!> @file poismg_fast_mod.f90 !--------------------------------------------------------------------------------! ! This file is part of PALM. ! ! PALM is free software: you can redistribute it and/or modify it under the terms ! of the GNU General Public License as published by the Free Software Foundation, ! either version 3 of the License, or (at your option) any later version. ! ! PALM is distributed in the hope that it will be useful, but WITHOUT ANY ! WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR ! A PARTICULAR PURPOSE. See the GNU General Public License for more details. ! ! You should have received a copy of the GNU General Public License along with ! PALM. If not, see . ! ! Copyright 1997-2016 Leibniz Universitaet Hannover !--------------------------------------------------------------------------------! ! ! Current revisions: ! ----------------- ! ! ! Former revisions: ! ----------------- ! $Id: poismg_fast_mod.f90 1905 2016-05-11 13:07:37Z hellstea $ ! ! 1904 2016-05-11 13:06:12Z suehring ! Bugfix: enable special_exchange_horiz only for finer grid levels. ! Some formatting adjustments and variable descriptions. ! ! 1898 2016-05-03 11:27:17Z suehring ! Bugfix: bottom and top boundary condition in resid_fast ! Bugfix: restriction at nzb+1 ! formatting adjustments, variable descriptions added in some declaration blocks ! ! 1850 2016-04-08 13:29:27Z maronga ! Module renamed ! ! ! 1762 2016-02-25 12:31:13Z hellstea ! Introduction of nested domain feature ! ! 1682 2015-10-07 23:56:08Z knoop ! Code annotations made doxygen readable ! ! 1609 2015-07-03 15:37:58Z maronga ! Bugfix: allow compilation without __parallel. ! ! 1575 2015-03-27 09:56:27Z raasch ! Initial revision. ! Routine re-written and optimised based on poismg. ! ! Following optimisations have been made: ! - vectorisation (for Intel-CPUs) of the red-black algorithm by resorting ! array elements with even and odd indices ! - explicit boundary conditions for building walls removed (solver is ! running through the buildings ! - reduced data transfer in case of ghost point exchange, because only ! "red" or "black" data points need to be exchanged. This is not applied ! for coarser grid levels, since for then the transfer time is latency bound ! ! ! Description: ! ------------ !> Solves the Poisson equation for the perturbation pressure with a multigrid !> V- or W-Cycle scheme. !> !> This multigrid method was originally developed for PALM by Joerg Uhlenbrock, !> September 2000 - July 2001. It has been optimised for speed by Klaus !> Ketelsen in November 2014. !> !> @attention Loop unrolling and cache optimization in SOR-Red/Black method !> still does not give the expected speedup! !> !> @todo Further work required. !------------------------------------------------------------------------------! MODULE poismg_mod USE cpulog, & ONLY: cpu_log, log_point_s USE kinds USE pegrid PRIVATE INTEGER, SAVE :: ind_even_odd !< border index between even and odd k index INTEGER, DIMENSION(:), SAVE, ALLOCATABLE :: even_odd_level !< stores ind_even_odd for all MG levels REAL(wp), DIMENSION(:,:), SAVE, ALLOCATABLE :: f1_mg_b, f2_mg_b, f3_mg_b !< blocked version of f1_mg ... INTERFACE poismg_fast MODULE PROCEDURE poismg_fast END INTERFACE poismg_fast INTERFACE sort_k_to_even_odd_blocks MODULE PROCEDURE sort_k_to_even_odd_blocks MODULE PROCEDURE sort_k_to_even_odd_blocks_int MODULE PROCEDURE sort_k_to_even_odd_blocks_1d END INTERFACE sort_k_to_even_odd_blocks PUBLIC poismg_fast CONTAINS !------------------------------------------------------------------------------! ! Description: ! ------------ !> Solves the Poisson equation for the perturbation pressure with a multigrid !> V- or W-Cycle scheme. !------------------------------------------------------------------------------! SUBROUTINE poismg_fast( r ) USE arrays_3d, & ONLY: d, p_loc USE control_parameters, & ONLY: gathered_size, grid_level, grid_level_count, & maximum_grid_level, message_string, mgcycles, mg_cycles, & mg_switch_to_pe0_level, residual_limit, subdomain_size USE cpulog, & ONLY: cpu_log, log_point_s USE indices, & ONLY: nxl, nxlg, nxl_mg, nxr, nxrg, nxr_mg, nys, nysg, nys_mg, nyn,& nyng, nyn_mg, nzb, nzt, nzt_mg IMPLICIT NONE REAL(wp) :: maxerror !< REAL(wp) :: maximum_mgcycles !< REAL(wp) :: residual_norm !< REAL(wp), DIMENSION(nzb:nzt+1,nys-1:nyn+1,nxl-1:nxr+1) :: r !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: p3 !< CALL cpu_log( log_point_s(29), 'poismg_fast', 'start' ) ! !-- Initialize arrays and variables used in this subroutine !-- If the number of grid points of the gathered grid, which is collected !-- on PE0, is larger than the number of grid points of an PE, than array !-- p3 will be enlarged. IF ( gathered_size > subdomain_size ) THEN ALLOCATE( p3(nzb:nzt_mg(mg_switch_to_pe0_level)+1,nys_mg( & mg_switch_to_pe0_level)-1:nyn_mg(mg_switch_to_pe0_level)+1,& nxl_mg(mg_switch_to_pe0_level)-1:nxr_mg( & mg_switch_to_pe0_level)+1) ) ELSE ALLOCATE ( p3(nzb:nzt+1,nysg:nyng,nxlg:nxrg) ) ENDIF p3 = 0.0_wp ! !-- Ghost boundaries have to be added to divergence array. !-- Exchange routine needs to know the grid level! grid_level = maximum_grid_level CALL exchange_horiz( d, 1) d(nzb,:,:) = d(nzb+1,:,:) ! !-- Initiation of the multigrid scheme. Does n cycles until the !-- residual is smaller than the given limit. The accuracy of the solution !-- of the poisson equation will increase with the number of cycles. !-- If the number of cycles is preset by the user, this number will be !-- carried out regardless of the accuracy. grid_level_count = 0 mgcycles = 0 IF ( mg_cycles == -1 ) THEN maximum_mgcycles = 0 residual_norm = 1.0_wp ELSE maximum_mgcycles = mg_cycles residual_norm = 0.0_wp ENDIF ! !-- Initial settings for sorting k-dimension from sequential order (alternate !-- even/odd) into blocks of even and odd or vice versa CALL init_even_odd_blocks ! !-- Sort input arrays in even/odd blocks along k-dimension CALL sort_k_to_even_odd_blocks( d, grid_level ) CALL sort_k_to_even_odd_blocks( p_loc, grid_level ) ! !-- The complete multigrid cycles are running in block mode, i.e. over !-- seperate data blocks of even and odd indices DO WHILE ( residual_norm > residual_limit .OR. & mgcycles < maximum_mgcycles ) CALL next_mg_level_fast( d, p_loc, p3, r) ! !-- Calculate the residual if the user has not preset the number of !-- cycles to be performed IF ( maximum_mgcycles == 0 ) THEN CALL resid_fast( d, p_loc, r ) maxerror = SUM( r(nzb+1:nzt,nys:nyn,nxl:nxr)**2 ) #if defined( __parallel ) IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLREDUCE( maxerror, residual_norm, 1, MPI_REAL, & MPI_SUM, comm2d, ierr) #else residual_norm = maxerror #endif residual_norm = SQRT( residual_norm ) ENDIF mgcycles = mgcycles + 1 ! !-- If the user has not limited the number of cycles, stop the run in case !-- of insufficient convergence IF ( mgcycles > 1000 .AND. mg_cycles == -1 ) THEN message_string = 'no sufficient convergence within 1000 cycles' CALL message( 'poismg_fast', 'PA0283', 1, 2, 0, 6, 0 ) ENDIF ENDDO DEALLOCATE( p3 ) ! !-- Result has to be sorted back from even/odd blocks to sequential order CALL sort_k_to_sequential( p_loc ) ! !-- Unset the grid level. Variable is used to determine the MPI datatypes for !-- ghost point exchange grid_level = 0 CALL cpu_log( log_point_s(29), 'poismg_fast', 'stop' ) END SUBROUTINE poismg_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Computes the residual of the perturbation pressure. !------------------------------------------------------------------------------! SUBROUTINE resid_fast( f_mg, p_mg, r ) USE arrays_3d, & ONLY: f1_mg, f2_mg, f3_mg USE control_parameters, & ONLY: bc_lr_cyc, bc_ns_cyc, grid_level, ibc_p_b, ibc_p_t, inflow_l,& inflow_n, inflow_r, inflow_s, nest_bound_l, nest_bound_n, & nest_bound_r, nest_bound_s, outflow_l, outflow_n, outflow_r, & outflow_s USE grid_variables, & ONLY: ddx2_mg, ddy2_mg USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< index indicating grid level INTEGER(iwp) :: km1 !< index variable along z dimension (k-1) INTEGER(iwp) :: kp1 !< index variable along z dimension (k+1) REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: f_mg !< velocity divergence REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: p_mg !< perturbation pressure REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: r !< residuum of perturbation pressure ! !-- Calculate the residual l = grid_level CALL cpu_log( log_point_s(53), 'resid', 'start' ) !$OMP PARALLEL PRIVATE (i,j,k,km1,kp1) !$OMP DO DO i = nxl_mg(l), nxr_mg(l) DO j = nys_mg(l), nyn_mg(l) !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd r(k,j,i) = f_mg(k,j,i) & - ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & - ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & - f2_mg_b(k,l) * p_mg(kp1,j,i) & - f3_mg_b(k,l) * p_mg(km1,j,i) & + f1_mg_b(k,l) * p_mg(k,j,i) ENDDO !DIR$ IVDEP DO k = nzb+1, ind_even_odd km1 = k+ind_even_odd kp1 = k+ind_even_odd+1 r(k,j,i) = f_mg(k,j,i) & - ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & - ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & - f2_mg_b(k,l) * p_mg(kp1,j,i) & - f3_mg_b(k,l) * p_mg(km1,j,i) & + f1_mg_b(k,l) * p_mg(k,j,i) ENDDO ENDDO ENDDO !$OMP END PARALLEL ! !-- Horizontal boundary conditions CALL exchange_horiz( r, 1) IF ( .NOT. bc_lr_cyc ) THEN IF ( inflow_l .OR. outflow_l .OR. nest_bound_l ) THEN r(:,:,nxl_mg(l)-1) = r(:,:,nxl_mg(l)) ENDIF IF ( inflow_r .OR. outflow_r .OR. nest_bound_r ) THEN r(:,:,nxr_mg(l)+1) = r(:,:,nxr_mg(l)) ENDIF ENDIF IF ( .NOT. bc_ns_cyc ) THEN IF ( inflow_n .OR. outflow_n .OR. nest_bound_n ) THEN r(:,nyn_mg(l)+1,:) = r(:,nyn_mg(l),:) ENDIF IF ( inflow_s .OR. outflow_s .OR. nest_bound_s ) THEN r(:,nys_mg(l)-1,:) = r(:,nys_mg(l),:) ENDIF ENDIF ! !-- Boundary conditions at bottom and top of the domain. Points may be within !-- buildings, but that doesn't matter. IF ( ibc_p_b == 1 ) THEN ! !-- equivalent to r(nzb,:,: ) = r(nzb+1,:,:) r(nzb,:,: ) = r(ind_even_odd+1,:,:) ELSE r(nzb,:,: ) = 0.0_wp ENDIF IF ( ibc_p_t == 1 ) THEN ! !-- equivalent to r(nzt_mg(l)+1,:,: ) = r(nzt_mg(l),:,:) r(nzt_mg(l)+1,:,: ) = r(ind_even_odd,:,:) ELSE r(nzt_mg(l)+1,:,: ) = 0.0_wp ENDIF CALL cpu_log( log_point_s(53), 'resid', 'stop' ) END SUBROUTINE resid_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Interpolates the residual on the next coarser grid with "full weighting" !> scheme !------------------------------------------------------------------------------! SUBROUTINE restrict_fast( f_mg, r ) USE control_parameters, & ONLY: bc_lr_cyc, bc_ns_cyc, grid_level, ibc_p_b, ibc_p_t, inflow_l,& inflow_n, inflow_r, inflow_s, nest_bound_l, nest_bound_n, & nest_bound_r, nest_bound_s, outflow_l, outflow_n, outflow_r, & outflow_s USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: i !< index variable along x on finer grid INTEGER(iwp) :: ic !< index variable along x on coarser grid INTEGER(iwp) :: j !< index variable along y on finer grid INTEGER(iwp) :: jc !< index variable along y on coarser grid INTEGER(iwp) :: k !< index variable along z on finer grid INTEGER(iwp) :: kc !< index variable along z on coarser grid INTEGER(iwp) :: l !< index indicating finer grid level INTEGER(iwp) :: km1 !< index variable along z dimension (k-1 on finer level) INTEGER(iwp) :: kp1 !< index variable along z dimension (k+1 on finer level) REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & f_mg !< Residual on coarser grid level REAL(wp), DIMENSION(nzb:nzt_mg(grid_level+1)+1, & nys_mg(grid_level+1)-1:nyn_mg(grid_level+1)+1, & nxl_mg(grid_level+1)-1:nxr_mg(grid_level+1)+1) :: & r !< Residual on finer grid level ! !-- Interpolate the residual l = grid_level CALL cpu_log( log_point_s(54), 'restrict', 'start' ) ! !-- No wall treatment !$OMP PARALLEL PRIVATE (i,j,k,ic,jc,kc,km1,kp1) DO ic = nxl_mg(l), nxr_mg(l) i = 2*ic !$OMP DO SCHEDULE( STATIC ) DO jc = nys_mg(l), nyn_mg(l) ! !-- Calculation for the first point along k j = 2*jc ! !-- Calculation for the other points along k !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l+1) ! Fine grid at this point km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd kc = k-ind_even_odd ! Coarse grid index f_mg(kc,jc,ic) = 1.0_wp / 64.0_wp * ( & 8.0_wp * r(k,j,i) & + 4.0_wp * ( r(k,j,i-1) + r(k,j,i+1) + & r(k,j+1,i) + r(k,j-1,i) ) & + 2.0_wp * ( r(k,j-1,i-1) + r(k,j+1,i-1) + & r(k,j-1,i+1) + r(k,j+1,i+1) ) & + 4.0_wp * r(km1,j,i) & + 2.0_wp * ( r(km1,j,i-1) + r(km1,j,i+1) + & r(km1,j+1,i) + r(km1,j-1,i) ) & + ( r(km1,j-1,i-1) + r(km1,j+1,i-1) + & r(km1,j-1,i+1) + r(km1,j+1,i+1) ) & + 4.0_wp * r(kp1,j,i) & + 2.0_wp * ( r(kp1,j,i-1) + r(kp1,j,i+1) + & r(kp1,j+1,i) + r(kp1,j-1,i) ) & + ( r(kp1,j-1,i-1) + r(kp1,j+1,i-1) + & r(kp1,j-1,i+1) + r(kp1,j+1,i+1) ) & ) ENDDO ENDDO !$OMP ENDDO nowait ENDDO !$OMP END PARALLEL ! !-- Ghost point exchange CALL exchange_horiz( f_mg, 1) ! !-- Horizontal boundary conditions IF ( .NOT. bc_lr_cyc ) THEN IF ( inflow_l .OR. outflow_l .OR. nest_bound_l ) THEN f_mg(:,:,nxl_mg(l)-1) = f_mg(:,:,nxl_mg(l)) ENDIF IF ( inflow_r .OR. outflow_r .OR. nest_bound_r ) THEN f_mg(:,:,nxr_mg(l)+1) = f_mg(:,:,nxr_mg(l)) ENDIF ENDIF IF ( .NOT. bc_ns_cyc ) THEN IF ( inflow_n .OR. outflow_n .OR. nest_bound_n ) THEN f_mg(:,nyn_mg(l)+1,:) = f_mg(:,nyn_mg(l),:) ENDIF IF ( inflow_s .OR. outflow_s .OR. nest_bound_s ) THEN f_mg(:,nys_mg(l)-1,:) = f_mg(:,nys_mg(l),:) ENDIF ENDIF ! !-- Boundary conditions at bottom and top of the domain. !-- These points are not handled by the above loop. Points may be within !-- buildings, but that doesn't matter. Remark: f_mg is ordered sequentielly !-- after interpolation on coarse grid (is ordered in odd-even blocks further !-- below). IF ( ibc_p_b == 1 ) THEN f_mg(nzb,:,: ) = f_mg(nzb+1,:,:) ELSE f_mg(nzb,:,: ) = 0.0_wp ENDIF IF ( ibc_p_t == 1 ) THEN f_mg(nzt_mg(l)+1,:,: ) = f_mg(nzt_mg(l),:,:) ELSE f_mg(nzt_mg(l)+1,:,: ) = 0.0_wp ENDIF CALL cpu_log( log_point_s(54), 'restrict', 'stop' ) ! !-- Since residual is in sequential order after interpolation, an additional !-- sorting in odd-even blocks along z dimension is required at this point. CALL sort_k_to_even_odd_blocks( f_mg , l) END SUBROUTINE restrict_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Interpolates the correction of the perturbation pressure !> to the next finer grid. !------------------------------------------------------------------------------! SUBROUTINE prolong_fast( p, temp ) USE control_parameters, & ONLY: bc_lr_cyc, bc_ns_cyc, grid_level, ibc_p_b, ibc_p_t, inflow_l,& inflow_n, inflow_r, inflow_s, nest_bound_l, nest_bound_n, & nest_bound_r, nest_bound_s, outflow_l, outflow_n, & outflow_r, outflow_s USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: i !< index variable along x on coarser grid level INTEGER(iwp) :: j !< index variable along y on coarser grid level INTEGER(iwp) :: k !< index variable along z on coarser grid level INTEGER(iwp) :: l !< index indicating finer grid level INTEGER(iwp) :: kp1 !< index variable along z INTEGER(iwp) :: ke !< index for prolog even INTEGER(iwp) :: ko !< index for prolog odd REAL(wp), DIMENSION(nzb:nzt_mg(grid_level-1)+1, & nys_mg(grid_level-1)-1:nyn_mg(grid_level-1)+1, & nxl_mg(grid_level-1)-1:nxr_mg(grid_level-1)+1 ) :: & p !< perturbation pressure on coarser grid level REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & temp !< perturbation pressure on finer grid level CALL cpu_log( log_point_s(55), 'prolong', 'start' ) ! !-- First, store elements of the coarser grid on the next finer grid l = grid_level ind_even_odd = even_odd_level(grid_level-1) !$OMP PARALLEL PRIVATE (i,j,k,kp1,ke,ko) !$OMP DO DO i = nxl_mg(l-1), nxr_mg(l-1) DO j = nys_mg(l-1), nyn_mg(l-1) !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l-1) kp1 = k - ind_even_odd ke = 2 * ( k-ind_even_odd - 1 ) + 1 ko = 2 * k - 1 ! !-- Points of the coarse grid are directly stored on the next finer !-- grid temp(ko,2*j,2*i) = p(k,j,i) ! !-- Points between two coarse-grid points temp(ko,2*j,2*i+1) = 0.5_wp * ( p(k,j,i) + p(k,j,i+1) ) temp(ko,2*j+1,2*i) = 0.5_wp * ( p(k,j,i) + p(k,j+1,i) ) temp(ke,2*j,2*i) = 0.5_wp * ( p(k,j,i) + p(kp1,j,i) ) ! !-- Points in the center of the planes stretched by four points !-- of the coarse grid cube temp(ko,2*j+1,2*i+1) = 0.25_wp * ( p(k,j,i) + p(k,j,i+1) + & p(k,j+1,i) + p(k,j+1,i+1) ) temp(ke,2*j,2*i+1) = 0.25_wp * ( p(k,j,i) + p(k,j,i+1) + & p(kp1,j,i) + p(kp1,j,i+1) ) temp(ke,2*j+1,2*i) = 0.25_wp * ( p(k,j,i) + p(k,j+1,i) + & p(kp1,j,i) + p(kp1,j+1,i) ) ! !-- Points in the middle of coarse grid cube temp(ke,2*j+1,2*i+1) = 0.125_wp * & ( p(k,j,i) + p(k,j,i+1) + & p(k,j+1,i) + p(k,j+1,i+1) + & p(kp1,j,i) + p(kp1,j,i+1) + & p(kp1,j+1,i) + p(kp1,j+1,i+1) ) ENDDO !DIR$ IVDEP DO k = nzb+1, ind_even_odd kp1 = k + ind_even_odd + 1 ke = 2 * k ko = 2 * ( k + ind_even_odd ) ! !-- Points of the coarse grid are directly stored on the next finer !-- grid temp(ko,2*j,2*i) = p(k,j,i) ! !-- Points between two coarse-grid points temp(ko,2*j,2*i+1) = 0.5_wp * ( p(k,j,i) + p(k,j,i+1) ) temp(ko,2*j+1,2*i) = 0.5_wp * ( p(k,j,i) + p(k,j+1,i) ) temp(ke,2*j,2*i) = 0.5_wp * ( p(k,j,i) + p(kp1,j,i) ) ! !-- Points in the center of the planes stretched by four points !-- of the coarse grid cube temp(ko,2*j+1,2*i+1) = 0.25_wp * ( p(k,j,i) + p(k,j,i+1) + & p(k,j+1,i) + p(k,j+1,i+1) ) temp(ke,2*j,2*i+1) = 0.25_wp * ( p(k,j,i) + p(k,j,i+1) + & p(kp1,j,i) + p(kp1,j,i+1) ) temp(ke,2*j+1,2*i) = 0.25_wp * ( p(k,j,i) + p(k,j+1,i) + & p(kp1,j,i) + p(kp1,j+1,i) ) ! !-- Points in the middle of coarse grid cube temp(ke,2*j+1,2*i+1) = 0.125_wp * & ( p(k,j,i) + p(k,j,i+1) + & p(k,j+1,i) + p(k,j+1,i+1) + & p(kp1,j,i) + p(kp1,j,i+1) + & p(kp1,j+1,i) + p(kp1,j+1,i+1) ) ENDDO ENDDO ENDDO !$OMP END PARALLEL ind_even_odd = even_odd_level(grid_level) ! !-- Horizontal boundary conditions CALL exchange_horiz( temp, 1) IF ( .NOT. bc_lr_cyc ) THEN IF ( inflow_l .OR. outflow_l .OR. nest_bound_l ) THEN temp(:,:,nxl_mg(l)-1) = temp(:,:,nxl_mg(l)) ENDIF IF ( inflow_r .OR. outflow_r .OR. nest_bound_r ) THEN temp(:,:,nxr_mg(l)+1) = temp(:,:,nxr_mg(l)) ENDIF ENDIF IF ( .NOT. bc_ns_cyc ) THEN IF ( inflow_n .OR. outflow_n .OR. nest_bound_n ) THEN temp(:,nyn_mg(l)+1,:) = temp(:,nyn_mg(l),:) ENDIF IF ( inflow_s .OR. outflow_s .OR. nest_bound_s ) THEN temp(:,nys_mg(l)-1,:) = temp(:,nys_mg(l),:) ENDIF ENDIF ! !-- Bottom and top boundary conditions IF ( ibc_p_b == 1 ) THEN ! !-- equivalent to temp(nzb,:,: ) = temp(nzb+1,:,:) temp(nzb,:,: ) = temp(ind_even_odd+1,:,:) ELSE temp(nzb,:,: ) = 0.0_wp ENDIF IF ( ibc_p_t == 1 ) THEN ! !-- equivalent to temp(nzt_mg(l)+1,:,: ) = temp(nzt_mg(l),:,:) temp(nzt_mg(l)+1,:,: ) = temp(ind_even_odd,:,:) ELSE temp(nzt_mg(l)+1,:,: ) = 0.0_wp ENDIF CALL cpu_log( log_point_s(55), 'prolong', 'stop' ) END SUBROUTINE prolong_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Relaxation method for the multigrid scheme. A Gauss-Seidel iteration with !> 3D-Red-Black decomposition (GS-RB) is used. !------------------------------------------------------------------------------! SUBROUTINE redblack_fast( f_mg, p_mg ) USE arrays_3d, & ONLY: f1_mg, f2_mg, f3_mg USE control_parameters, & ONLY: bc_lr_cyc, bc_ns_cyc, grid_level, ibc_p_b, ibc_p_t, inflow_l,& inflow_n, inflow_r, inflow_s, nest_bound_l, nest_bound_n, & nest_bound_r, nest_bound_s, ngsrb, outflow_l, outflow_n, & outflow_r, outflow_s USE grid_variables, & ONLY: ddx2_mg, ddy2_mg USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: color !< grid point color, either red or black INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: ic !< index variable along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: jc !< index variable along y INTEGER(iwp) :: jj !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< grid level INTEGER(iwp) :: n !< loop variable Gauß-Seidel iterations INTEGER(iwp) :: km1 !< index variable (k-1) INTEGER(iwp) :: kp1 !< index variable (k+1) LOGICAL :: unroll !< flag indicating whether loop unrolling is possible REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & f_mg !< residual of perturbation pressure REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & p_mg !< perturbation pressure l = grid_level unroll = ( MOD( nyn_mg(l)-nys_mg(l)+1, 4 ) == 0 .AND. & MOD( nxr_mg(l)-nxl_mg(l)+1, 2 ) == 0 ) DO n = 1, ngsrb DO color = 1, 2 IF ( .NOT. unroll ) THEN CALL cpu_log( log_point_s(36), 'redblack_no_unroll_f', 'start' ) ! !-- Without unrolling of loops, no cache optimization !$OMP PARALLEL PRIVATE (i,j,k,km1,kp1) !$OMP DO DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO ENDDO ENDDO !$OMP DO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO ENDDO ENDDO !$OMP DO DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 !DIR$ IVDEP DO k = nzb+1, ind_even_odd km1 = k+ind_even_odd kp1 = k+ind_even_odd+1 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO ENDDO ENDDO !$OMP DO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 !DIR$ IVDEP DO k = nzb+1, ind_even_odd km1 = k+ind_even_odd kp1 = k+ind_even_odd+1 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO ENDDO ENDDO !$OMP END PARALLEL CALL cpu_log( log_point_s(36), 'redblack_no_unroll_f', 'stop' ) ELSE ! !-- Loop unrolling along y, only one i loop for better cache use CALL cpu_log( log_point_s(38), 'redblack_unroll_f', 'start' ) !$OMP PARALLEL PRIVATE (i,j,k,ic,jc,km1,kp1,jj) !$OMP DO DO ic = nxl_mg(l), nxr_mg(l), 2 DO jc = nys_mg(l), nyn_mg(l), 4 i = ic jj = jc+2-color !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd j = jj p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) j = jj+2 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO i = ic+1 jj = jc+color-1 !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) km1 = k-ind_even_odd-1 kp1 = k-ind_even_odd j = jj p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) j = jj+2 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO i = ic jj = jc+color-1 !DIR$ IVDEP DO k = nzb+1, ind_even_odd km1 = k+ind_even_odd kp1 = k+ind_even_odd+1 j = jj p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) j = jj+2 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO i = ic+1 jj = jc+2-color !DIR$ IVDEP DO k = nzb+1, ind_even_odd km1 = k+ind_even_odd kp1 = k+ind_even_odd+1 j = jj p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) j = jj+2 p_mg(k,j,i) = 1.0_wp / f1_mg_b(k,l) * ( & ddx2_mg(l) * & ( p_mg(k,j,i+1) + p_mg(k,j,i-1) ) & + ddy2_mg(l) * & ( p_mg(k,j+1,i) + p_mg(k,j-1,i) ) & + f2_mg_b(k,l) * p_mg(kp1,j,i) & + f3_mg_b(k,l) * p_mg(km1,j,i) & - f_mg(k,j,i) ) ENDDO ENDDO ENDDO !$OMP END PARALLEL CALL cpu_log( log_point_s(38), 'redblack_unroll_f', 'stop' ) ENDIF ! !-- Horizontal boundary conditions CALL special_exchange_horiz( p_mg, color ) IF ( .NOT. bc_lr_cyc ) THEN IF ( inflow_l .OR. outflow_l .OR. nest_bound_l ) THEN p_mg(:,:,nxl_mg(l)-1) = p_mg(:,:,nxl_mg(l)) ENDIF IF ( inflow_r .OR. outflow_r .OR. nest_bound_r ) THEN p_mg(:,:,nxr_mg(l)+1) = p_mg(:,:,nxr_mg(l)) ENDIF ENDIF IF ( .NOT. bc_ns_cyc ) THEN IF ( inflow_n .OR. outflow_n .OR. nest_bound_n ) THEN p_mg(:,nyn_mg(l)+1,:) = p_mg(:,nyn_mg(l),:) ENDIF IF ( inflow_s .OR. outflow_s .OR. nest_bound_s ) THEN p_mg(:,nys_mg(l)-1,:) = p_mg(:,nys_mg(l),:) ENDIF ENDIF ! !-- Bottom and top boundary conditions IF ( ibc_p_b == 1 ) THEN ! !-- equivalent to p_mg(nzb,:,: ) = p_mg(nzb+1,:,:) p_mg(nzb,:,: ) = p_mg(ind_even_odd+1,:,:) ELSE p_mg(nzb,:,: ) = 0.0_wp ENDIF IF ( ibc_p_t == 1 ) THEN ! !-- equivalent to p_mg(nzt_mg(l)+1,:,: ) = p_mg(nzt_mg(l),:,:) p_mg(nzt_mg(l)+1,:,: ) = p_mg(ind_even_odd,:,:) ELSE p_mg(nzt_mg(l)+1,:,: ) = 0.0_wp ENDIF ENDDO ENDDO END SUBROUTINE redblack_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Sort k-Dimension from sequential into blocks of even and odd. !> This is required to vectorize the red-black subroutine. !> Version for 3D-REAL arrays !------------------------------------------------------------------------------! SUBROUTINE sort_k_to_even_odd_blocks( p_mg , glevel ) USE control_parameters, & ONLY: grid_level USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp), INTENT(IN) :: glevel !< grid level REAL(wp), DIMENSION(nzb:nzt_mg(glevel)+1, & nys_mg(glevel)-1:nyn_mg(glevel)+1, & nxl_mg(glevel)-1:nxr_mg(glevel)+1) :: & p_mg !< array to be sorted ! !-- Local variables INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< grid level INTEGER(iwp) :: ind !< index variable along z REAL(wp), DIMENSION(nzb:nzt_mg(glevel)+1) :: tmp !< odd-even sorted temporary array CALL cpu_log( log_point_s(52), 'sort_k_to_even_odd', 'start' ) l = glevel ind_even_odd = even_odd_level(l) !$OMP PARALLEL PRIVATE (i,j,k,ind,tmp) !$OMP DO DO i = nxl_mg(l)-1, nxr_mg(l)+1 DO j = nys_mg(l)-1, nyn_mg(l)+1 ! !-- Sort the data with even k index ind = nzb-1 DO k = nzb, nzt_mg(l), 2 ind = ind + 1 tmp(ind) = p_mg(k,j,i) ENDDO ! !-- Sort the data with odd k index DO k = nzb+1, nzt_mg(l)+1, 2 ind = ind + 1 tmp(ind) = p_mg(k,j,i) ENDDO p_mg(:,j,i) = tmp ENDDO ENDDO !$OMP END PARALLEL CALL cpu_log( log_point_s(52), 'sort_k_to_even_odd', 'stop' ) END SUBROUTINE sort_k_to_even_odd_blocks !------------------------------------------------------------------------------! ! Description: ! ------------ !> Sort k-Dimension from sequential into blocks of even and odd. !> This is required to vectorize the red-black subroutine. !> Version for 1D-REAL arrays !------------------------------------------------------------------------------! SUBROUTINE sort_k_to_even_odd_blocks_1d( f_mg, f_mg_b, glevel ) USE indices, & ONLY: nzb, nzt_mg IMPLICIT NONE INTEGER(iwp), INTENT(IN) :: glevel !< grid level REAL(wp), DIMENSION(nzb+1:nzt_mg(glevel)) :: f_mg !< 1D input array REAL(wp), DIMENSION(nzb:nzt_mg(glevel)+1) :: f_mg_b !< 1D output array ! !-- Local variables INTEGER(iwp) :: ind !< index variable along z INTEGER(iwp) :: k !< index variable along z ind = nzb - 1 ! !-- Sort the data with even k index DO k = nzb, nzt_mg(glevel), 2 ind = ind + 1 IF ( k >= nzb+1 .AND. k <= nzt_mg(glevel) ) THEN f_mg_b(ind) = f_mg(k) ENDIF ENDDO ! !-- Sort the data with odd k index DO k = nzb+1, nzt_mg(glevel)+1, 2 ind = ind + 1 IF( k >= nzb+1 .AND. k <= nzt_mg(glevel) ) THEN f_mg_b(ind) = f_mg(k) ENDIF ENDDO END SUBROUTINE sort_k_to_even_odd_blocks_1d !------------------------------------------------------------------------------! ! Description: ! ------------ !> Sort k-Dimension from sequential into blocks of even and odd. !> This is required to vectorize the red-black subroutine. !> Version for 2D-INTEGER arrays !------------------------------------------------------------------------------! SUBROUTINE sort_k_to_even_odd_blocks_int( i_mg , glevel ) USE control_parameters, & ONLY: grid_level USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp), INTENT(IN) :: glevel !< grid level INTEGER(iwp), DIMENSION(nzb:nzt_mg(glevel)+1, & nys_mg(glevel)-1:nyn_mg(glevel)+1, & nxl_mg(glevel)-1:nxr_mg(glevel)+1) :: & i_mg !< array to be sorted ! !-- Local variables INTEGER(iwp) :: i !< index variabel along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< grid level INTEGER(iwp) :: ind !< index variable along z INTEGER(iwp),DIMENSION(nzb:nzt_mg(glevel)+1) :: tmp !< temporary odd-even sorted array CALL cpu_log( log_point_s(52), 'sort_k_to_even_odd', 'start' ) l = glevel ind_even_odd = even_odd_level(l) DO i = nxl_mg(l)-1, nxr_mg(l)+1 DO j = nys_mg(l)-1, nyn_mg(l)+1 ! !-- Sort the data with even k index ind = nzb-1 DO k = nzb, nzt_mg(l), 2 ind = ind + 1 tmp(ind) = i_mg(k,j,i) ENDDO ! !++ ATTENTION: Check reason for this error. Remove it or replace WRITE !++ by PALM message #if defined ( __parallel ) IF ( ind /= ind_even_odd ) THEN WRITE (0,*) 'ERROR ==> illegal ind_even_odd ',ind,ind_even_odd,l CALL MPI_ABORT(MPI_COMM_WORLD,i,j) ENDIF #endif ! !-- Sort the data with odd k index DO k = nzb+1, nzt_mg(l)+1, 2 ind = ind + 1 tmp(ind) = i_mg(k,j,i) ENDDO i_mg(:,j,i) = tmp ENDDO ENDDO CALL cpu_log( log_point_s(52), 'sort_k_to_even_odd', 'stop' ) END SUBROUTINE sort_k_to_even_odd_blocks_int !------------------------------------------------------------------------------! ! Description: ! ------------ !> Sort k-dimension from blocks of even and odd into sequential !------------------------------------------------------------------------------! SUBROUTINE sort_k_to_sequential( p_mg ) USE control_parameters, & ONLY: grid_level USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & p_mg !< array to be sorted ! !-- Local variables INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< grid level INTEGER(iwp) :: ind !< index variable along z REAL(wp),DIMENSION(nzb:nzt_mg(grid_level)+1) :: tmp l = grid_level !$OMP PARALLEL PRIVATE (i,j,k,ind,tmp) !$OMP DO DO i = nxl_mg(l)-1, nxr_mg(l)+1 DO j = nys_mg(l)-1, nyn_mg(l)+1 ind = nzb - 1 tmp = p_mg(:,j,i) DO k = nzb, nzt_mg(l), 2 ind = ind + 1 p_mg(k,j,i) = tmp(ind) ENDDO DO k = nzb+1, nzt_mg(l)+1, 2 ind = ind + 1 p_mg(k,j,i) = tmp(ind) ENDDO ENDDO ENDDO !$OMP END PARALLEL END SUBROUTINE sort_k_to_sequential !------------------------------------------------------------------------------! ! Description: ! ------------ !> Gather subdomain data from all PEs. !------------------------------------------------------------------------------! SUBROUTINE mg_gather_fast( f2, f2_sub ) USE control_parameters, & ONLY: grid_level USE cpulog, & ONLY: cpu_log, log_point_s USE indices, & ONLY: mg_loc_ind, nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: i !< INTEGER(iwp) :: il !< INTEGER(iwp) :: ir !< INTEGER(iwp) :: j !< INTEGER(iwp) :: jn !< INTEGER(iwp) :: js !< INTEGER(iwp) :: k !< INTEGER(iwp) :: nwords !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: f2 !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: f2_l !< REAL(wp), DIMENSION(nzb:mg_loc_ind(5,myid)+1, & mg_loc_ind(3,myid)-1:mg_loc_ind(4,myid)+1, & mg_loc_ind(1,myid)-1:mg_loc_ind(2,myid)+1) :: f2_sub !< #if defined( __parallel ) CALL cpu_log( log_point_s(34), 'mg_gather', 'start' ) f2_l = 0.0_wp ! !-- Store the local subdomain array on the total array js = mg_loc_ind(3,myid) IF ( south_border_pe ) js = js - 1 jn = mg_loc_ind(4,myid) IF ( north_border_pe ) jn = jn + 1 il = mg_loc_ind(1,myid) IF ( left_border_pe ) il = il - 1 ir = mg_loc_ind(2,myid) IF ( right_border_pe ) ir = ir + 1 DO i = il, ir DO j = js, jn DO k = nzb, nzt_mg(grid_level)+1 f2_l(k,j,i) = f2_sub(k,j,i) ENDDO ENDDO ENDDO ! !-- Find out the number of array elements of the total array nwords = SIZE( f2 ) ! !-- Gather subdomain data from all PEs IF ( collective_wait ) CALL MPI_BARRIER( comm2d, ierr ) CALL MPI_ALLREDUCE( f2_l(nzb,nys_mg(grid_level)-1,nxl_mg(grid_level)-1), & f2(nzb,nys_mg(grid_level)-1,nxl_mg(grid_level)-1), & nwords, MPI_REAL, MPI_SUM, comm2d, ierr ) CALL cpu_log( log_point_s(34), 'mg_gather', 'stop' ) #endif END SUBROUTINE mg_gather_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> @todo It might be possible to improve the speed of this routine by using !> non-blocking communication !------------------------------------------------------------------------------! SUBROUTINE mg_scatter_fast( p2, p2_sub ) USE control_parameters, & ONLY: grid_level USE cpulog, & ONLY: cpu_log, log_point_s USE indices, & ONLY: mg_loc_ind, nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE INTEGER(iwp) :: nwords !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level-1)+1, & nys_mg(grid_level-1)-1:nyn_mg(grid_level-1)+1, & nxl_mg(grid_level-1)-1:nxr_mg(grid_level-1)+1) :: p2 !< REAL(wp), DIMENSION(nzb:mg_loc_ind(5,myid)+1, & mg_loc_ind(3,myid)-1:mg_loc_ind(4,myid)+1, & mg_loc_ind(1,myid)-1:mg_loc_ind(2,myid)+1) :: p2_sub !< ! !-- Find out the number of array elements of the subdomain array nwords = SIZE( p2_sub ) #if defined( __parallel ) CALL cpu_log( log_point_s(35), 'mg_scatter', 'start' ) p2_sub = p2(:,mg_loc_ind(3,myid)-1:mg_loc_ind(4,myid)+1, & mg_loc_ind(1,myid)-1:mg_loc_ind(2,myid)+1) CALL cpu_log( log_point_s(35), 'mg_scatter', 'stop' ) #endif END SUBROUTINE mg_scatter_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> This is where the multigrid technique takes place. V- and W- Cycle are !> implemented and steered by the parameter "gamma". Parameter "nue" determines !> the convergence of the multigrid iterative solution. There are nue times !> RB-GS iterations. It should be set to "1" or "2", considering the time effort !> one would like to invest. Last choice shows a very good converging factor, !> but leads to an increase in computing time. !------------------------------------------------------------------------------! RECURSIVE SUBROUTINE next_mg_level_fast( f_mg, p_mg, p3, r ) USE control_parameters, & ONLY: bc_lr_dirrad, bc_lr_raddir, bc_ns_dirrad, bc_ns_raddir, & gamma_mg, grid_level, grid_level_count, ibc_p_b, ibc_p_t, & inflow_l, inflow_n, inflow_r, inflow_s, maximum_grid_level, & mg_switch_to_pe0_level, mg_switch_to_pe0, nest_domain, & nest_bound_l, nest_bound_n, nest_bound_r, nest_bound_s, & ngsrb, outflow_l, outflow_n, outflow_r, outflow_s USE indices, & ONLY: mg_loc_ind, nxl, nxl_mg, nxr, nxr_mg, nys, nys_mg, nyn, & nyn_mg, nzb, nzt, nzt_mg IMPLICIT NONE INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: nxl_mg_save !< INTEGER(iwp) :: nxr_mg_save !< INTEGER(iwp) :: nyn_mg_save !< INTEGER(iwp) :: nys_mg_save !< INTEGER(iwp) :: nzt_mg_save !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: f_mg !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: p_mg !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: p3 !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: r !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level-1)+1, & nys_mg(grid_level-1)-1:nyn_mg(grid_level-1)+1, & nxl_mg(grid_level-1)-1:nxr_mg(grid_level-1)+1) :: f2 !< REAL(wp), DIMENSION(nzb:nzt_mg(grid_level-1)+1, & nys_mg(grid_level-1)-1:nyn_mg(grid_level-1)+1, & nxl_mg(grid_level-1)-1:nxr_mg(grid_level-1)+1) :: p2 !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: f2_sub !< REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: p2_sub !< ! !-- Restriction to the coarsest grid 10 IF ( grid_level == 1 ) THEN ! !-- Solution on the coarsest grid. Double the number of Gauss-Seidel !-- iterations in order to get a more accurate solution. ngsrb = 2 * ngsrb ind_even_odd = even_odd_level(grid_level) CALL redblack_fast( f_mg, p_mg ) ngsrb = ngsrb / 2 ELSEIF ( grid_level /= 1 ) THEN grid_level_count(grid_level) = grid_level_count(grid_level) + 1 ! !-- Solution on the actual grid level ind_even_odd = even_odd_level(grid_level) CALL redblack_fast( f_mg, p_mg ) ! !-- Determination of the actual residual CALL resid_fast( f_mg, p_mg, r ) !-- Restriction of the residual (finer grid values!) to the next coarser !-- grid. Therefore, the grid level has to be decremented now. nxl..nzt have !-- to be set to the coarse grid values, because these variables are needed !-- for the exchange of ghost points in routine exchange_horiz grid_level = grid_level - 1 nxl = nxl_mg(grid_level) nys = nys_mg(grid_level) nxr = nxr_mg(grid_level) nyn = nyn_mg(grid_level) nzt = nzt_mg(grid_level) IF ( grid_level == mg_switch_to_pe0_level ) THEN ! !-- From this level on, calculations are done on PE0 only. !-- First, carry out restriction on the subdomain. !-- Therefore, indices of the level have to be changed to subdomain !-- values in between (otherwise, the restrict routine would expect !-- the gathered array) nxl_mg_save = nxl_mg(grid_level) nxr_mg_save = nxr_mg(grid_level) nys_mg_save = nys_mg(grid_level) nyn_mg_save = nyn_mg(grid_level) nzt_mg_save = nzt_mg(grid_level) nxl_mg(grid_level) = mg_loc_ind(1,myid) nxr_mg(grid_level) = mg_loc_ind(2,myid) nys_mg(grid_level) = mg_loc_ind(3,myid) nyn_mg(grid_level) = mg_loc_ind(4,myid) nzt_mg(grid_level) = mg_loc_ind(5,myid) nxl = mg_loc_ind(1,myid) nxr = mg_loc_ind(2,myid) nys = mg_loc_ind(3,myid) nyn = mg_loc_ind(4,myid) nzt = mg_loc_ind(5,myid) ALLOCATE( f2_sub(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) ) CALL restrict_fast( f2_sub, r ) ! !-- Restore the correct indices of this level nxl_mg(grid_level) = nxl_mg_save nxr_mg(grid_level) = nxr_mg_save nys_mg(grid_level) = nys_mg_save nyn_mg(grid_level) = nyn_mg_save nzt_mg(grid_level) = nzt_mg_save nxl = nxl_mg(grid_level) nxr = nxr_mg(grid_level) nys = nys_mg(grid_level) nyn = nyn_mg(grid_level) nzt = nzt_mg(grid_level) ! !-- Gather all arrays from the subdomains on PE0 CALL mg_gather_fast( f2, f2_sub ) ! !-- Set switch for routine exchange_horiz, that no ghostpoint exchange !-- has to be carried out from now on mg_switch_to_pe0 = .TRUE. ! !-- In case of non-cyclic lateral boundary conditions, both in- and !-- outflow conditions have to be used on all PEs after the switch, !-- because then they have the total domain. IF ( bc_lr_dirrad ) THEN inflow_l = .TRUE. inflow_r = .FALSE. outflow_l = .FALSE. outflow_r = .TRUE. ELSEIF ( bc_lr_raddir ) THEN inflow_l = .FALSE. inflow_r = .TRUE. outflow_l = .TRUE. outflow_r = .FALSE. ELSEIF ( nest_domain ) THEN nest_bound_l = .TRUE. nest_bound_r = .TRUE. ENDIF IF ( bc_ns_dirrad ) THEN inflow_n = .TRUE. inflow_s = .FALSE. outflow_n = .FALSE. outflow_s = .TRUE. ELSEIF ( bc_ns_raddir ) THEN inflow_n = .FALSE. inflow_s = .TRUE. outflow_n = .TRUE. outflow_s = .FALSE. ELSEIF ( nest_domain ) THEN nest_bound_s = .TRUE. nest_bound_n = .TRUE. ENDIF DEALLOCATE( f2_sub ) ELSE CALL restrict_fast( f2, r ) ind_even_odd = even_odd_level(grid_level) ! must be after restrict ENDIF p2 = 0.0_wp ! !-- Repeat the same procedure till the coarsest grid is reached CALL next_mg_level_fast( f2, p2, p3, r ) ENDIF ! !-- Now follows the prolongation IF ( grid_level >= 2 ) THEN ! !-- Prolongation of the new residual. The values are transferred !-- from the coarse to the next finer grid. IF ( grid_level == mg_switch_to_pe0_level+1 ) THEN #if defined( __parallel ) ! !-- At this level, the new residual first has to be scattered from !-- PE0 to the other PEs ALLOCATE( p2_sub(nzb:mg_loc_ind(5,myid)+1, & mg_loc_ind(3,myid)-1:mg_loc_ind(4,myid)+1, & mg_loc_ind(1,myid)-1:mg_loc_ind(2,myid)+1) ) CALL mg_scatter_fast( p2, p2_sub ) ! !-- Therefore, indices of the previous level have to be changed to !-- subdomain values in between (otherwise, the prolong routine would !-- expect the gathered array) nxl_mg_save = nxl_mg(grid_level-1) nxr_mg_save = nxr_mg(grid_level-1) nys_mg_save = nys_mg(grid_level-1) nyn_mg_save = nyn_mg(grid_level-1) nzt_mg_save = nzt_mg(grid_level-1) nxl_mg(grid_level-1) = mg_loc_ind(1,myid) nxr_mg(grid_level-1) = mg_loc_ind(2,myid) nys_mg(grid_level-1) = mg_loc_ind(3,myid) nyn_mg(grid_level-1) = mg_loc_ind(4,myid) nzt_mg(grid_level-1) = mg_loc_ind(5,myid) ! !-- Set switch for routine exchange_horiz, that ghostpoint exchange !-- has to be carried again out from now on mg_switch_to_pe0 = .FALSE. ! !-- For non-cyclic lateral boundary conditions, restore the !-- in-/outflow conditions inflow_l = .FALSE.; inflow_r = .FALSE. inflow_n = .FALSE.; inflow_s = .FALSE. outflow_l = .FALSE.; outflow_r = .FALSE. outflow_n = .FALSE.; outflow_s = .FALSE. IF ( pleft == MPI_PROC_NULL ) THEN IF ( bc_lr_dirrad ) THEN inflow_l = .TRUE. ELSEIF ( bc_lr_raddir ) THEN outflow_l = .TRUE. ELSEIF ( nest_domain ) THEN nest_bound_l = .TRUE. ENDIF ENDIF IF ( pright == MPI_PROC_NULL ) THEN IF ( bc_lr_dirrad ) THEN outflow_r = .TRUE. ELSEIF ( bc_lr_raddir ) THEN inflow_r = .TRUE. ELSEIF ( nest_domain ) THEN nest_bound_r = .TRUE. ENDIF ENDIF IF ( psouth == MPI_PROC_NULL ) THEN IF ( bc_ns_dirrad ) THEN outflow_s = .TRUE. ELSEIF ( bc_ns_raddir ) THEN inflow_s = .TRUE. ELSEIF ( nest_domain ) THEN nest_bound_s = .TRUE. ENDIF ENDIF IF ( pnorth == MPI_PROC_NULL ) THEN IF ( bc_ns_dirrad ) THEN inflow_n = .TRUE. ELSEIF ( bc_ns_raddir ) THEN outflow_n = .TRUE. ELSEIF ( nest_domain ) THEN nest_bound_n = .TRUE. ENDIF ENDIF CALL prolong_fast( p2_sub, p3 ) ! !-- Restore the correct indices of the previous level nxl_mg(grid_level-1) = nxl_mg_save nxr_mg(grid_level-1) = nxr_mg_save nys_mg(grid_level-1) = nys_mg_save nyn_mg(grid_level-1) = nyn_mg_save nzt_mg(grid_level-1) = nzt_mg_save DEALLOCATE( p2_sub ) #endif ELSE CALL prolong_fast( p2, p3 ) ENDIF ! !-- Computation of the new pressure correction. Therefore, !-- values from prior grids are added up automatically stage by stage. DO i = nxl_mg(grid_level)-1, nxr_mg(grid_level)+1 DO j = nys_mg(grid_level)-1, nyn_mg(grid_level)+1 DO k = nzb, nzt_mg(grid_level)+1 p_mg(k,j,i) = p_mg(k,j,i) + p3(k,j,i) ENDDO ENDDO ENDDO ! !-- Relaxation of the new solution CALL redblack_fast( f_mg, p_mg ) ENDIF ! !-- The following few lines serve the steering of the multigrid scheme IF ( grid_level == maximum_grid_level ) THEN GOTO 20 ELSEIF ( grid_level /= maximum_grid_level .AND. grid_level /= 1 .AND. & grid_level_count(grid_level) /= gamma_mg ) THEN GOTO 10 ENDIF ! !-- Reset counter for the next call of poismg_fast grid_level_count(grid_level) = 0 ! !-- Continue with the next finer level. nxl..nzt have to be !-- set to the finer grid values, because these variables are needed for the !-- exchange of ghost points in routine exchange_horiz grid_level = grid_level + 1 ind_even_odd = even_odd_level(grid_level) nxl = nxl_mg(grid_level) nxr = nxr_mg(grid_level) nys = nys_mg(grid_level) nyn = nyn_mg(grid_level) nzt = nzt_mg(grid_level) 20 CONTINUE END SUBROUTINE next_mg_level_fast !------------------------------------------------------------------------------! ! Description: ! ------------ !> Initial settings for sorting k-dimension from sequential order (alternate !> even/odd) into blocks of even and odd or vice versa !------------------------------------------------------------------------------! SUBROUTINE init_even_odd_blocks USE arrays_3d, & ONLY: f1_mg, f2_mg, f3_mg USE control_parameters, & ONLY: grid_level, maximum_grid_level USE indices, & ONLY: nzb, nzt, nzt_mg USE indices, & ONLY: nxl_mg, nxr_mg, nys_mg, nyn_mg, nzb, nzt_mg IMPLICIT NONE ! !-- Local variables INTEGER(iwp) :: i !< INTEGER(iwp) :: l !< LOGICAL, SAVE :: lfirst = .TRUE. IF ( .NOT. lfirst ) RETURN ALLOCATE( even_odd_level(maximum_grid_level) ) ALLOCATE( f1_mg_b(nzb:nzt+1,maximum_grid_level), & f2_mg_b(nzb:nzt+1,maximum_grid_level), & f3_mg_b(nzb:nzt+1,maximum_grid_level) ) ! !-- Set border index between the even and odd block DO i = maximum_grid_level, 1, -1 even_odd_level(i) = nzt_mg(i) / 2 ENDDO ! !-- Sort grid coefficients used in red/black scheme and for calculating the !-- residual to block (even/odd) structure DO l = maximum_grid_level, 1 , -1 CALL sort_k_to_even_odd_blocks( f1_mg(nzb+1:nzt_mg(grid_level),l), & f1_mg_b(nzb:nzt_mg(grid_level)+1,l), & l ) CALL sort_k_to_even_odd_blocks( f2_mg(nzb+1:nzt_mg(grid_level),l), & f2_mg_b(nzb:nzt_mg(grid_level)+1,l), & l ) CALL sort_k_to_even_odd_blocks( f3_mg(nzb+1:nzt_mg(grid_level),l), & f3_mg_b(nzb:nzt_mg(grid_level)+1,l), & l ) ENDDO lfirst = .FALSE. END SUBROUTINE init_even_odd_blocks !------------------------------------------------------------------------------! ! Description: ! ------------ !> Special exchange_horiz subroutine for use in redblack. Transfers only !> "red" or "black" data points. !------------------------------------------------------------------------------! SUBROUTINE special_exchange_horiz ( p_mg, color ) USE control_parameters, & ONLY: bc_lr_cyc, bc_ns_cyc, grid_level, ibc_p_b, ibc_p_t, & inflow_l, inflow_n, inflow_r, inflow_s, maximum_grid_level, & mg_switch_to_pe0_level, outflow_l, outflow_n, outflow_r, & outflow_s, synchronous_exchange USE indices, & ONLY: mg_loc_ind, nxl, nxl_mg, nxr, nxr_mg, nys, nys_mg, nyn, & nyn_mg, nzb, nzt, nzt_mg IMPLICIT NONE REAL(wp), DIMENSION(nzb:nzt_mg(grid_level)+1, & nys_mg(grid_level)-1:nyn_mg(grid_level)+1, & nxl_mg(grid_level)-1:nxr_mg(grid_level)+1) :: & p_mg !< treated array INTEGER(iwp), intent(IN) :: color !< flag for grid point type (red or black) ! !-- Local variables INTEGER(iwp) :: i !< index variable along x INTEGER(iwp) :: i1 !< index variable along x on coarse level INTEGER(iwp) :: i2 !< index variable along x on coarse level INTEGER(iwp) :: j !< index variable along y INTEGER(iwp) :: j1 !< index variable along y on coarse level INTEGER(iwp) :: j2 !< index variable along y on coarse level INTEGER(iwp) :: k !< index variable along z INTEGER(iwp) :: l !< short for grid level INTEGER(iwp) :: jys !< index for lower local PE boundary along y INTEGER(iwp) :: jyn !< index for upper local PE boundary along y INTEGER(iwp) :: ixl !< index for lower local PE boundary along x INTEGER(iwp) :: ixr !< index for upper local PE boundary along x LOGICAL :: sendrecv_in_background_save !< dummy to reset sendrecv_in_background to prescribed value LOGICAL :: synchronous_exchange_save !< dummy to reset synchronous_exchange to prescribed value REAL(wp), DIMENSION(:,:,:), ALLOCATABLE :: temp !< temporary array on next coarser grid level #if defined ( __parallel ) sendrecv_in_background_save = sendrecv_in_background sendrecv_in_background = .FALSE. synchronous_exchange_save = synchronous_exchange synchronous_exchange = .FALSE. l = grid_level ind_even_odd = even_odd_level(grid_level) ! !-- Restricted transfer only on finer levels with enough data. !-- Restricted transfer is not possible for levels smaller or equal to !-- 'switch to PE0 levels', since array bounds does not fit. Moreover, !-- it is not possible for the coarsest grid level, since the dimensions !-- of temp are not defined. For such cases, normal exchange_horiz is called. IF ( l > 1 .AND. l > mg_switch_to_pe0_level + 1 .AND. & ( ngp_xz(grid_level) >= 900 .OR. ngp_yz(grid_level) >= 900 ) ) THEN jys = nys_mg(grid_level-1) jyn = nyn_mg(grid_level-1) ixl = nxl_mg(grid_level-1) ixr = nxr_mg(grid_level-1) ALLOCATE( temp(nzb:nzt_mg(l-1)+1,jys-1:jyn+1,ixl-1:ixr+1) ) ! !-- Handling the even k Values !-- Collecting data for the north - south exchange !-- Since only every second value has to be transfered, data are stored !-- on the next coarser grid level, because the arrays on that level !-- have just the required size i1 = nxl_mg(grid_level-1) i2 = nxl_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,jys,i1) = p_mg(k,j,i) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,jyn,i2) = p_mg(k,j,i) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,jys,i1) = p_mg(k,j,i) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,jyn,i2) = p_mg(k,j,i) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO grid_level = grid_level-1 nxl = nxl_mg(grid_level) nys = nys_mg(grid_level) nxr = nxr_mg(grid_level) nyn = nyn_mg(grid_level) nzt = nzt_mg(grid_level) send_receive = 'ns' CALL exchange_horiz( temp, 1 ) grid_level = grid_level+1 i1 = nxl_mg(grid_level-1) i2 = nxl_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,nyn_mg(l)+1,i) = temp(k-ind_even_odd,jyn+1,i1) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,nys_mg(l)-1,i) = temp(k-ind_even_odd,jys-1,i2) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,nyn_mg(l)+1,i) = temp(k-ind_even_odd,jyn+1,i1) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,nys_mg(l)-1,i) = temp(k-ind_even_odd,jys-1,i2) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO ! !-- Collecting data for the left - right exchange !-- Since only every second value has to be transfered, data are stored !-- on the next coarser grid level, because the arrays on that level !-- have just the required size j1 = nys_mg(grid_level-1) j2 = nys_mg(grid_level-1) DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 DO i = nxl_mg(l), nxr_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,j1,ixl) = p_mg(k,j,i) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,j2,ixr) = p_mg(k,j,i) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 DO i = nxl_mg(l)+1, nxr_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,j1,ixl) = p_mg(k,j,i) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) temp(k-ind_even_odd,j2,ixr) = p_mg(k,j,i) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO grid_level = grid_level-1 send_receive = 'lr' CALL exchange_horiz( temp, 1 ) grid_level = grid_level+1 j1 = nys_mg(grid_level-1) j2 = nys_mg(grid_level-1) DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 DO i = nxl_mg(l), nxr_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,j,nxr_mg(l)+1) = temp(k-ind_even_odd,j1,ixr+1) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,j,nxl_mg(l)-1) = temp(k-ind_even_odd,j2,ixl-1) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 DO i = nxl_mg(l)+1, nxr_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,j,nxr_mg(l)+1) = temp(k-ind_even_odd,j1,ixr+1) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = ind_even_odd+1, nzt_mg(l) p_mg(k,j,nxl_mg(l)-1) = temp(k-ind_even_odd,j2,ixl-1) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO ! !-- Now handling the even k values !-- Collecting data for the north - south exchange !-- Since only every second value has to be transfered, data are stored !-- on the next coarser grid level, because the arrays on that level !-- have just the required size i1 = nxl_mg(grid_level-1) i2 = nxl_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,jys,i1) = p_mg(k,j,i) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,jyn,i2) = p_mg(k,j,i) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,jys,i1) = p_mg(k,j,i) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,jyn,i2) = p_mg(k,j,i) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO grid_level = grid_level-1 send_receive = 'ns' CALL exchange_horiz( temp, 1 ) grid_level = grid_level+1 i1 = nxl_mg(grid_level-1) i2 = nxl_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,nyn_mg(l)+1,i) = temp(k,jyn+1,i1) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,nys_mg(l)-1,i) = temp(k,jys-1,i2) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( j == nys_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,nyn_mg(l)+1,i) = temp(k,jyn+1,i1) ENDDO i1 = i1 + 1 ENDIF IF ( j == nyn_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,nys_mg(l)-1,i) = temp(k,jys-1,i2) ENDDO i2 = i2 + 1 ENDIF ENDDO ENDDO j1 = nys_mg(grid_level-1) j2 = nys_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,j1,ixl) = p_mg(k,j,i) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,j2,ixr) = p_mg(k,j,i) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,j1,ixl) = p_mg(k,j,i) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd temp(k,j2,ixr) = p_mg(k,j,i) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO grid_level = grid_level-1 send_receive = 'lr' CALL exchange_horiz( temp, 1 ) grid_level = grid_level+1 nxl = nxl_mg(grid_level) nys = nys_mg(grid_level) nxr = nxr_mg(grid_level) nyn = nyn_mg(grid_level) nzt = nzt_mg(grid_level) j1 = nys_mg(grid_level-1) j2 = nys_mg(grid_level-1) DO i = nxl_mg(l), nxr_mg(l), 2 DO j = nys_mg(l) + (color-1), nyn_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,j,nxr_mg(l)+1) = temp(k,j1,ixr+1) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,j,nxl_mg(l)-1) = temp(k,j2,ixl-1) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO DO i = nxl_mg(l)+1, nxr_mg(l), 2 DO j = nys_mg(l) + 2 - color, nyn_mg(l), 2 IF ( i == nxl_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,j,nxr_mg(l)+1) = temp(k,j1,ixr+1) ENDDO j1 = j1 + 1 ENDIF IF ( i == nxr_mg(l) ) THEN !DIR$ IVDEP DO k = nzb+1, ind_even_odd p_mg(k,j,nxl_mg(l)-1) = temp(k,j2,ixl-1) ENDDO j2 = j2 + 1 ENDIF ENDDO ENDDO DEALLOCATE( temp ) ELSE ! !-- Standard horizontal ghost boundary exchange for small coarse grid !-- levels, where the transfer time is latency bound CALL exchange_horiz( p_mg, 1 ) ENDIF ! !-- Reset values to default PALM setup sendrecv_in_background = sendrecv_in_background_save synchronous_exchange = synchronous_exchange_save send_receive = 'al' #else ! !-- Standard horizontal ghost boundary exchange for small coarse grid !-- levels, where the transfer time is latency bound CALL exchange_horiz( p_mg, 1 ) #endif END SUBROUTINE special_exchange_horiz END MODULE poismg_mod