Millepede-II V04-16-01
minresModule.f90
Go to the documentation of this file.
1
3
4!+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
5! File minresModule.f90
6!
38
40
41 use minresdatamodule, only : dp
42
43 implicit none
44 public :: minres
45
46contains
47
48 !+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
49
294 subroutine minres( n, Aprod, Msolve, b, shift, checkA, precon, &
295 x, itnlim, nout, rtol, &
296 istop, itn, Anorm, Acond, rnorm, Arnorm, ynorm )
297
298 integer, intent(in) :: n, itnlim, nout
299 logical, intent(in) :: checka, precon
300 real(dp), intent(in) :: b(n)
301 real(dp), intent(in) :: shift, rtol
302 real(dp), intent(out) :: x(n)
303 integer, intent(out) :: istop, itn
304 real(dp), intent(out) :: anorm, acond, rnorm, arnorm, ynorm
305
306 interface
307 subroutine aprod (n,x,y) ! y := A*x
309 integer, intent(in) :: n
310 real(dp), intent(in) :: x(n)
311 real(dp), intent(out) :: y(n)
312 end subroutine aprod
313
314 subroutine msolve(n,x,y) ! Solve M*y = x
316 integer, intent(in) :: n
317 real(dp), intent(in) :: x(n)
318 real(dp), intent(out) :: y(n)
319 end subroutine msolve
320 end interface
321
322! Local arrays and variables
323 real(dp) :: r1(n), r2(n), v(n), w(n), w1(n), w2(n), y(n)
324 real(dp) :: alfa , beta , beta1 , cs , &
325 dbar , delta , denom , diag , &
326 eps , epsa , epsln , epsr , epsx , &
327 gamma , gbar , gmax , gmin , &
328 oldb , oldeps, qrnorm, phi , phibar, &
329 rhs1 , rhs2 , rnorml, rootl , &
330 arnorml, relarnorml, &
331 s , sn , t , tnorm2, ynorm2, z
332 integer :: i
333 logical :: debug, prnt
334
335 ! Local constants
336 real(dp), parameter :: zero = 0.0, one = 1.0
337 real(dp), parameter :: ten = 10.0
338 character(len=*), parameter :: enter = ' Enter MINRES. '
339 character(len=*), parameter :: exitt = ' Exit MINRES. '
340 character(len=*), parameter :: msg(-1:8) = &
341 (/ 'beta2 = 0. If M = I, b and x are eigenvectors of A', & ! -1
342 'beta1 = 0. The exact solution is x = 0 ', & ! 0
343 'Requested accuracy achieved, as determined by rtol ', & ! 1
344 'Reasonable accuracy achieved, given eps ', & ! 2
345 'x has converged to an eigenvector ', & ! 3
346 'Acond has exceeded 0.1/eps ', & ! 4
347 'The iteration limit was reached ', & ! 5
348 'Aprod does not define a symmetric matrix ', & ! 6
349 'Msolve does not define a symmetric matrix ', & ! 7
350 'Msolve does not define a pos-def preconditioner ' /) ! 8
351 !-------------------------------------------------------------------
352
353 !obsolete intrinsic :: abs, dot_product, epsilon, min, max, sqrt
354
355 ! Print heading and initialize.
356
357 debug = .false.
358 eps = epsilon(eps)
359 if (nout > 0) then
360 write(nout, 1000) enter, n, checka, precon, itnlim, rtol, shift
361 end if
362 istop = 0
363 itn = 0
364 anorm = zero
365 acond = zero
366 rnorm = zero
367 ynorm = zero
368 x(1:n) = zero
369 arnorml = zero
370 gmin = zero
371 gmax = zero
372
373 !-------------------------------------------------------------------
374 ! Set up y and v for the first Lanczos vector v1.
375 ! y = beta1 P' v1, where P = C**(-1).
376 ! v is really P' v1.
377 !-------------------------------------------------------------------
378 y = b
379 r1 = b
380 if ( precon ) call msolve( n, b, y )
381 beta1 = dot_product(b,y)
382
383 if (beta1 < zero) then ! M must be indefinite.
384 istop = 8
385 go to 900
386 end if
387
388 if (beta1 == zero) then ! b = 0 exactly. Stop with x = 0.
389 istop = 0
390 go to 900
391 end if
392
393 beta1 = sqrt( beta1 ) ! Normalize y to get v1 later.
394
395 !-------------------------------------------------------------------
396 ! See if Msolve is symmetric.
397 !-------------------------------------------------------------------
398 if (checka .and. precon) then
399 call msolve( n, y, r2 )
400 s = dot_product(y ,y )
401 t = dot_product(r1,r2)
402 z = abs(s - t)
403 epsa = (s + eps) * eps**0.33333
404 if (z > epsa) then
405 istop = 7
406 go to 900
407 end if
408 end if
409
410 !-------------------------------------------------------------------
411 ! See if Aprod is symmetric. Initialize Arnorm.
412 !-------------------------------------------------------------------
413 if (checka) then
414 call aprod ( n, y, w )
415 call aprod ( n, w, r2 )
416 s = dot_product(w,w )
417 t = dot_product(y,r2)
418 z = abs(s - t)
419 epsa = (s + eps) * eps**0.33333
420 if (z > epsa) then
421 istop = 6
422 go to 900
423 end if
424 arnorml = sqrt(s);
425 else
426 call aprod ( n, y, w )
427 arnorml = sqrt( dot_product(w,w) )
428 end if
429
430 !-------------------------------------------------------------------
431 ! Initialize other quantities.
432 !-------------------------------------------------------------------
433 oldb = zero
434 beta = beta1
435 dbar = zero
436 epsln = zero
437 qrnorm = beta1
438 phibar = beta1
439 rhs1 = beta1
440 rhs2 = zero
441 tnorm2 = zero
442 ynorm2 = zero
443 cs = - one
444 sn = zero
445 w(1:n) = zero
446 w2(1:n)= zero
447 r2(1:n)= r1
448
449 if (debug) then
450 write(*,*) ' '
451 write(*,*) 'b ', b
452 write(*,*) 'beta ', beta
453 write(*,*) ' '
454 end if
455
456 !===================================================================
457 ! Main iteration loop.
458 !===================================================================
459 do
460 itn = itn + 1 ! k = itn = 1 first time through
461
462 !----------------------------------------------------------------
463 ! Obtain quantities for the next Lanczos vector vk+1, k = 1, 2,...
464 ! The general iteration is similar to the case k = 1 with v0 = 0:
465 !
466 ! p1 = Operator * v1 - beta1 * v0,
467 ! alpha1 = v1'p1,
468 ! q2 = p2 - alpha1 * v1,
469 ! beta2^2 = q2'q2,
470 ! v2 = (1/beta2) q2.
471 !
472 ! Again, y = betak P vk, where P = C**(-1).
473 ! .... more description needed.
474 !----------------------------------------------------------------
475 s = one / beta ! Normalize previous vector (in y).
476 v = s*y(1:n) ! v = vk if P = I
477
478 call aprod ( n, v, y )
479 y = y - shift*v ! call daxpy ( n, (- shift), v, 1, y, 1 )
480 if (itn >= 2) then
481 y = y - (beta/oldb)*r1 ! call daxpy ( n, (- beta/oldb), r1, 1, y, 1 )
482 end if
483
484 alfa = dot_product(v,y) ! alphak
485 y = y - (alfa/beta)*r2 ! call daxpy ( n, (- alfa/beta), r2, 1, y, 1 )
486 r1 = r2
487 r2 = y
488 if ( precon ) call msolve( n, r2, y )
489
490 oldb = beta ! oldb = betak
491 beta = dot_product(r2,y) ! beta = betak+1^2
492 if (beta < zero) then
493 istop = 6
494 go to 900
495 end if
496
497 beta = sqrt( beta ) ! beta = betak+1
498 tnorm2 = tnorm2 + alfa**2 + oldb**2 + beta**2
499
500 if (itn == 1) then ! Initialize a few things.
501 if (beta/beta1 <= ten*eps) then ! beta2 = 0 or ~ 0.
502 istop = -1 ! Terminate later.
503 end if
504 !tnorm2 = alfa**2
505 gmax = abs( alfa ) ! alpha1
506 gmin = gmax ! alpha1
507 end if
508
509 ! Apply previous rotation Qk-1 to get
510 ! [deltak epslnk+1] = [cs sn][dbark 0 ]
511 ! [gbar k dbar k+1] [sn -cs][alfak betak+1].
512
513 oldeps = epsln
514 delta = cs * dbar + sn * alfa ! delta1 = 0 deltak
515 gbar = sn * dbar - cs * alfa ! gbar 1 = alfa1 gbar k
516 epsln = sn * beta ! epsln2 = 0 epslnk+1
517 dbar = - cs * beta ! dbar 2 = beta2 dbar k+1
518
519 ! Compute the next plane rotation Qk
520
521 gamma = sqrt( gbar**2 + beta**2 ) ! gammak
522 cs = gbar / gamma ! ck
523 sn = beta / gamma ! sk
524 phi = cs * phibar ! phik
525 phibar = sn * phibar ! phibark+1
526
527 if (debug) then
528 write(*,*) ' '
529 write(*,*) 'v ', v
530 write(*,*) 'alfa ', alfa
531 write(*,*) 'beta ', beta
532 write(*,*) 'gamma', gamma
533 write(*,*) 'delta', delta
534 write(*,*) 'gbar ', gbar
535 write(*,*) 'epsln', epsln
536 write(*,*) 'dbar ', dbar
537 write(*,*) 'phi ', phi
538 write(*,*) 'phiba', phibar
539 write(*,*) ' '
540 end if
541
542 ! Update x.
543
544 denom = one/gamma
545
546 do i = 1, n
547 w1(i) = w2(i)
548 w2(i) = w(i)
549 w(i) = ( v(i) - oldeps*w1(i) - delta*w2(i) ) * denom
550 x(i) = x(i) + phi * w(i)
551 end do
552
553 ! Go round again.
554
555 gmax = max( gmax, gamma )
556 gmin = min( gmin, gamma )
557 z = rhs1 / gamma
558 ynorm2 = z**2 + ynorm2
559 rhs1 = rhs2 - delta * z
560 rhs2 = - epsln * z
561
562 ! Estimate various norms and test for convergence.
563
564 anorm = sqrt( tnorm2 )
565 ynorm = sqrt( ynorm2 )
566 epsa = anorm * eps
567 epsx = anorm * ynorm * eps
568 epsr = anorm * ynorm * rtol
569 diag = gbar
570 if (diag == zero) diag = epsa
571
572 qrnorm = phibar
573 rnorml = rnorm
574 rnorm = qrnorm
575 rootl = sqrt( gbar**2 +dbar**2 ) ! norm([gbar; dbar]);
576 arnorml = rnorml*rootl ! ||A r_{k-1} ||
577 relarnorml = rootl / anorm; ! ||Ar|| / (||A|| ||r||)
578 !relArnorml = Arnorml / Anorm; ! ||Ar|| / ||A||
579
580 ! Estimate cond(A).
581 ! In this version we look at the diagonals of R in the
582 ! factorization of the lower Hessenberg matrix, Q * H = R,
583 ! where H is the tridiagonal matrix from Lanczos with one
584 ! extra row, beta(k+1) e_k^T.
585
586 acond = gmax / gmin
587
588 ! See if any of the stopping criteria are satisfied.
589 ! In rare cases, istop is already -1 from above (Abar = const*I).
590
591 if (istop == 0) then
592 if (itn >= itnlim ) istop = 5
593 if (acond >= 0.1d+0/eps) istop = 4
594 if (epsx >= beta1 ) istop = 3
595 ! original
596 !if (qrnorm <= epsx .or. relArnorml <= epsx) istop = 2
597 !if (qrnorm <= epsr .or. relArnorml <= epsr) istop = 1
598 ! C. Kleinwort, DESY, 131002
599 if (qrnorm <= epsx .or. relarnorml <= eps) istop = 2
600 if (qrnorm <= epsr .or. relarnorml <= rtol) istop = 1
601 end if
602
603
604 ! See if it is time to print something.
605
606 if (nout > 0) then
607 prnt = .false.
608 if (n <= 40 ) prnt = .true.
609 if (itn <= 10 ) prnt = .true.
610 if (itn >= itnlim - 10) prnt = .true.
611 if (mod(itn,10) == 0) prnt = .true.
612 if (qrnorm <= ten * epsx) prnt = .true.
613 if (qrnorm <= ten * epsr) prnt = .true.
614 if (relarnorml<= ten*epsx) prnt = .true.
615 if (relarnorml<= ten*epsr) prnt = .true.
616 if (acond >= 1.0d-2/eps ) prnt = .true.
617 if (istop /= 0 ) prnt = .true.
618
619 if ( prnt ) then
620 if ( itn == 1) write(nout, 1200)
621 write(nout, 1300) itn, x(1), qrnorm, anorm, acond
622 if (mod(itn,10) == 0) write(nout, 1500)
623 end if
624 end if
625 if (istop /= 0) exit
626
627 end do
628 !===================================================================
629 ! End of iteration loop.
630 !===================================================================
631
632 ! Display final status.
633
634900 arnorm = arnorml
635 if (nout > 0) then
636 write(nout, 2000) exitt, istop, itn, &
637 exitt, anorm, acond, &
638 exitt, rnorm, ynorm, arnorm
639 write(nout, 3000) exitt, msg(istop)
640 end if
641
642 return
643
644 1000 format(// 1p, a, 5x, 'Solution of symmetric Ax = b' &
645 / ' n =', i7, 5x, 'checkA =', l4, 12x, &
646 'precon =', l4 &
647 / ' itnlim =', i7, 5x, 'rtol =', e11.2, 5x, &
648 'shift =', e23.14)
649 1200 format(// 5x, 'itn', 8x, 'x(1)', 10x, &
650 'norm(r)', 3x, 'norm(A)', 3x, 'cond(A)')
651 1300 format(1p, i8, e19.10, 3e10.2)
652 1500 format(1x)
653 2000 format(/ 1p, a, 5x, 'istop =', i3, 14x, 'itn =', i8 &
654 / a, 5x, 'Anorm =', e12.4, 5x, 'Acond =', e12.4 &
655 / a, 5x, 'rnorm =', e12.4, 5x, 'ynorm =', e12.4, 5x, 'Arnorml =', e12.4)
656 3000 format( a, 5x, a )
657
658 end subroutine minres
659
660end module minresmodule
subroutine precon(p, n, c, cu, a, s, nrkd)
Constrained preconditioner, decomposition.
Definition: mpnum.f90:2711
Defines real(kind=dp) and a few constants for use in other modules.
integer, parameter, public dp
MINRES solves symmetric systems Ax = b or min ||Ax - b||_2, where the matrix A may be indefinite and/...
subroutine, public minres(n, Aprod, Msolve, b, shift, checkA, precon, x, itnlim, nout, rtol, istop, itn, Anorm, Acond, rnorm, Arnorm, ynorm)
Solution of linear equation system.