[1878] | 1 | MODULE solpcg |
---|
| 2 | !!====================================================================== |
---|
| 3 | !! *** MODULE solfet |
---|
| 4 | !! Ocean solver : preconditionned conjugate gradient solver |
---|
| 5 | !!===================================================================== |
---|
| 6 | |
---|
| 7 | !!---------------------------------------------------------------------- |
---|
| 8 | !! sol_pcg : preconditionned conjugate gradient solver |
---|
| 9 | !!---------------------------------------------------------------------- |
---|
| 10 | USE oce ! ocean dynamics and tracers variables |
---|
| 11 | USE dom_oce ! ocean space and time domain variables |
---|
| 12 | USE sol_oce ! ocean solver variables |
---|
| 13 | USE lib_mpp ! distributed memory computing |
---|
| 14 | USE lbclnk ! ocean lateral boundary conditions (or mpp link) |
---|
| 15 | USE in_out_manager ! I/O manager |
---|
| 16 | |
---|
| 17 | IMPLICIT NONE |
---|
| 18 | PRIVATE |
---|
| 19 | |
---|
| 20 | PUBLIC sol_pcg ! |
---|
| 21 | |
---|
| 22 | !! * Substitutions |
---|
| 23 | # include "vectopt_loop_substitute.h90" |
---|
| 24 | !!---------------------------------------------------------------------- |
---|
| 25 | !! NEMO/OPA 3.2 , LOCEAN-IPSL (2009) |
---|
| 26 | !! $Id: solpcg.F90 1601 2009-08-11 10:09:19Z ctlod $ |
---|
| 27 | !! Software governed by the CeCILL licence (modipsl/doc/NEMO_CeCILL.txt) |
---|
| 28 | !!---------------------------------------------------------------------- |
---|
| 29 | CONTAINS |
---|
| 30 | |
---|
| 31 | SUBROUTINE sol_pcg( kindic ) |
---|
| 32 | !!---------------------------------------------------------------------- |
---|
| 33 | !! *** ROUTINE sol_pcg *** |
---|
| 34 | !! |
---|
| 35 | !! ** Purpose : Solve the ellipic equation for the transport |
---|
| 36 | !! divergence system using a diagonal preconditionned |
---|
| 37 | !! conjugate gradient method. |
---|
| 38 | !! |
---|
| 39 | !! ** Method : Diagonal preconditionned conjugate gradient method. |
---|
| 40 | !! the algorithm is multitasked. (case of 5 points matrix) |
---|
| 41 | !! define pa = q^-1 * a |
---|
| 42 | !! pgcb = q^-1 * gcb |
---|
| 43 | !! < . ; . >_q = ( . )^t q ( . ) |
---|
| 44 | !! where q is the preconditioning matrix = diagonal matrix of the |
---|
| 45 | !! diagonal elements of a |
---|
| 46 | !! Initialization : |
---|
| 47 | !! x(o) = gcx |
---|
| 48 | !! r(o) = d(o) = pgcb - pa.x(o) |
---|
| 49 | !! rr(o)= < r(o) , r(o) >_q |
---|
| 50 | !! Iteration 1 : |
---|
| 51 | !! standard PCG algorithm |
---|
| 52 | !! Iteration n > 1 : |
---|
| 53 | !! s(n) = pa.r(n) |
---|
| 54 | !! gam(n) = < r(n) , r(n) >_q |
---|
| 55 | !! del(n) = < r(n) , s(n) >_q |
---|
| 56 | !! bet(n) = gam(n) / gam(n-1) |
---|
| 57 | !! d(n) = r(n) + bet(n) d(n-1) |
---|
| 58 | !! z(n) = s(n) + bet(n) z(n-1) |
---|
| 59 | !! sig(n) = del(n) - bet(n)*bet(n)*sig(n-1) |
---|
| 60 | !! alp(n) = gam(n) / sig(n) |
---|
| 61 | !! x(n+1) = x(n) + alp(n) d(n) |
---|
| 62 | !! r(n+1) = r(n) - alp(n) z(n) |
---|
| 63 | !! Convergence test : |
---|
| 64 | !! rr(n+1) / < gcb , gcb >_q =< epsr |
---|
| 65 | !! |
---|
| 66 | !! ** Action : - niter : solver number of iteration done |
---|
| 67 | !! - res : solver residu reached |
---|
| 68 | !! - gcx() : solution of the elliptic system |
---|
| 69 | !! |
---|
| 70 | !! References : |
---|
| 71 | !! Madec et al. 1988, Ocean Modelling, issue 78, 1-6. |
---|
| 72 | !! D Azevedo et al. 1993, Computer Science Technical Report, Tennessee U. |
---|
| 73 | !! |
---|
| 74 | !! History : |
---|
| 75 | !! ! 90-10 (G. Madec) Original code |
---|
| 76 | !! ! 91-11 (G. Madec) |
---|
| 77 | !! ! 93-04 (M. Guyon) loops and suppress pointers |
---|
| 78 | !! ! 95-09 (M. Imbard, J. Escobar) mpp exchange |
---|
| 79 | !! ! 96-05 (G. Madec) merge sor and pcg formulations |
---|
| 80 | !! ! 96-11 (A. Weaver) correction to preconditioning |
---|
| 81 | !! 8.5 ! 02-08 (G. Madec) F90: Free form |
---|
| 82 | !! ! 08-01 (R. Benshila) mpp optimization |
---|
| 83 | !!---------------------------------------------------------------------- |
---|
| 84 | INTEGER, INTENT( inout ) :: kindic ! solver indicator, < 0 if the conver- |
---|
| 85 | ! ! gence is not reached: the model is |
---|
| 86 | ! ! stopped in step |
---|
| 87 | ! ! set to zero before the call of solpcg |
---|
| 88 | !! |
---|
| 89 | INTEGER :: ji, jj, jn ! dummy loop indices |
---|
| 90 | REAL(wp) :: zgcad ! temporary scalars |
---|
| 91 | REAL(wp), DIMENSION(2) :: zsum |
---|
| 92 | REAL(wp), DIMENSION(jpi,jpj) :: zgcr |
---|
| 93 | !!---------------------------------------------------------------------- |
---|
| 94 | |
---|
| 95 | ! Initialization of the algorithm with standard PCG |
---|
| 96 | ! ------------------------------------------------- |
---|
| 97 | |
---|
| 98 | CALL lbc_lnk( gcx, c_solver_pt, 1. ) ! lateral boundary condition |
---|
| 99 | |
---|
| 100 | ! gcr = gcb-a.gcx |
---|
| 101 | ! gcdes = gcr |
---|
| 102 | |
---|
| 103 | DO jj = 2, jpjm1 |
---|
| 104 | DO ji = fs_2, fs_jpim1 ! vector opt. |
---|
| 105 | zgcad = bmask(ji,jj) * ( gcb(ji,jj ) - gcx(ji ,jj ) & |
---|
| 106 | & - gcp(ji,jj,1) * gcx(ji ,jj-1) & |
---|
| 107 | & - gcp(ji,jj,2) * gcx(ji-1,jj ) & |
---|
| 108 | & - gcp(ji,jj,3) * gcx(ji+1,jj ) & |
---|
| 109 | & - gcp(ji,jj,4) * gcx(ji ,jj+1) ) |
---|
| 110 | gcr (ji,jj) = zgcad |
---|
| 111 | gcdes(ji,jj) = zgcad |
---|
| 112 | END DO |
---|
| 113 | END DO |
---|
| 114 | |
---|
| 115 | ! rnorme = (gcr,gcr) |
---|
| 116 | rnorme = SUM( gcr(:,:) * gcdmat(:,:) * gcr(:,:) ) |
---|
| 117 | IF( lk_mpp ) CALL mpp_sum( rnorme ) ! sum over the global domain |
---|
| 118 | |
---|
| 119 | CALL lbc_lnk( gcdes, c_solver_pt, 1. ) ! lateral boundary condition |
---|
| 120 | |
---|
| 121 | ! gccd = matrix . gcdes |
---|
| 122 | DO jj = 2, jpjm1 |
---|
| 123 | DO ji = fs_2, fs_jpim1 ! vector opt. |
---|
| 124 | gccd(ji,jj) = bmask(ji,jj)*( gcdes(ji,jj) & |
---|
| 125 | & +gcp(ji,jj,1)*gcdes(ji,jj-1)+gcp(ji,jj,2)*gcdes(ji-1,jj) & |
---|
| 126 | & +gcp(ji,jj,4)*gcdes(ji,jj+1)+gcp(ji,jj,3)*gcdes(ji+1,jj) ) |
---|
| 127 | END DO |
---|
| 128 | END DO |
---|
| 129 | |
---|
| 130 | ! alph = (gcr,gcr)/(gcdes,gccd) |
---|
| 131 | radd = SUM( gcdes(:,:) * gcdmat(:,:) * gccd(:,:) ) |
---|
| 132 | IF( lk_mpp ) CALL mpp_sum( radd ) ! sum over the global domain |
---|
| 133 | alph = rnorme /radd |
---|
| 134 | |
---|
| 135 | ! gcx = gcx + alph * gcdes |
---|
| 136 | ! gcr = gcr - alph * gccd |
---|
| 137 | DO jj = 2, jpjm1 |
---|
| 138 | DO ji = fs_2, fs_jpim1 ! vector opt. |
---|
| 139 | gcx(ji,jj) = bmask(ji,jj) * ( gcx(ji,jj) + alph * gcdes(ji,jj) ) |
---|
| 140 | gcr(ji,jj) = bmask(ji,jj) * ( gcr(ji,jj) - alph * gccd (ji,jj) ) |
---|
| 141 | END DO |
---|
| 142 | END DO |
---|
| 143 | |
---|
| 144 | ! Algorithm wtih Eijkhout rearrangement |
---|
| 145 | ! ------------------------------------- |
---|
| 146 | |
---|
| 147 | ! !================ |
---|
| 148 | DO jn = 1, nn_nmax ! Iterative loop |
---|
| 149 | ! !================ |
---|
| 150 | |
---|
| 151 | CALL lbc_lnk( gcr, c_solver_pt, 1. ) ! lateral boundary condition |
---|
| 152 | |
---|
| 153 | ! zgcr = matrix . gcr |
---|
| 154 | DO jj = 2, jpjm1 |
---|
| 155 | DO ji = fs_2, fs_jpim1 ! vector opt. |
---|
| 156 | zgcr(ji,jj) = bmask(ji,jj)*( gcr(ji,jj) & |
---|
| 157 | & +gcp(ji,jj,1)*gcr(ji,jj-1)+gcp(ji,jj,2)*gcr(ji-1,jj) & |
---|
| 158 | & +gcp(ji,jj,4)*gcr(ji,jj+1)+gcp(ji,jj,3)*gcr(ji+1,jj) ) |
---|
| 159 | END DO |
---|
| 160 | END DO |
---|
| 161 | |
---|
| 162 | ! rnorme = (gcr,gcr) |
---|
| 163 | rr = rnorme |
---|
| 164 | zsum(1) = SUM( gcr(:,:) * gcdmat(:,:) * gcr(:,:) ) |
---|
| 165 | |
---|
| 166 | ! zgcad = (zgcr,gcr) |
---|
| 167 | zsum(2) = SUM( gcr(2:jpim1,2:jpjm1) * gcdmat(2:jpim1,2:jpjm1) * zgcr(2:jpim1,2:jpjm1) ) |
---|
| 168 | |
---|
| 169 | IF( lk_mpp ) CALL mpp_sum( zsum, 2 ) ! sum over the global domain |
---|
| 170 | rnorme = zsum(1) |
---|
| 171 | zgcad = zsum(2) |
---|
| 172 | |
---|
| 173 | ! test of convergence |
---|
| 174 | IF( rnorme < epsr .OR. jn == nn_nmax ) THEN |
---|
| 175 | res = SQRT( rnorme ) |
---|
| 176 | niter = jn |
---|
| 177 | ncut = 999 |
---|
| 178 | ENDIF |
---|
| 179 | |
---|
| 180 | ! beta = (rk+1,rk+1)/(rk,rk) |
---|
| 181 | beta = rnorme / rr |
---|
| 182 | radd = zgcad - beta*beta*radd |
---|
| 183 | alph = rnorme / radd |
---|
| 184 | |
---|
| 185 | ! gcx = gcx + alph * gcdes |
---|
| 186 | ! gcr = gcr - alph * gccd |
---|
| 187 | DO jj = 2, jpjm1 |
---|
| 188 | DO ji = fs_2, fs_jpim1 ! vector opt. |
---|
| 189 | gcdes(ji,jj) = gcr (ji,jj) + beta * gcdes(ji,jj) |
---|
| 190 | gccd (ji,jj) = zgcr(ji,jj) + beta * gccd (ji,jj) |
---|
| 191 | gcx (ji,jj) = gcx (ji,jj) + alph * gcdes(ji,jj) |
---|
| 192 | gcr (ji,jj) = gcr (ji,jj) - alph * gccd (ji,jj) |
---|
| 193 | END DO |
---|
| 194 | END DO |
---|
| 195 | |
---|
| 196 | ! indicator of non-convergence or explosion |
---|
| 197 | IF( jn == nn_nmax .OR. SQRT(epsr)/eps > 1.e+20 ) kindic = -2 |
---|
| 198 | IF( ncut == 999 ) GOTO 999 |
---|
| 199 | |
---|
| 200 | ! !================ |
---|
| 201 | END DO ! End Loop |
---|
| 202 | ! !================ |
---|
| 203 | |
---|
| 204 | 999 CONTINUE |
---|
| 205 | |
---|
| 206 | |
---|
| 207 | ! Output in gcx with lateral b.c. applied |
---|
| 208 | ! --------------------------------------- |
---|
| 209 | |
---|
| 210 | CALL lbc_lnk( gcx, c_solver_pt, 1. ) |
---|
| 211 | |
---|
| 212 | END SUBROUTINE sol_pcg |
---|
| 213 | |
---|
| 214 | !!===================================================================== |
---|
| 215 | END MODULE solpcg |
---|