! $Id$

  module Bands
  USE parallel_lmdz
    INTEGER, parameter :: bands_caldyn=1
    INTEGER, parameter :: bands_vanleer=2
    INTEGER, parameter :: bands_dissip=3
    
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_Caldyn
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_vanleer
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_vanleer2
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_dissip
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_physic
    INTEGER,DIMENSION(:),ALLOCATABLE :: jj_Nb_physic_bis
   
    TYPE(distrib),SAVE,TARGET :: distrib_Caldyn
    TYPE(distrib),SAVE,TARGET :: distrib_vanleer
    TYPE(distrib),SAVE,TARGET :: distrib_vanleer2
    TYPE(distrib),SAVE,TARGET :: distrib_dissip
    TYPE(distrib),SAVE,TARGET :: distrib_physic
    TYPE(distrib),SAVE,TARGET :: distrib_physic_bis

    INTEGER,DIMENSION(:),ALLOCATABLE :: distrib_phys
  
  CONTAINS
  
  SUBROUTINE AllocateBands
    USE parallel_lmdz
    IMPLICIT NONE
    
    allocate(jj_Nb_Caldyn(0:MPI_Size-1))
    allocate(jj_Nb_vanleer(0:MPI_Size-1))
    allocate(jj_Nb_vanleer2(0:MPI_Size-1))
    allocate(jj_Nb_dissip(0:MPI_Size-1))
    allocate(jj_Nb_physic(0:MPI_Size-1))
    allocate(jj_Nb_physic_bis(0:MPI_Size-1))
    allocate(distrib_phys(0:MPI_Size-1))
  
  END SUBROUTINE  AllocateBands
  
  SUBROUTINE Read_distrib
    USE parallel_lmdz
USE lmdz_dimensions, ONLY: iim, jjm, llm, ndm
    IMPLICIT NONE


      INTEGER :: i,j
      CHARACTER (LEN=4) :: siim,sjjm,sllm,sproc
      CHARACTER (LEN=255) :: filename
      INTEGER :: unit_number=10
      INTEGER :: ierr
    
      CALL AllocateBands
      WRITE(siim,'(i3)') iim
      WRITE(sjjm,'(i3)') jjm
      WRITE(sllm,'(i3)') llm
      WRITE(sproc,'(i3)') mpi_size
      filename='Bands_'//TRIM(ADJUSTL(siim))//'x'//TRIM(ADJUSTL(sjjm))//'x'//TRIM(ADJUSTL(sllm))//'_'  &
                        //TRIM(ADJUSTL(sproc))//'prc.dat'    
       
      OPEN(UNIT=unit_number,FILE=trim(filename),STATUS='old',FORM='formatted',IOSTAT=ierr)
      
      IF (ierr==0) THEN
         DO i=0,mpi_size-1
          read (unit_number,*) j,jj_nb_caldyn(i)
        enddo
      
        DO i=0,mpi_size-1
          read (unit_number,*) j,jj_nb_vanleer(i)
        enddo
      
        DO i=0,mpi_size-1
          read (unit_number,*) j,jj_nb_dissip(i)
        enddo
      
        DO i=0,mpi_size-1
          read (unit_number,*) j,distrib_phys(i)
        enddo

    CLOSE(unit_number)
  
      else
        DO i=0,mpi_size-1
          jj_nb_caldyn(i)=(jjm+1)/mpi_size
      IF (i<MOD(jjm+1,mpi_size)) jj_nb_caldyn(i)=jj_nb_caldyn(i)+1
        enddo
      
        jj_nb_vanleer(:)=jj_nb_caldyn(:)
        jj_nb_dissip(:)=jj_nb_caldyn(:)
        
    DO i=0,mpi_size-1
      distrib_phys(i)=(iim*(jjm-1)+2)/mpi_size
      IF (i<MOD(iim*(jjm-1)+2,mpi_size)) distrib_phys(i)=distrib_phys(i)+1
    enddo
      endif
      
!      distrib_phys(:)=jj_nb_caldyn(:)*iim
!      distrib_phys(0) = distrib_phys(0) - (iim-1)
!      distrib_phys(mpi_size-1) = distrib_phys(mpi_size-1) - (iim-1)
      
   END SUBROUTINE  Read_distrib
   
   
   SUBROUTINE  Set_Bands 
     USE parallel_lmdz
USE lmdz_dimensions, ONLY: iim, jjm, llm, ndm
     IMPLICIT NONE

     INTEGER :: i, ij
     INTEGER :: jj_para_begin(0:mpi_size-1)
     INTEGER :: jj_para_end(0:mpi_size-1)
        
      DO i=0,mpi_size-1
         jj_nb_vanleer2(i)=(jjm+1)/mpi_size
     IF (i<MOD(jjm+1,mpi_size)) jj_nb_vanleer2(i)=jj_nb_vanleer2(i)+1
      enddo
          
      jj_para_begin(0)=1
      ij=distrib_phys(0)+iim-1
      jj_para_end(0)=((ij-1)/iim)+1
      
      DO i=1,mpi_Size-1
        ij=ij+1
        jj_para_begin(i)=((ij-1)/iim)+1
        ij=ij+distrib_phys(i)-1
        jj_para_end(i)=((ij-1)/iim)+1
      ENDDO
 
       DO i=0,MPI_Size-1
        jj_Nb_physic(i)=jj_para_end(i)-jj_para_begin(i)+1
        IF (i/=0) THEN
          IF (jj_para_begin(i)==jj_para_end(i-1)) THEN
            jj_Nb_physic(i-1)=jj_Nb_physic(i-1)-1
          endif
        endif
      enddo
      
      DO i=0,MPI_Size-1
        jj_Nb_physic_bis(i)=jj_para_end(i)-jj_para_begin(i)+1
        IF (i/=0) THEN
          IF (jj_para_begin(i)==jj_para_end(i-1)) THEN
            jj_Nb_physic_bis(i)=jj_Nb_physic_bis(i)-1
          else
        jj_Nb_physic_bis(i-1)=jj_Nb_physic_bis(i-1)+1
        jj_Nb_physic_bis(i)=jj_Nb_physic_bis(i)-1
      ENDIF
        endif
      enddo

      CALL create_distrib(jj_Nb_Caldyn,distrib_caldyn)
      CALL create_distrib(jj_Nb_vanleer,distrib_vanleer)
      CALL create_distrib(jj_Nb_vanleer2,distrib_vanleer2)
      CALL create_distrib(jj_Nb_dissip,distrib_dissip)
      CALL create_distrib(jj_Nb_physic,distrib_physic)
      CALL create_distrib(jj_Nb_physic_bis,distrib_physic_bis)
      
      distrib_physic_bis%jjb_u=distrib_physic%jjb_u
      distrib_physic_bis%jje_u=distrib_physic%jje_u
      distrib_physic_bis%jjnb_u=distrib_physic%jjnb_u

      distrib_physic_bis%ijb_u=distrib_physic%ijb_u
      distrib_physic_bis%ije_u=distrib_physic%ije_u
      distrib_physic_bis%ijnb_u=distrib_physic%ijnb_u

      distrib_physic_bis%jjb_v=distrib_physic%jjb_v
      distrib_physic_bis%jje_v=distrib_physic%jje_v
      distrib_physic_bis%jjnb_v=distrib_physic%jjnb_v

      distrib_physic_bis%ijb_v=distrib_physic%ijb_v
      distrib_physic_bis%ije_v=distrib_physic%ije_v
      distrib_physic_bis%ijnb_v=distrib_physic%ijnb_v
     
    END SUBROUTINE  Set_Bands


    SUBROUTINE AdjustBands_caldyn(new_dist)
      USE times
      USE parallel_lmdz
      IMPLICIT NONE
      TYPE(distrib),INTENT(INOUT) :: new_dist

      REAL :: minvalue,maxvalue
      INTEGER :: min_proc,max_proc
      INTEGER :: i,j
      REAL,ALLOCATABLE,DIMENSION(:) :: value
      INTEGER,ALLOCATABLE,DIMENSION(:) :: index
      REAL :: tmpvalue
      INTEGER :: tmpindex
      
      allocate(value(0:mpi_size-1))
      allocate(index(0:mpi_size-1))
        
  
      CALL allgather_timer_average

      DO i=0,mpi_size-1
        value(i)=timer_average(jj_nb_caldyn(i),timer_caldyn,i)
    index(i)=i
      enddo
      
      DO i=0,mpi_size-2
        DO j=i+1,mpi_size-1
      IF (value(i)>value(j)) THEN
        tmpvalue=value(i)
        value(i)=value(j)
        value(j)=tmpvalue

        tmpindex=index(i)
        index(i)=index(j)
        index(j)=tmpindex
       endif
     enddo
      enddo
      
      maxvalue=value(mpi_size-1)
      max_proc=index(mpi_size-1)           
           
      DO i=0,mpi_size-2
        minvalue=value(i)
        min_proc=index(i)
        IF (jj_nb_caldyn(max_proc)>2) THEN
          IF (timer_iteration(jj_nb_caldyn(min_proc)+1,timer_caldyn,min_proc)<=1 ) THEN
             jj_nb_caldyn(min_proc)=jj_nb_caldyn(min_proc)+1
             jj_nb_caldyn(max_proc)=jj_nb_caldyn(max_proc)-1
         exit
           else
             IF (timer_average(jj_nb_caldyn(min_proc)+1,timer_caldyn,min_proc)                 &
            -timer_delta(jj_nb_caldyn(min_proc)+1,timer_caldyn,min_proc) < maxvalue) THEN
               jj_nb_caldyn(min_proc)=jj_nb_caldyn(min_proc)+1
               jj_nb_caldyn(max_proc)=jj_nb_caldyn(max_proc)-1
               exit
         endif
           endif
         endif
      enddo
      
      deallocate(value)
      deallocate(index)
      CALL create_distrib(jj_nb_caldyn,new_dist)
        
    END SUBROUTINE  AdjustBands_caldyn
    
    SUBROUTINE AdjustBands_vanleer(new_dist)
      USE times
      USE parallel_lmdz
      IMPLICIT NONE
      TYPE(distrib),INTENT(INOUT) :: new_dist

      REAL :: minvalue,maxvalue
      INTEGER :: min_proc,max_proc
      INTEGER :: i,j
      REAL,ALLOCATABLE,DIMENSION(:) :: value
      INTEGER,ALLOCATABLE,DIMENSION(:) :: index
      REAL :: tmpvalue
      INTEGER :: tmpindex
      
      allocate(value(0:mpi_size-1))
      allocate(index(0:mpi_size-1))
        
  
      CALL allgather_timer_average

      DO i=0,mpi_size-1
        value(i)=timer_average(jj_nb_vanleer(i),timer_vanleer,i)
    index(i)=i
      enddo
      
      DO i=0,mpi_size-2
        DO j=i+1,mpi_size-1
      IF (value(i)>value(j)) THEN
        tmpvalue=value(i)
        value(i)=value(j)
        value(j)=tmpvalue

        tmpindex=index(i)
        index(i)=index(j)
        index(j)=tmpindex
       endif
     enddo
      enddo
      
      maxvalue=value(mpi_size-1)
      max_proc=index(mpi_size-1)           
           
      DO i=0,mpi_size-2
        minvalue=value(i)
        min_proc=index(i)

        IF (jj_nb_vanleer(max_proc)>2) THEN
          IF (timer_average(jj_nb_vanleer(min_proc)+1,timer_vanleer,min_proc)==0. .OR. &
             timer_average(jj_nb_vanleer(max_proc)-1,timer_vanleer,max_proc)==0.) THEN
             jj_nb_vanleer(min_proc)=jj_nb_vanleer(min_proc)+1
             jj_nb_vanleer(max_proc)=jj_nb_vanleer(max_proc)-1
         exit
           else
             IF (timer_average(jj_nb_vanleer(min_proc)+1,timer_vanleer,min_proc) < maxvalue) THEN
               jj_nb_vanleer(min_proc)=jj_nb_vanleer(min_proc)+1
               jj_nb_vanleer(max_proc)=jj_nb_vanleer(max_proc)-1
               exit
         endif
           endif
         endif
      enddo
      
      deallocate(value)
      deallocate(index)
 
      CALL create_distrib(jj_nb_vanleer,new_dist)
         
    END SUBROUTINE  AdjustBands_vanleer

    SUBROUTINE AdjustBands_dissip(new_dist)
      USE times
      USE parallel_lmdz
      IMPLICIT NONE
      TYPE(distrib),INTENT(INOUT) :: new_dist
      
      REAL :: minvalue,maxvalue
      INTEGER :: min_proc,max_proc
      INTEGER :: i,j
      REAL,ALLOCATABLE,DIMENSION(:) :: value
      INTEGER,ALLOCATABLE,DIMENSION(:) :: index
      REAL :: tmpvalue
      INTEGER :: tmpindex
      
      allocate(value(0:mpi_size-1))
      allocate(index(0:mpi_size-1))
        
  
      CALL allgather_timer_average

      DO i=0,mpi_size-1
        value(i)=timer_average(jj_nb_dissip(i),timer_dissip,i)
    index(i)=i
      enddo
      
      DO i=0,mpi_size-2
        DO j=i+1,mpi_size-1
      IF (value(i)>value(j)) THEN
        tmpvalue=value(i)
        value(i)=value(j)
        value(j)=tmpvalue

        tmpindex=index(i)
        index(i)=index(j)
        index(j)=tmpindex
       endif
     enddo
      enddo
      
      maxvalue=value(mpi_size-1)
      max_proc=index(mpi_size-1)           
           
      DO i=0,mpi_size-2
        minvalue=value(i)
        min_proc=index(i)

        IF (jj_nb_dissip(max_proc)>3) THEN
          IF (timer_iteration(jj_nb_dissip(min_proc)+1,timer_dissip,min_proc)<=1) THEN
             jj_nb_dissip(min_proc)=jj_nb_dissip(min_proc)+1
             jj_nb_dissip(max_proc)=jj_nb_dissip(max_proc)-1
         exit
           else
             IF (timer_average(jj_nb_dissip(min_proc)+1,timer_dissip,min_proc)         &
            - timer_delta(jj_nb_dissip(min_proc)+1,timer_dissip,min_proc) < maxvalue) THEN
               jj_nb_dissip(min_proc)=jj_nb_dissip(min_proc)+1
               jj_nb_dissip(max_proc)=jj_nb_dissip(max_proc)-1
               exit
         endif
           endif
         endif
      enddo
      
      deallocate(value)
      deallocate(index)
  
      CALL create_distrib(jj_nb_dissip,new_dist)
         
    END SUBROUTINE  AdjustBands_dissip

    SUBROUTINE AdjustBands_physic
      USE times

      ! Ehouarn: what follows is only related to // physics
      USE lmdz_cppkeys_wrapper, ONLY: CPPKEY_PHYS
      USE lmdz_phys_para, ONLY: klon_mpi_para_nb

      USE parallel_lmdz
      IMPLICIT NONE

      INTEGER :: i,Index
      REAL,ALLOCATABLE,DIMENSION(:) :: value
      INTEGER,ALLOCATABLE,DIMENSION(:) :: Inc
      REAL :: medium
      INTEGER :: NbTot,sgn
      
      allocate(value(0:mpi_size-1))
      allocate(Inc(0:mpi_size-1))
        
  
      CALL allgather_timer_average
      
      medium=0
      DO i=0,mpi_size-1
        value(i)=timer_average(jj_nb_physic(i),timer_physic,i)
    medium=medium+value(i)
      enddo    
      
      medium=medium/mpi_size      
      NbTot=0
      IF (CPPKEY_PHYS) THEN
      DO i=0,mpi_size-1
        Inc(i)=nint(klon_mpi_para_nb(i)*(medium-value(i))/value(i))
        NbTot=NbTot+Inc(i)  
      enddo
      
      IF (NbTot>=0) THEN
        Sgn=1
      else
        Sgn=-1
    NbTot=-NbTot
      endif
      
      Index=0
      DO i=1,NbTot
        Inc(Index)=Inc(Index)-Sgn
    Index=Index+1
    IF (Index>mpi_size-1) Index=0
      enddo
      
      DO i=0,mpi_size-1
        distrib_phys(i)=klon_mpi_para_nb(i)+inc(i)
      enddo
     END IF
         
    END SUBROUTINE  AdjustBands_physic

    SUBROUTINE WriteBands
    USE parallel_lmdz
USE lmdz_dimensions, ONLY: iim, jjm, llm, ndm
    IMPLICIT NONE


      INTEGER :: i,j
      CHARACTER (LEN=4) :: siim,sjjm,sllm,sproc
      CHARACTER (LEN=255) :: filename
      INTEGER :: unit_number=10
      INTEGER :: ierr
  
      WRITE(siim,'(i3)') iim
      WRITE(sjjm,'(i3)') jjm
      WRITE(sllm,'(i3)') llm
      WRITE(sproc,'(i3)') mpi_size

      filename='Bands_'//TRIM(ADJUSTL(siim))//'x'//TRIM(ADJUSTL(sjjm))//'x'//TRIM(ADJUSTL(sllm))//'_'  &
                        //TRIM(ADJUSTL(sproc))//'prc.dat'    
      
      OPEN(UNIT=unit_number,FILE=trim(filename),STATUS='replace',FORM='formatted',IOSTAT=ierr)
      
      IF (ierr==0) THEN
!    write (unit_number,*) '*** Bandes caldyn ***'
    DO i=0,mpi_size-1
          write (unit_number,*) i,jj_nb_caldyn(i)
        enddo
        
!    write (unit_number,*) '*** Bandes vanleer ***'
        DO i=0,mpi_size-1
          write (unit_number,*) i,jj_nb_vanleer(i)
        enddo
       
!        write (unit_number,*) '*** Bandes dissip ***'
        DO i=0,mpi_size-1
          write (unit_number,*) i,jj_nb_dissip(i)
        enddo
        
    DO i=0,mpi_size-1
          write (unit_number,*) i,distrib_phys(i)
        enddo

        CLOSE(unit_number)   
      else 
        print *,'probleme lors de l ecriture des bandes'
      endif
       
    END SUBROUTINE  WriteBands
  
  END MODULE Bands