! ! $Id$ ! MODULE parallel_lmdz USE mod_const_mpi USE lmdz_mpi, ONLY : using_mpi use IOIPSL INTEGER,PARAMETER :: halo_max=3 LOGICAL,SAVE :: using_omp ! .true. if using OpenMP LOGICAL,SAVE :: is_master ! .true. if the core is both MPI & OpenMP master !$OMP THREADPRIVATE(is_master) integer, save :: mpi_size integer, save :: mpi_rank integer, save :: jj_begin integer, save :: jj_end integer, save :: jj_nb integer, save :: ij_begin integer, save :: ij_end logical, save :: pole_nord logical, save :: pole_sud integer,save :: jjb_u integer,save :: jje_u integer,save :: jjnb_u integer,save :: jjb_v integer,save :: jje_v integer,save :: jjnb_v integer,save :: ijb_u integer,save :: ije_u integer,save :: ijnb_u integer,save :: ijb_v integer,save :: ije_v integer,save :: ijnb_v integer, allocatable, save, dimension(:) :: jj_begin_para integer, allocatable, save, dimension(:) :: jj_end_para integer, allocatable, save, dimension(:) :: jj_nb_para integer, save :: OMP_CHUNK integer, save :: omp_rank integer, save :: omp_size !$OMP THREADPRIVATE(omp_rank) TYPE distrib integer :: jj_begin integer :: jj_end integer :: jj_nb integer :: ij_begin integer :: ij_end integer :: jjb_u integer :: jje_u integer :: jjnb_u integer :: jjb_v integer :: jje_v integer :: jjnb_v integer :: ijb_u integer :: ije_u integer :: ijnb_u integer :: ijb_v integer :: ije_v integer :: ijnb_v integer, pointer :: jj_begin_para(:) => NULL() integer, pointer :: jj_end_para(:) => NULL() integer, pointer :: jj_nb_para(:) => NULL() END TYPE distrib INTERFACE ASSIGNMENT (=) MODULE PROCEDURE copy_distrib END INTERFACE TYPE(distrib),SAVE :: current_dist contains subroutine init_parallel USE iniprint_mod_h USE vampir USE lmdz_mpi USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none integer :: ierr integer :: i,j integer :: type_size integer, dimension(3) :: blocklen,type integer :: comp_id character(len=4) :: num character(len=20) :: filename #ifdef CPP_OMP INTEGER :: OMP_GET_NUM_THREADS EXTERNAL OMP_GET_NUM_THREADS INTEGER :: OMP_GET_THREAD_NUM EXTERNAL OMP_GET_THREAD_NUM #endif #ifdef CPP_OMP using_OMP=.TRUE. #else using_OMP=.FALSE. #endif call InitVampir IF (using_mpi) THEN call MPI_COMM_SIZE(COMM_LMDZ,mpi_size,ierr) call MPI_COMM_RANK(COMM_LMDZ,mpi_rank,ierr) ELSE mpi_size=1 mpi_rank=0 ENDIF ! Open text output file with mpi_rank in suffix of file name IF (lunout /= 5 .and. lunout /= 6) THEN WRITE(num,'(I4.4)') mpi_rank filename='lmdz.out_'//num IF (mpi_rank .NE. 0) THEN OPEN(UNIT=lunout,FILE=TRIM(filename),ACTION='write', & STATUS='unknown',FORM='formatted',IOSTAT=ierr) ENDIF ENDIF allocate(jj_begin_para(0:mpi_size-1)) allocate(jj_end_para(0:mpi_size-1)) allocate(jj_nb_para(0:mpi_size-1)) do i=0,mpi_size-1 jj_nb_para(i)=(jjm+1)/mpi_size if ( i < MOD((jjm+1),mpi_size) ) jj_nb_para(i)=jj_nb_para(i)+1 if (jj_nb_para(i) <= 1 ) then write(lunout,*)"Arret : le nombre de bande de lattitude par process est trop faible (<2)." write(lunout,*)" ---> diminuez le nombre de CPU ou augmentez la taille en lattitude" IF (using_mpi) call MPI_ABORT(COMM_LMDZ,-1, ierr) endif enddo ! jj_nb_para(0)=11 ! jj_nb_para(1)=25 ! jj_nb_para(2)=25 ! jj_nb_para(3)=12 j=1 do i=0,mpi_size-1 jj_begin_para(i)=j jj_end_para(i)=j+jj_Nb_para(i)-1 j=j+jj_Nb_para(i) enddo jj_begin = jj_begin_para(mpi_rank) jj_end = jj_end_para(mpi_rank) jj_nb = jj_nb_para(mpi_rank) ij_begin=(jj_begin-1)*iip1+1 ij_end=jj_end*iip1 if (mpi_rank.eq.0) then pole_nord=.TRUE. else pole_nord=.FALSE. endif if (mpi_rank.eq.mpi_size-1) then pole_sud=.TRUE. else pole_sud=.FALSE. endif write(lunout,*)"init_parallel: jj_begin",jj_begin write(lunout,*)"init_parallel: jj_end",jj_end write(lunout,*)"init_parallel: ij_begin",ij_begin write(lunout,*)"init_parallel: ij_end",ij_end jjb_u=MAX(jj_begin-halo_max,1) jje_u=MIN(jj_end+halo_max,jjp1) jjnb_u=jje_u-jjb_u+1 jjb_v=MAX(jj_begin-halo_max,1) jje_v=MIN(jj_end+halo_max,jjm) jjnb_v=jje_v-jjb_v+1 ijb_u=MAX(ij_begin-halo_max*iip1,1) ije_u=MIN(ij_end+halo_max*iip1,ip1jmp1) ijnb_u=ije_u-ijb_u+1 ijb_v=MAX(ij_begin-halo_max*iip1,1) ije_v=MIN(ij_end+halo_max*iip1,ip1jm) ijnb_v=ije_v-ijb_v+1 !$OMP PARALLEL #ifdef CPP_OMP !$OMP MASTER omp_size=OMP_GET_NUM_THREADS() !$OMP END MASTER !$OMP BARRIER omp_rank=OMP_GET_THREAD_NUM() !Config Key = omp_chunk !Config Desc = taille des blocs openmp !Config Def = 1 !Config Help = defini la taille des packets d'it�ration openmp !Config distribue a chaque tache lors de l'entree dans une !Config boucle parallelisee !$OMP MASTER omp_chunk=(llm+1)/omp_size IF (MOD(llm+1,omp_size)/=0) omp_chunk=omp_chunk+1 CALL getin('omp_chunk',omp_chunk) !$OMP END MASTER !$OMP BARRIER #else omp_size=1 omp_rank=0 #endif !$OMP END PARALLEL CALL create_distrib(jj_nb_para,current_dist) IF ((mpi_rank==0).and.(omp_rank==0)) THEN is_master=.true. ELSE is_master=.false. ENDIF end subroutine init_parallel SUBROUTINE create_distrib(jj_nb_new,d) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h IMPLICIT NONE INTEGER,INTENT(IN) :: jj_Nb_New(0:MPI_Size-1) TYPE(distrib),INTENT(INOUT) :: d INTEGER :: i IF (.NOT. ASSOCIATED(d%jj_nb_para)) ALLOCATE(d%jj_nb_para(0:MPI_Size-1)) IF (.NOT. ASSOCIATED(d%jj_begin_para)) ALLOCATE(d%jj_begin_para(0:MPI_Size-1)) IF (.NOT. ASSOCIATED(d%jj_end_para)) ALLOCATE(d%jj_end_para(0:MPI_Size-1)) d%jj_Nb_Para=jj_Nb_New d%jj_begin_para(0)=1 d%jj_end_para(0)=d%jj_Nb_Para(0) do i=1,mpi_size-1 d%jj_begin_para(i)=d%jj_end_para(i-1)+1 d%jj_end_para(i)=d%jj_begin_para(i)+d%jj_Nb_para(i)-1 enddo d%jj_begin = d%jj_begin_para(mpi_rank) d%jj_end = d%jj_end_para(mpi_rank) d%jj_nb = d%jj_nb_para(mpi_rank) d%ij_begin=(d%jj_begin-1)*iip1+1 d%ij_end=d%jj_end*iip1 d%jjb_u=MAX(d%jj_begin-halo_max,1) d%jje_u=MIN(d%jj_end+halo_max,jjp1) d%jjnb_u=d%jje_u-d%jjb_u+1 d%jjb_v=MAX(d%jj_begin-halo_max,1) d%jje_v=MIN(d%jj_end+halo_max,jjm) d%jjnb_v=d%jje_v-d%jjb_v+1 d%ijb_u=MAX(d%ij_begin-halo_max*iip1,1) d%ije_u=MIN(d%ij_end+halo_max*iip1,ip1jmp1) d%ijnb_u=d%ije_u-d%ijb_u+1 d%ijb_v=MAX(d%ij_begin-halo_max*iip1,1) d%ije_v=MIN(d%ij_end+halo_max*iip1,ip1jm) d%ijnb_v=d%ije_v-d%ijb_v+1 END SUBROUTINE create_distrib SUBROUTINE Set_Distrib(d) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h IMPLICIT NONE TYPE(distrib),INTENT(IN) :: d jj_begin = d%jj_begin jj_end = d%jj_end jj_nb = d%jj_nb ij_begin = d%ij_begin ij_end = d%ij_end jjb_u = d%jjb_u jje_u = d%jje_u jjnb_u = d%jjnb_u jjb_v = d%jjb_v jje_v = d%jje_v jjnb_v = d%jjnb_v ijb_u = d%ijb_u ije_u = d%ije_u ijnb_u = d%ijnb_u ijb_v = d%ijb_v ije_v = d%ije_v ijnb_v = d%ijnb_v jj_begin_para(:) = d%jj_begin_para(:) jj_end_para(:) = d%jj_end_para(:) jj_nb_para(:) = d%jj_nb_para(:) current_dist=d END SUBROUTINE Set_Distrib SUBROUTINE copy_distrib(dist,new_dist) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h IMPLICIT NONE TYPE(distrib),INTENT(INOUT) :: dist TYPE(distrib),INTENT(IN) :: new_dist dist%jj_begin = new_dist%jj_begin dist%jj_end = new_dist%jj_end dist%jj_nb = new_dist%jj_nb dist%ij_begin = new_dist%ij_begin dist%ij_end = new_dist%ij_end dist%jjb_u = new_dist%jjb_u dist%jje_u = new_dist%jje_u dist%jjnb_u = new_dist%jjnb_u dist%jjb_v = new_dist%jjb_v dist%jje_v = new_dist%jje_v dist%jjnb_v = new_dist%jjnb_v dist%ijb_u = new_dist%ijb_u dist%ije_u = new_dist%ije_u dist%ijnb_u = new_dist%ijnb_u dist%ijb_v = new_dist%ijb_v dist%ije_v = new_dist%ije_v dist%ijnb_v = new_dist%ijnb_v dist%jj_begin_para(:) = new_dist%jj_begin_para(:) dist%jj_end_para(:) = new_dist%jj_end_para(:) dist%jj_nb_para(:) = new_dist%jj_nb_para(:) END SUBROUTINE copy_distrib SUBROUTINE get_current_distrib(d) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h IMPLICIT NONE TYPE(distrib),INTENT(OUT) :: d d=current_dist END SUBROUTINE get_current_distrib subroutine Finalize_parallel USE lmdz_mpi ! ug Pour les sorties XIOS USE wxios USE control_mod, only : ok_dyn_xios USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h #ifdef CPP_COUPLE ! Use of Oasis-MCT coupler #if defined CPP_OMCT use mod_prism #else use mod_prism_proto #endif ! Ehouarn: surface_data module is in 'phylmd' ... use surface_data, only : type_ocean implicit none #else implicit none ! without the surface_data module, we declare (and set) a dummy 'type_ocean' character(len=6),parameter :: type_ocean="dummy" #endif ! #endif of #ifdef CPP_EARTH integer :: ierr integer :: i if (allocated(jj_begin_para)) deallocate(jj_begin_para) if (allocated(jj_end_para)) deallocate(jj_end_para) if (allocated(jj_nb_para)) deallocate(jj_nb_para) if (type_ocean == 'couple') then #ifdef CPP_COUPLE IF (using_xios) THEN !Fermeture propre de XIOS ! close xios dynamic context if is defined (call LMDZDYN) IF (ok_dyn_xios) THEN CALL xios_context_finalize() ENDIF CALL wxios_close() CALL prism_terminate_proto(ierr) IF (ierr .ne. PRISM_Ok) THEN CALL abort_gcm('Finalize_parallel',' Probleme dans prism_terminate_proto ',1) ENDIF ELSE call prism_terminate_proto(ierr) IF (ierr .ne. PRISM_Ok) THEN call abort_gcm('Finalize_parallel',' Probleme dans prism_terminate_proto ',1) endif ENDIF #else call abort_gcm('Finalize_parallel','type_ocean = couple but CPP_COUPLE not defined',1) #endif else IF (using_xios) THEN !Fermeture propre de XIOS IF (ok_dyn_xios) THEN CALL xios_context_finalize() ENDIF CALL wxios_close() ENDIF IF (using_mpi) call MPI_FINALIZE(ierr) end if end subroutine Finalize_parallel subroutine Pack_Data(Field,ij,ll,row,Buffer) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none integer, intent(in) :: ij,ll,row real,dimension(ij,ll),intent(in) ::Field real,dimension(ll*iip1*row), intent(out) :: Buffer integer :: Pos integer :: i,l Pos=0 do l=1,ll do i=1,row*iip1 Pos=Pos+1 Buffer(Pos)=Field(i,l) enddo enddo end subroutine Pack_data subroutine Unpack_Data(Field,ij,ll,row,Buffer) USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none integer, intent(in) :: ij,ll,row real,dimension(ij,ll),intent(out) ::Field real,dimension(ll*iip1*row), intent(in) :: Buffer integer :: Pos integer :: i,l Pos=0 do l=1,ll do i=1,row*iip1 Pos=Pos+1 Field(i,l)=Buffer(Pos) enddo enddo end subroutine UnPack_data SUBROUTINE barrier USE lmdz_mpi IMPLICIT NONE INTEGER :: ierr !$OMP CRITICAL (MPI) IF (using_mpi) CALL MPI_Barrier(COMM_LMDZ,ierr) !$OMP END CRITICAL (MPI) END SUBROUTINE barrier subroutine exchange_hallo(Field,ij,ll,up,down) USE lmdz_mpi USE Vampir USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none INTEGER :: ij,ll REAL, dimension(ij,ll) :: Field INTEGER :: up,down INTEGER :: ierr LOGICAL :: SendUp,SendDown LOGICAL :: RecvUp,RecvDown INTEGER, DIMENSION(4) :: Request INTEGER, DIMENSION(MPI_STATUS_SIZE,4) :: Status INTEGER :: NbRequest REAL, dimension(:),allocatable :: Buffer_Send_up,Buffer_Send_down REAL, dimension(:),allocatable :: Buffer_Recv_up,Buffer_Recv_down INTEGER :: Buffer_size IF (using_mpi) THEN CALL barrier call VTb(VThallo) SendUp=.TRUE. SendDown=.TRUE. RecvUp=.TRUE. RecvDown=.TRUE. IF (pole_nord) THEN SendUp=.FALSE. RecvUp=.FALSE. ENDIF IF (pole_sud) THEN SendDown=.FALSE. RecvDown=.FALSE. ENDIF if (up.eq.0) then SendDown=.FALSE. RecvUp=.FALSE. endif if (down.eq.0) then SendUp=.FALSE. RecvDown=.FALSE. endif NbRequest=0 IF (SendUp) THEN NbRequest=NbRequest+1 buffer_size=down*iip1*ll allocate(Buffer_Send_up(Buffer_size)) call PACK_Data(Field(ij_begin,1),ij,ll,down,Buffer_Send_up) !$OMP CRITICAL (MPI) call MPI_ISEND(Buffer_send_up,Buffer_Size,MPI_REAL8,MPI_Rank-1,1, & COMM_LMDZ,Request(NbRequest),ierr) !$OMP END CRITICAL (MPI) ENDIF IF (SendDown) THEN NbRequest=NbRequest+1 buffer_size=up*iip1*ll allocate(Buffer_Send_down(Buffer_size)) call PACK_Data(Field(ij_end+1-up*iip1,1),ij,ll,up,Buffer_send_down) !$OMP CRITICAL (MPI) call MPI_ISEND(Buffer_send_down,Buffer_Size,MPI_REAL8,MPI_Rank+1,1, & COMM_LMDZ,Request(NbRequest),ierr) !$OMP END CRITICAL (MPI) ENDIF IF (RecvUp) THEN NbRequest=NbRequest+1 buffer_size=up*iip1*ll allocate(Buffer_recv_up(Buffer_size)) !$OMP CRITICAL (MPI) call MPI_IRECV(Buffer_recv_up,Buffer_size,MPI_REAL8,MPI_Rank-1,1, & COMM_LMDZ,Request(NbRequest),ierr) !$OMP END CRITICAL (MPI) ENDIF IF (RecvDown) THEN NbRequest=NbRequest+1 buffer_size=down*iip1*ll allocate(Buffer_recv_down(Buffer_size)) !$OMP CRITICAL (MPI) call MPI_IRECV(Buffer_recv_down,Buffer_size,MPI_REAL8,MPI_Rank+1,1, & COMM_LMDZ,Request(NbRequest),ierr) !$OMP END CRITICAL (MPI) ENDIF if (NbRequest > 0) call MPI_WAITALL(NbRequest,Request,Status,ierr) IF (RecvUp) call Unpack_Data(Field(ij_begin-up*iip1,1),ij,ll,up,Buffer_Recv_up) IF (RecvDown) call Unpack_Data(Field(ij_end+1,1),ij,ll,down,Buffer_Recv_down) call VTe(VThallo) call barrier ENDIF ! using_mpi RETURN end subroutine exchange_Hallo subroutine Gather_Field(Field,ij,ll,rank) USE iniprint_mod_h USE lmdz_mpi USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none INTEGER :: ij,ll,rank REAL, dimension(ij,ll) :: Field REAL, dimension(:),allocatable :: Buffer_send REAL, dimension(:),allocatable :: Buffer_Recv INTEGER, dimension(0:MPI_Size-1) :: Recv_count, displ INTEGER :: ierr INTEGER ::i IF (using_mpi) THEN if (ij==ip1jmp1) then allocate(Buffer_send(iip1*ll*(jj_end-jj_begin+1))) call Pack_Data(Field(ij_begin,1),ij,ll,jj_end-jj_begin+1,Buffer_send) else if (ij==ip1jm) then allocate(Buffer_send(iip1*ll*(min(jj_end,jjm)-jj_begin+1))) call Pack_Data(Field(ij_begin,1),ij,ll,min(jj_end,jjm)-jj_begin+1,Buffer_send) else write(lunout,*)ij CALL abort_gcm("parallel_lmdz","erreur dans Gather_Field",1) endif if (MPI_Rank==rank) then allocate(Buffer_Recv(ij*ll)) !CDIR NOVECTOR do i=0,MPI_Size-1 if (ij==ip1jmp1) then Recv_count(i)=(jj_end_para(i)-jj_begin_para(i)+1)*ll*iip1 else if (ij==ip1jm) then Recv_count(i)=(min(jj_end_para(i),jjm)-jj_begin_para(i)+1)*ll*iip1 else CALL abort_gcm("parallel_lmdz","erreur dans Gather_Field",1) endif if (i==0) then displ(i)=0 else displ(i)=displ(i-1)+Recv_count(i-1) endif enddo else ! Ehouarn: When in debug mode, ifort complains (for call MPI_GATHERV ! below) about Buffer_Recv() being not allocated. ! So make a dummy allocation. allocate(Buffer_Recv(1)) endif ! of if (MPI_Rank==rank) !$OMP CRITICAL (MPI) call MPI_GATHERV(Buffer_send,(min(ij_end,ij)-ij_begin+1)*ll,MPI_REAL8, & Buffer_Recv,Recv_count,displ,MPI_REAL8,rank,COMM_LMDZ,ierr) !$OMP END CRITICAL (MPI) if (MPI_Rank==rank) then if (ij==ip1jmp1) then do i=0,MPI_Size-1 call Unpack_Data(Field((jj_begin_para(i)-1)*iip1+1,1),ij,ll, & jj_end_para(i)-jj_begin_para(i)+1,Buffer_Recv(displ(i)+1)) enddo else if (ij==ip1jm) then do i=0,MPI_Size-1 call Unpack_Data(Field((jj_begin_para(i)-1)*iip1+1,1),ij,ll, & min(jj_end_para(i),jjm)-jj_begin_para(i)+1,Buffer_Recv(displ(i)+1)) enddo endif endif ENDIF ! using_mpi end subroutine Gather_Field subroutine AllGather_Field(Field,ij,ll) USE lmdz_mpi USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none INTEGER :: ij,ll REAL, dimension(ij,ll) :: Field INTEGER :: ierr IF (using_mpi) THEN call Gather_Field(Field,ij,ll,0) !$OMP CRITICAL (MPI) call MPI_BCAST(Field,ij*ll,MPI_REAL8,0,COMM_LMDZ,ierr) !$OMP END CRITICAL (MPI) ENDIF end subroutine AllGather_Field subroutine Broadcast_Field(Field,ij,ll,rank) USE lmdz_mpi USE dimensions_mod, ONLY: iim, jjm, llm, ndm USE paramet_mod_h implicit none INTEGER :: ij,ll REAL, dimension(ij,ll) :: Field INTEGER :: rank INTEGER :: ierr IF (using_mpi) THEN !$OMP CRITICAL (MPI) call MPI_BCAST(Field,ij*ll,MPI_REAL8,rank,COMM_LMDZ,ierr) !$OMP END CRITICAL (MPI) ENDIF end subroutine Broadcast_Field ! Subroutine verif_hallo(Field,ij,ll,up,down) ! USE lmdz_mpi ! USE dimensions_mod, ONLY: iim, jjm, llm, ndm ! USE paramet_mod_h, ONLY: iip1, iip2, iip3, jjp1, llmp1, llmp2, llmm1, kftd, ip1jm, ip1jmp1, & ! ip1jmi1, ijp1llm, ijmllm, mvar, jcfil, jcfllm !implicit none ! ! ! ! INTEGER :: ij,ll ! REAL, dimension(ij,ll) :: Field ! INTEGER :: up,down ! ! REAL,dimension(ij,ll): NewField ! ! NewField=0 ! ! ijb=ij_begin ! ije=ij_end ! if (pole_nord) ! NewField(ij_be end MODULE parallel_lmdz