
| 205 |
& ipntr, workd, info) |
| 206 |
c |
| 207 |
c %----------------------------------------------------% |
| 208 |
c | Include files for debugging and timing information | |
| 209 |
c %----------------------------------------------------% |
| 210 |
c |
| 211 |
include 'debug.h' |
| 212 |
include 'stat.h' |
| 213 |
c |
| 214 |
c %------------------% |
| 215 |
c | Scalar Arguments | |
| 216 |
c %------------------% |
| 217 |
c |
| 218 |
character bmat*1 |
| 219 |
integer ido, info, k, ldh, ldv, n, mode, np |
| 220 |
Double precision |
| 221 |
& rnorm |
| 222 |
c |
| 223 |
c %-----------------% |
| 224 |
c | Array Arguments | |
| 225 |
c %-----------------% |
| 226 |
c |
| 227 |
integer ipntr(3) |
| 228 |
Double precision |
| 229 |
& h(ldh,2), resid(n), v(ldv,k+np), workd(3*n) |
| 230 |
c |
| 231 |
c %------------% |
| 232 |
c | Parameters | |
| 233 |
c %------------% |
| 234 |
c |
| 235 |
Double precision |
| 236 |
& one, zero |
| 237 |
parameter (one = 1.0D+0, zero = 0.0D+0) |
| 238 |
c |
| 239 |
c %---------------% |
| 240 |
c | Local Scalars | |
| 241 |
c %---------------% |
| 242 |
c |
| 243 |
logical first, orth1, orth2, rstart, step3, step4 |
| 244 |
integer i, ierr, ipj, irj, ivj, iter, itry, j, msglvl, |
| 245 |
& infol, jj |
| 246 |
Double precision |
| 247 |
& rnorm1, wnorm, safmin, temp1 |
| 248 |
save orth1, orth2, rstart, step3, step4, |
| 249 |
& ierr, ipj, irj, ivj, iter, itry, j, msglvl, |
| 250 |
& rnorm1, safmin, wnorm |
| 251 |
c |
| 252 |
c %-----------------------% |
| 253 |
c | Local Array Arguments | |
| 254 |
c %-----------------------% |
| 255 |
c |
| 256 |
Double precision |
| 257 |
& xtemp(2) |
| 258 |
c |
| 259 |
c %----------------------% |
| 260 |
c | External Subroutines | |
| 261 |
c %----------------------% |
| 262 |
c |
| 263 |
external daxpy, dcopy, dscal, dgemv, dgetv0, dmout, |
| 264 |
& dlascl, second |
| 265 |
c |
| 266 |
c %--------------------% |
| 267 |
c | External Functions | |
| 268 |
c %--------------------% |
| 269 |
c |
| 270 |
Double precision |
| 271 |
& ddot, dnrm2, dlamch |
| 272 |
external ddot, dnrm2, dlamch |
| 273 |
c |
| 274 |
c %-----------------% |
| 275 |
c | Data statements | |
| 276 |
c %-----------------% |
| 277 |
c |
| 278 |
data first / .true. / |
| 279 |
c |
| 280 |
c %-----------------------% |
| 281 |
c | Executable Statements | |
| 282 |
c %-----------------------% |
| 283 |
c |
| 284 |
if (first) then |
| 285 |
first = .false. |
| 286 |
c |
| 287 |
c %--------------------------------% |
| 288 |
c | safmin = safe minimum is such | |
| 289 |
c | that 1/sfmin does not overflow | |
| 290 |
c %--------------------------------% |
| 291 |
c |
| 292 |
safmin = dlamch('safmin') |
| 293 |
end if |
| 294 |
c |
| 295 |
if (ido .eq. 0) then |
| 296 |
c |
| 297 |
c %-------------------------------% |
| 298 |
c | Initialize timing statistics | |
| 299 |
c | & message level for debugging | |
| 300 |
c %-------------------------------% |
| 301 |
c |
| 302 |
call second (t0) |
| 303 |
msglvl = msaitr |
| 304 |
c |
| 305 |
c %------------------------------% |
| 306 |
c | Initial call to this routine | |
| 307 |
c %------------------------------% |
| 308 |
c |
| 309 |
info = 0 |
| 310 |
step3 = .false. |
| 311 |
step4 = .false. |
| 312 |
rstart = .false. |
| 313 |
orth1 = .false. |
| 314 |
orth2 = .false. |
| 315 |
c |
| 316 |
c %--------------------------------% |
| 317 |
c | Pointer to the current step of | |
| 318 |
c | the factorization to build | |
| 319 |
c %--------------------------------% |
| 320 |
c |
| 321 |
j = k + 1 |
| 322 |
c |
| 323 |
c %------------------------------------------% |
| 324 |
c | Pointers used for reverse communication | |
| 325 |
c | when using WORKD. | |
| 326 |
c %------------------------------------------% |
| 327 |
c |
| 328 |
ipj = 1 |
| 329 |
irj = ipj + n |
| 330 |
ivj = irj + n |
| 331 |
end if |
| 332 |
c |
| 333 |
c %-------------------------------------------------% |
| 334 |
c | When in reverse communication mode one of: | |
| 335 |
c | STEP3, STEP4, ORTH1, ORTH2, RSTART | |
| 336 |
c | will be .true. | |
| 337 |
c | STEP3: return from computing OP*v_{j}. | |
| 338 |
c | STEP4: return from computing B-norm of OP*v_{j} | |
| 339 |
c | ORTH1: return from computing B-norm of r_{j+1} | |
| 340 |
c | ORTH2: return from computing B-norm of | |
| 341 |
c | correction to the residual vector. | |
| 342 |
c | RSTART: return from OP computations needed by | |
| 343 |
c | dgetv0. | |
| 344 |
c %-------------------------------------------------% |
| 345 |
c |
| 346 |
if (step3) go to 50 |
| 347 |
if (step4) go to 60 |
| 348 |
if (orth1) go to 70 |
| 349 |
if (orth2) go to 90 |
| 350 |
if (rstart) go to 30 |
| 351 |
c |
| 352 |
c %------------------------------% |
| 353 |
c | Else this is the first step. | |
| 354 |
c %------------------------------% |
| 355 |
c |
| 356 |
c %--------------------------------------------------------------% |
| 357 |
c | | |
| 358 |
c | A R N O L D I I T E R A T I O N L O O P | |
| 359 |
c | | |
| 360 |
c | Note: B*r_{j-1} is already in WORKD(1:N)=WORKD(IPJ:IPJ+N-1) | |
| 361 |
c %--------------------------------------------------------------% |
| 362 |
c |
| 363 |
1000 continue |
| 364 |
c |
| 365 |
c if (msglvl .gt. 2) then |
| 366 |
c call ivout (logfil, 1, j, ndigit, |
| 367 |
c & '_saitr: generating Arnoldi vector no.') |
| 368 |
c call dvout (logfil, 1, rnorm, ndigit, |
| 369 |
c & '_saitr: B-norm of the current residual =') |
| 370 |
c end if |
| 371 |
c |
| 372 |
c %---------------------------------------------------------% |
| 373 |
c | Check for exact zero. Equivalent to determining whether | |
| 374 |
c | a j-step Arnoldi factorization is present. | |
| 375 |
c %---------------------------------------------------------% |
| 376 |
c |
| 377 |
if (rnorm .gt. zero) go to 40 |
| 378 |
c |
| 379 |
c %---------------------------------------------------% |
| 380 |
c | Invariant subspace found, generate a new starting | |
| 381 |
c | vector which is orthogonal to the current Arnoldi | |
| 382 |
c | basis and continue the iteration. | |
| 383 |
c %---------------------------------------------------% |
| 384 |
c |
| 385 |
c if (msglvl .gt. 0) then |
| 386 |
c call ivout (logfil, 1, j, ndigit, |
| 387 |
c & '_saitr: ****** restart at step ******') |
| 388 |
c end if |
| 389 |
c |
| 390 |
c %---------------------------------------------% |
| 391 |
c | ITRY is the loop variable that controls the | |
| 392 |
c | maximum amount of times that a restart is | |
| 393 |
c | attempted. NRSTRT is used by stat.h | |
| 394 |
c %---------------------------------------------% |
| 395 |
c |
| 396 |
nrstrt = nrstrt + 1 |
| 397 |
itry = 1 |
| 398 |
20 continue |
| 399 |
rstart = .true. |
| 400 |
ido = 0 |
| 401 |
30 continue |
| 402 |
c |
| 403 |
c %--------------------------------------% |
| 404 |
c | If in reverse communication mode and | |
| 405 |
c | RSTART = .true. flow returns here. | |
| 406 |
c %--------------------------------------% |
| 407 |
c |
| 408 |
call dgetv0 (ido, bmat, itry, .false., n, j, v, ldv, |
| 409 |
& resid, rnorm, ipntr, workd, ierr) |
| 410 |
if (ido .ne. 99) go to 9000 |
| 411 |
if (ierr .lt. 0) then |
| 412 |
itry = itry + 1 |
| 413 |
if (itry .le. 3) go to 20 |
| 414 |
c |
| 415 |
c %------------------------------------------------% |
| 416 |
c | Give up after several restart attempts. | |
| 417 |
c | Set INFO to the size of the invariant subspace | |
| 418 |
c | which spans OP and exit. | |
| 419 |
c %------------------------------------------------% |
| 420 |
c |
| 421 |
info = j - 1 |
| 422 |
call second (t1) |
| 423 |
tsaitr = tsaitr + (t1 - t0) |
| 424 |
ido = 99 |
| 425 |
go to 9000 |
| 426 |
end if |
| 427 |
c |
| 428 |
40 continue |
| 429 |
c |
| 430 |
c %---------------------------------------------------------% |
| 431 |
c | STEP 2: v_{j} = r_{j-1}/rnorm and p_{j} = p_{j}/rnorm | |
| 432 |
c | Note that p_{j} = B*r_{j-1}. In order to avoid overflow | |
| 433 |
c | when reciprocating a small RNORM, test against lower | |
| 434 |
c | machine bound. | |
| 435 |
c %---------------------------------------------------------% |
| 436 |
c |
| 437 |
call dcopy (n, resid, 1, v(1,j), 1) |
| 438 |
if (rnorm .ge. safmin) then |
| 439 |
temp1 = one / rnorm |
| 440 |
call dscal (n, temp1, v(1,j), 1) |
| 441 |
call dscal (n, temp1, workd(ipj), 1) |
| 442 |
else |
| 443 |
c |
| 444 |
c %-----------------------------------------% |
| 445 |
c | To scale both v_{j} and p_{j} carefully | |
| 446 |
c | use LAPACK routine SLASCL | |
| 447 |
c %-----------------------------------------% |
| 448 |
c |
| 449 |
call dlascl ('General', i, i, rnorm, one, n, 1, |
| 450 |
& v(1,j), n, infol) |
| 451 |
call dlascl ('General', i, i, rnorm, one, n, 1, |
| 452 |
& workd(ipj), n, infol) |
| 453 |
end if |
| 454 |
c |
| 455 |
c %------------------------------------------------------% |
| 456 |
c | STEP 3: r_{j} = OP*v_{j}; Note that p_{j} = B*v_{j} | |
| 457 |
c | Note that this is not quite yet r_{j}. See STEP 4 | |
| 458 |
c %------------------------------------------------------% |
| 459 |
c |
| 460 |
step3 = .true. |
| 461 |
nopx = nopx + 1 |
| 462 |
call second (t2) |
| 463 |
call dcopy (n, v(1,j), 1, workd(ivj), 1) |
| 464 |
ipntr(1) = ivj |
| 465 |
ipntr(2) = irj |
| 466 |
ipntr(3) = ipj |
| 467 |
ido = 1 |
| 468 |
c |
| 469 |
c %-----------------------------------% |
| 470 |
c | Exit in order to compute OP*v_{j} | |
| 471 |
c %-----------------------------------% |
| 472 |
c |
| 473 |
go to 9000 |
| 474 |
50 continue |
| 475 |
c |
| 476 |
c %-----------------------------------% |
| 477 |
c | Back from reverse communication; | |
| 478 |
c | WORKD(IRJ:IRJ+N-1) := OP*v_{j}. | |
| 479 |
c %-----------------------------------% |
| 480 |
c |
| 481 |
call second (t3) |
| 482 |
tmvopx = tmvopx + (t3 - t2) |
| 483 |
c |
| 484 |
step3 = .false. |
| 485 |
c |
| 486 |
c %------------------------------------------% |
| 487 |
c | Put another copy of OP*v_{j} into RESID. | |
| 488 |
c %------------------------------------------% |
| 489 |
c |
| 490 |
call dcopy (n, workd(irj), 1, resid, 1) |
| 491 |
c |
| 492 |
c %-------------------------------------------% |
| 493 |
c | STEP 4: Finish extending the symmetric | |
| 494 |
c | Arnoldi to length j. If MODE = 2 | |
| 495 |
c | then B*OP = B*inv(B)*A = A and | |
| 496 |
c | we don't need to compute B*OP. | |
| 497 |
c | NOTE: If MODE = 2 WORKD(IVJ:IVJ+N-1) is | |
| 498 |
c | assumed to have A*v_{j}. | |
| 499 |
c %-------------------------------------------% |
| 500 |
c |
| 501 |
if (mode .eq. 2) go to 65 |
| 502 |
call second (t2) |
| 503 |
if (bmat .eq. 'G') then |
| 504 |
nbx = nbx + 1 |
| 505 |
step4 = .true. |
| 506 |
ipntr(1) = irj |
| 507 |
ipntr(2) = ipj |
| 508 |
ido = 2 |
| 509 |
c |
| 510 |
c %-------------------------------------% |
| 511 |
c | Exit in order to compute B*OP*v_{j} | |
| 512 |
c %-------------------------------------% |
| 513 |
c |
| 514 |
go to 9000 |
| 515 |
else if (bmat .eq. 'I') then |
| 516 |
call dcopy(n, resid, 1 , workd(ipj), 1) |
| 517 |
end if |
| 518 |
60 continue |
| 519 |
c |
| 520 |
c %-----------------------------------% |
| 521 |
c | Back from reverse communication; | |
| 522 |
c | WORKD(IPJ:IPJ+N-1) := B*OP*v_{j}. | |
| 523 |
c %-----------------------------------% |
| 524 |
c |
| 525 |
if (bmat .eq. 'G') then |
| 526 |
call second (t3) |
| 527 |
tmvbx = tmvbx + (t3 - t2) |
| 528 |
end if |
| 529 |
c |
| 530 |
step4 = .false. |
| 531 |
c |
| 532 |
c %-------------------------------------% |
| 533 |
c | The following is needed for STEP 5. | |
| 534 |
c | Compute the B-norm of OP*v_{j}. | |
| 535 |
c %-------------------------------------% |
| 536 |
c |
| 537 |
65 continue |
| 538 |
if (mode .eq. 2) then |
| 539 |
c |
| 540 |
c %----------------------------------% |
| 541 |
c | Note that the B-norm of OP*v_{j} | |
| 542 |
c | is the inv(B)-norm of A*v_{j}. | |
| 543 |
c %----------------------------------% |
| 544 |
c |
| 545 |
wnorm = ddot (n, resid, 1, workd(ivj), 1) |
| 546 |
wnorm = sqrt(abs(wnorm)) |
| 547 |
else if (bmat .eq. 'G') then |
| 548 |
wnorm = ddot (n, resid, 1, workd(ipj), 1) |
| 549 |
wnorm = sqrt(abs(wnorm)) |
| 550 |
else if (bmat .eq. 'I') then |
| 551 |
wnorm = dnrm2(n, resid, 1) |
| 552 |
end if |
| 553 |
c |
| 554 |
c %-----------------------------------------% |
| 555 |
c | Compute the j-th residual corresponding | |
| 556 |
c | to the j step factorization. | |
| 557 |
c | Use Classical Gram Schmidt and compute: | |
| 558 |
c | w_{j} <- V_{j}^T * B * OP * v_{j} | |
| 559 |
c | r_{j} <- OP*v_{j} - V_{j} * w_{j} | |
| 560 |
c %-----------------------------------------% |
| 561 |
c |
| 562 |
c |
| 563 |
c %------------------------------------------% |
| 564 |
c | Compute the j Fourier coefficients w_{j} | |
| 565 |
c | WORKD(IPJ:IPJ+N-1) contains B*OP*v_{j}. | |
| 566 |
c %------------------------------------------% |
| 567 |
c |
| 568 |
if (mode .ne. 2 ) then |
| 569 |
call dgemv('T', n, j, one, v, ldv, workd(ipj), 1, zero, |
| 570 |
& workd(irj), 1) |
| 571 |
else if (mode .eq. 2) then |
| 572 |
call dgemv('T', n, j, one, v, ldv, workd(ivj), 1, zero, |
| 573 |
& workd(irj), 1) |
| 574 |
end if |
| 575 |
c |
| 576 |
c %--------------------------------------% |
| 577 |
c | Orthgonalize r_{j} against V_{j}. | |
| 578 |
c | RESID contains OP*v_{j}. See STEP 3. | |
| 579 |
c %--------------------------------------% |
| 580 |
c |
| 581 |
call dgemv('N', n, j, -one, v, ldv, workd(irj), 1, one, |
| 582 |
& resid, 1) |
| 583 |
c |
| 584 |
c %--------------------------------------% |
| 585 |
c | Extend H to have j rows and columns. | |
| 586 |
c %--------------------------------------% |
| 587 |
c |
| 588 |
h(j,2) = workd(irj + j - 1) |
| 589 |
if (j .eq. 1 .or. rstart) then |
| 590 |
h(j,1) = zero |
| 591 |
else |
| 592 |
h(j,1) = rnorm |
| 593 |
end if |
| 594 |
call second (t4) |
| 595 |
c |
| 596 |
orth1 = .true. |
| 597 |
iter = 0 |
| 598 |
c |
| 599 |
call second (t2) |
| 600 |
if (bmat .eq. 'G') then |
| 601 |
nbx = nbx + 1 |
| 602 |
call dcopy (n, resid, 1, workd(irj), 1) |
| 603 |
ipntr(1) = irj |
| 604 |
ipntr(2) = ipj |
| 605 |
ido = 2 |
| 606 |
c |
| 607 |
c %----------------------------------% |
| 608 |
c | Exit in order to compute B*r_{j} | |
| 609 |
c %----------------------------------% |
| 610 |
c |
| 611 |
go to 9000 |
| 612 |
else if (bmat .eq. 'I') then |
| 613 |
call dcopy (n, resid, 1, workd(ipj), 1) |
| 614 |
end if |
| 615 |
70 continue |
| 616 |
c |
| 617 |
c %---------------------------------------------------% |
| 618 |
c | Back from reverse communication if ORTH1 = .true. | |
| 619 |
c | WORKD(IPJ:IPJ+N-1) := B*r_{j}. | |
| 620 |
c %---------------------------------------------------% |
| 621 |
c |
| 622 |
if (bmat .eq. 'G') then |
| 623 |
call second (t3) |
| 624 |
tmvbx = tmvbx + (t3 - t2) |
| 625 |
end if |
| 626 |
c |
| 627 |
orth1 = .false. |
| 628 |
c |
| 629 |
c %------------------------------% |
| 630 |
c | Compute the B-norm of r_{j}. | |
| 631 |
c %------------------------------% |
| 632 |
c |
| 633 |
if (bmat .eq. 'G') then |
| 634 |
rnorm = ddot (n, resid, 1, workd(ipj), 1) |
| 635 |
rnorm = sqrt(abs(rnorm)) |
| 636 |
else if (bmat .eq. 'I') then |
| 637 |
rnorm = dnrm2(n, resid, 1) |
| 638 |
end if |
| 639 |
c |
| 640 |
c %-----------------------------------------------------------% |
| 641 |
c | STEP 5: Re-orthogonalization / Iterative refinement phase | |
| 642 |
c | Maximum NITER_ITREF tries. | |
| 643 |
c | | |
| 644 |
c | s = V_{j}^T * B * r_{j} | |
| 645 |
c | r_{j} = r_{j} - V_{j}*s | |
| 646 |
c | alphaj = alphaj + s_{j} | |
| 647 |
c | | |
| 648 |
c | The stopping criteria used for iterative refinement is | |
| 649 |
c | discussed in Parlett's book SEP, page 107 and in Gragg & | |
| 650 |
c | Reichel ACM TOMS paper; Algorithm 686, Dec. 1990. | |
| 651 |
c | Determine if we need to correct the residual. The goal is | |
| 652 |
c | to enforce ||v(:,1:j)^T * r_{j}|| .le. eps * || r_{j} || | |
| 653 |
c %-----------------------------------------------------------% |
| 654 |
c |
| 655 |
if (rnorm .gt. 0.717*wnorm) go to 100 |
| 656 |
nrorth = nrorth + 1 |
| 657 |
c |
| 658 |
c %---------------------------------------------------% |
| 659 |
c | Enter the Iterative refinement phase. If further | |
| 660 |
c | refinement is necessary, loop back here. The loop | |
| 661 |
c | variable is ITER. Perform a step of Classical | |
| 662 |
c | Gram-Schmidt using all the Arnoldi vectors V_{j} | |
| 663 |
c %---------------------------------------------------% |
| 664 |
c |
| 665 |
80 continue |
| 666 |
c |
| 667 |
if (msglvl .gt. 2) then |
| 668 |
xtemp(1) = wnorm |
| 669 |
xtemp(2) = rnorm |
| 670 |
c call dvout (logfil, 2, xtemp, ndigit, |
| 671 |
c & '_saitr: re-orthonalization ; wnorm and rnorm are') |
| 672 |
end if |
| 673 |
c |
| 674 |
c %----------------------------------------------------% |
| 675 |
c | Compute V_{j}^T * B * r_{j}. | |
| 676 |
c | WORKD(IRJ:IRJ+J-1) = v(:,1:J)'*WORKD(IPJ:IPJ+N-1). | |
| 677 |
c %----------------------------------------------------% |
| 678 |
c |
| 679 |
call dgemv ('T', n, j, one, v, ldv, workd(ipj), 1, |
| 680 |
& zero, workd(irj), 1) |
| 681 |
c |
| 682 |
c %----------------------------------------------% |
| 683 |
c | Compute the correction to the residual: | |
| 684 |
c | r_{j} = r_{j} - V_{j} * WORKD(IRJ:IRJ+J-1). | |
| 685 |
c | The correction to H is v(:,1:J)*H(1:J,1:J) + | |
| 686 |
c | v(:,1:J)*WORKD(IRJ:IRJ+J-1)*e'_j, but only | |
| 687 |
c | H(j,j) is updated. | |
| 688 |
c %----------------------------------------------% |
| 689 |
c |
| 690 |
call dgemv ('N', n, j, -one, v, ldv, workd(irj), 1, |
| 691 |
& one, resid, 1) |
| 692 |
c |
| 693 |
if (j .eq. 1 .or. rstart) h(j,1) = zero |
| 694 |
h(j,2) = h(j,2) + workd(irj + j - 1) |
| 695 |
c |
| 696 |
orth2 = .true. |
| 697 |
call second (t2) |
| 698 |
if (bmat .eq. 'G') then |
| 699 |
nbx = nbx + 1 |
| 700 |
call dcopy (n, resid, 1, workd(irj), 1) |
| 701 |
ipntr(1) = irj |
| 702 |
ipntr(2) = ipj |
| 703 |
ido = 2 |
| 704 |
c |
| 705 |
c %-----------------------------------% |
| 706 |
c | Exit in order to compute B*r_{j}. | |
| 707 |
c | r_{j} is the corrected residual. | |
| 708 |
c %-----------------------------------% |
| 709 |
c |
| 710 |
go to 9000 |
| 711 |
else if (bmat .eq. 'I') then |
| 712 |
call dcopy (n, resid, 1, workd(ipj), 1) |
| 713 |
end if |
| 714 |
90 continue |
| 715 |
c |
| 716 |
c %---------------------------------------------------% |
| 717 |
c | Back from reverse communication if ORTH2 = .true. | |
| 718 |
c %---------------------------------------------------% |
| 719 |
c |
| 720 |
if (bmat .eq. 'G') then |
| 721 |
call second (t3) |
| 722 |
tmvbx = tmvbx + (t3 - t2) |
| 723 |
end if |
| 724 |
c |
| 725 |
c %-----------------------------------------------------% |
| 726 |
c | Compute the B-norm of the corrected residual r_{j}. | |
| 727 |
c %-----------------------------------------------------% |
| 728 |
c |
| 729 |
if (bmat .eq. 'G') then |
| 730 |
rnorm1 = ddot (n, resid, 1, workd(ipj), 1) |
| 731 |
rnorm1 = sqrt(abs(rnorm1)) |
| 732 |
else if (bmat .eq. 'I') then |
| 733 |
rnorm1 = dnrm2(n, resid, 1) |
| 734 |
end if |
| 735 |
c |
| 736 |
c if (msglvl .gt. 0 .and. iter .gt. 0) then |
| 737 |
c call ivout (logfil, 1, j, ndigit, |
| 738 |
c & '_saitr: Iterative refinement for Arnoldi residual') |
| 739 |
c if (msglvl .gt. 2) then |
| 740 |
c xtemp(1) = rnorm |
| 741 |
c xtemp(2) = rnorm1 |
| 742 |
c call dvout (logfil, 2, xtemp, ndigit, |
| 743 |
c & '_saitr: iterative refinement ; rnorm and rnorm1 are') |
| 744 |
c end if |
| 745 |
c end if |
| 746 |
c |
| 747 |
c %-----------------------------------------% |
| 748 |
c | Determine if we need to perform another | |
| 749 |
c | step of re-orthogonalization. | |
| 750 |
c %-----------------------------------------% |
| 751 |
c |
| 752 |
if (rnorm1 .gt. 0.717*rnorm) then |
| 753 |
c |
| 754 |
c %--------------------------------% |
| 755 |
c | No need for further refinement | |
| 756 |
c %--------------------------------% |
| 757 |
c |
| 758 |
rnorm = rnorm1 |
| 759 |
c |
| 760 |
else |
| 761 |
c |
| 762 |
c %-------------------------------------------% |
| 763 |
c | Another step of iterative refinement step | |
| 764 |
c | is required. NITREF is used by stat.h | |
| 765 |
c %-------------------------------------------% |
| 766 |
c |
| 767 |
nitref = nitref + 1 |
| 768 |
rnorm = rnorm1 |
| 769 |
iter = iter + 1 |
| 770 |
if (iter .le. 1) go to 80 |
| 771 |
c |
| 772 |
c %-------------------------------------------------% |
| 773 |
c | Otherwise RESID is numerically in the span of V | |
| 774 |
c %-------------------------------------------------% |
| 775 |
c |
| 776 |
do 95 jj = 1, n |
| 777 |
resid(jj) = zero |
| 778 |
95 continue |
| 779 |
rnorm = zero |
| 780 |
end if |
| 781 |
c |
| 782 |
c %----------------------------------------------% |
| 783 |
c | Branch here directly if iterative refinement | |
| 784 |
c | wasn't necessary or after at most NITER_REF | |
| 785 |
c | steps of iterative refinement. | |
| 786 |
c %----------------------------------------------% |
| 787 |
c |
| 788 |
100 continue |
| 789 |
c |
| 790 |
rstart = .false. |
| 791 |
orth2 = .false. |
| 792 |
c |
| 793 |
call second (t5) |
| 794 |
titref = titref + (t5 - t4) |
| 795 |
c |
| 796 |
c %----------------------------------------------------------% |
| 797 |
c | Make sure the last off-diagonal element is non negative | |
| 798 |
c | If not perform a similarity transformation on H(1:j,1:j) | |
| 799 |
c | and scale v(:,j) by -1. | |
| 800 |
c %----------------------------------------------------------% |
| 801 |
c |
| 802 |
if (h(j,1) .lt. zero) then |
| 803 |
h(j,1) = -h(j,1) |
| 804 |
if ( j .lt. k+np) then |
| 805 |
call dscal(n, -one, v(1,j+1), 1) |
| 806 |
else |
| 807 |
call dscal(n, -one, resid, 1) |
| 808 |
end if |
| 809 |
end if |
| 810 |
c |
| 811 |
c %------------------------------------% |
| 812 |
c | STEP 6: Update j = j+1; Continue | |
| 813 |
c %------------------------------------% |
| 814 |
c |
| 815 |
j = j + 1 |
| 816 |
if (j .gt. k+np) then |
| 817 |
call second (t1) |
| 818 |
tsaitr = tsaitr + (t1 - t0) |
| 819 |
ido = 99 |
| 820 |
c |
| 821 |
c if (msglvl .gt. 1) then |
| 822 |
c call dvout (logfil, k+np, h(1,2), ndigit, |
| 823 |
c & '_saitr: main diagonal of matrix H of step K+NP.') |
| 824 |
c if (k+np .gt. 1) then |
| 825 |
c call dvout (logfil, k+np-1, h(2,1), ndigit, |
| 826 |
c & '_saitr: sub diagonal of matrix H of step K+NP.') |
| 827 |
c end if |
| 828 |
c end if |
| 829 |
c |
| 830 |
go to 9000 |
| 831 |
end if |
| 832 |
c |
| 833 |
c %--------------------------------------------------------% |
| 834 |
c | Loop back to extend the factorization by another step. | |
| 835 |
c %--------------------------------------------------------% |
| 836 |
c |
| 837 |
go to 1000 |
| 838 |
c |
| 839 |
c %---------------------------------------------------------------% |
| 840 |
c | | |
| 841 |
c | E N D O F M A I N I T E R A T I O N L O O P | |
| 842 |
c | | |
| 843 |
c %---------------------------------------------------------------% |
| 844 |
c |
| 845 |
9000 continue |
| 846 |
return |
| 847 |
c |
| 848 |
c %---------------% |
| 849 |
c | End of dsaitr | |
| 850 |
c %---------------% |
| 851 |
c |
| 852 |
end |