source: palm/trunk/SOURCE/exchange_horiz.f90 @ 1116

 SUBROUTINE exchange_horiz( ar, nbgp_local)

!--------------------------------------------------------------------------------!
! 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 <http://www.gnu.org/licenses/>.
!
! Copyright 1997-2012  Leibniz University Hannover
!--------------------------------------------------------------------------------!
!
! Current revisions:
! -----------------
! 
!
! Former revisions:
! -----------------
! $Id: exchange_horiz.f90 1114 2013-03-10 03:25:57Z hoffmann $
!
! 1113 2013-03-10 02:48:14Z raasch
! GPU-porting for single-core (1PE) mode
!
! 1036 2012-10-22 13:43:42Z raasch
! code put under GPL (PALM 3.9)
!
! 841 2012-02-28 12:29:49Z maronga
! Excluded routine from compilation of namelist_file_check
!
! 709 2011-03-30 09:31:40Z raasch
! formatting adjustments
!
! 707 2011-03-29 11:39:40Z raasch
! grid_level directly used as index for MPI data type arrays,
! bc_lr/ns replaced by bc_lr/ns_cyc
!
! 689 2011-02-20 19:31:12z gryschka 
! Bugfix for some logical expressions 
! (syntax was not compatible with all compilers)
!
! 683 2011-02-09 14:25:15Z raasch
! optional synchronous exchange (sendrecv) implemented, code partly reformatted
!
! 667 2010-12-23 12:06:00Z suehring/gryschka
! Dynamic exchange of ghost points with nbgp_local to ensure that no useless 
! ghost points exchanged in case of multigrid. type_yz(0) and type_xz(0)
! used for normal grid, the remaining types used for the several grid levels.
! Exchange is done via MPI-Vectors with a dynamic value of ghost points which
! depend on the advection scheme. Exchange of left and right PEs is 10% faster 
! with MPI-Vectors than without.
!
! 75 2007-03-22 09:54:05Z raasch
! Special cases for additional gridpoints along x or y in case of non-cyclic
! boundary conditions are not regarded any more
!
! RCS Log replace by Id keyword, revision history cleaned up
!
! Revision 1.16  2006/02/23 12:19:08  raasch
! anz_yz renamed ngp_yz
!
! Revision 1.1  1997/07/24 11:13:29  raasch
! Initial revision
!
!
! Description:
! ------------
! Exchange of lateral boundary values (parallel computers) and cyclic
! lateral boundary conditions, respectively.
!------------------------------------------------------------------------------!

    USE control_parameters
    USE cpulog
    USE indices
    USE interfaces
    USE pegrid

    IMPLICIT NONE


    INTEGER ::  i, j, k, nbgp_local
    REAL, DIMENSION(nzb:nzt+1,nys-nbgp_local:nyn+nbgp_local, &
                    nxl-nbgp_local:nxr+nbgp_local) ::  ar

#if ! defined( __check )
#if defined( __parallel )
    INTEGER, DIMENSION(4)                 ::  req
    INTEGER, DIMENSION(MPI_STATUS_SIZE,4) ::  wait_stat
#endif


    CALL cpu_log( log_point_s(2), 'exchange_horiz', 'start' )

#if defined( __parallel )

!
!-- Exchange of lateral boundary values
    IF ( pdims(1) == 1  .OR.  mg_switch_to_pe0 )  THEN
!
!--    One-dimensional decomposition along y, boundary values can be exchanged
!--    within the PE memory
       IF ( bc_lr_cyc )  THEN
          ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr)
          ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1)
       ENDIF

    ELSE

       IF ( synchronous_exchange )  THEN
!
!--       Send left boundary, receive right one (synchronous)
          CALL MPI_SENDRECV(                                                   &
              ar(nzb,nys-nbgp_local,nxl),   1, type_yz(grid_level), pleft,  0, &
              ar(nzb,nys-nbgp_local,nxr+1), 1, type_yz(grid_level), pright, 0, &
              comm2d, status, ierr )
!
!--       Send right boundary, receive left one (synchronous)
          CALL MPI_SENDRECV( ar(nzb,nys-nbgp_local,nxr+1-nbgp_local), 1, &
                             type_yz(grid_level), pright, 1,             &
                             ar(nzb,nys-nbgp_local,nxl-nbgp_local), 1,   &
                             type_yz(grid_level), pleft,  1,             &
                             comm2d, status, ierr )

       ELSE

          req = 0
!
!--       Send left boundary, receive right one (asynchronous)
          CALL MPI_ISEND( ar(nzb,nys-nbgp_local,nxl),   1, type_yz(grid_level), &
                          pleft, 0, comm2d, req(1), ierr )
          CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxr+1), 1, type_yz(grid_level), &
                          pright, 0, comm2d, req(2), ierr )
!
!--       Send right boundary, receive left one (asynchronous)
          CALL MPI_ISEND( ar(nzb,nys-nbgp_local,nxr+1-nbgp_local), 1,          &
                          type_yz(grid_level), pright, 1, comm2d, req(3), ierr )
          CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxl-nbgp_local),   1,          &
                          type_yz(grid_level), pleft,  1, comm2d, req(4), ierr )

          CALL MPI_WAITALL( 4, req, wait_stat, ierr )

       ENDIF

    ENDIF


    IF ( pdims(2) == 1  .OR.  mg_switch_to_pe0 )  THEN
!
!--    One-dimensional decomposition along x, boundary values can be exchanged
!--    within the PE memory
       IF ( bc_ns_cyc )  THEN
          ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:)
          ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:)
       ENDIF

    ELSE

       IF ( synchronous_exchange )  THEN
!
!--       Send front boundary, receive rear one (synchronous)
          CALL MPI_SENDRECV(                                                   &
              ar(nzb,nys,nxl-nbgp_local),   1, type_xz(grid_level), psouth, 0, &
              ar(nzb,nyn+1,nxl-nbgp_local), 1, type_xz(grid_level), pnorth, 0, &
              comm2d, status, ierr )
!
!--       Send rear boundary, receive front one (synchronous)
          CALL MPI_SENDRECV( ar(nzb,nyn-nbgp_local+1,nxl-nbgp_local), 1, &
                             type_xz(grid_level), pnorth, 1,             &
                             ar(nzb,nys-nbgp_local,nxl-nbgp_local),   1, &
                             type_xz(grid_level), psouth, 1,             &
                             comm2d, status, ierr )

       ELSE

          req = 0
!
!--       Send front boundary, receive rear one (asynchronous)
          CALL MPI_ISEND( ar(nzb,nys,nxl-nbgp_local),   1, type_xz(grid_level), &
                          psouth, 0, comm2d, req(1), ierr )
          CALL MPI_IRECV( ar(nzb,nyn+1,nxl-nbgp_local), 1, type_xz(grid_level), &
                          pnorth, 0, comm2d, req(2), ierr )
!
!--       Send rear boundary, receive front one (asynchronous)
          CALL MPI_ISEND( ar(nzb,nyn-nbgp_local+1,nxl-nbgp_local), 1,          &
                          type_xz(grid_level), pnorth, 1, comm2d, req(3), ierr )
          CALL MPI_IRECV( ar(nzb,nys-nbgp_local,nxl-nbgp_local),   1,          &
                          type_xz(grid_level), psouth, 1, comm2d, req(4), ierr )

          CALL MPI_WAITALL( 4, req, wait_stat, ierr )

       ENDIF

    ENDIF

#else

!
!-- Lateral boundary conditions in the non-parallel case.
!-- Case dependent, because in GPU mode still not all arrays are on device. This
!-- workaround has to be removed later. Also, since PGI compiler 12.5 has problems
!-- with array syntax, explicit loops are used.
    IF ( bc_lr == 'cyclic' )  THEN
       IF ( on_device )  THEN
          !$acc kernels present( ar )
          !$acc loop independent
          DO  i = 0, nbgp_local-1
             DO  j = nys-nbgp_local, nyn+nbgp_local
                !$acc loop vector( 32 )
                DO  k = nzb, nzt+1
                   ar(k,j,nxl-nbgp_local+i) = ar(k,j,nxr-nbgp_local+1+i)
                   ar(k,j,nxr+1+i)          = ar(k,j,nxl+i)
                ENDDO
             ENDDO
          ENDDO
          !$acc end kernels
       ELSE
          ar(:,:,nxl-nbgp_local:nxl-1) = ar(:,:,nxr-nbgp_local+1:nxr)
          ar(:,:,nxr+1:nxr+nbgp_local) = ar(:,:,nxl:nxl+nbgp_local-1)
       ENDIF
    ENDIF

    IF ( bc_ns == 'cyclic' )  THEN
       IF ( on_device )  THEN
          !$acc kernels present( ar )
          !$acc loop
          DO  i = nxl-nbgp_local, nxr+nbgp_local
             !$acc loop independent
             DO  j = 0, nbgp_local-1
                !$acc loop vector( 32 )
                DO  k = nzb, nzt+1
                   ar(k,nys-nbgp_local+j,i) = ar(k,nyn-nbgp_local+1+j,i)
                     ar(k,nyn+1+j,i)          = ar(k,nys+j,i)
                ENDDO
             ENDDO
          ENDDO
          !$acc end kernels
       ELSE
          ar(:,nys-nbgp_local:nys-1,:) = ar(:,nyn-nbgp_local+1:nyn,:)
          ar(:,nyn+1:nyn+nbgp_local,:) = ar(:,nys:nys+nbgp_local-1,:)
       ENDIF
    ENDIF

#endif
    CALL cpu_log( log_point_s(2), 'exchange_horiz', 'stop' )

#endif
 END SUBROUTINE exchange_horiz