module times
  INTEGER,PRIVATE,save :: Last_Count=0
  REAL, PRIVATE,save :: Last_cpuCount=0
  logical, PRIVATE,save :: AllTimer_IsActive=.FALSE.
  
  INTEGER, parameter :: nb_timer = 4
  INTEGER, parameter :: timer_caldyn  = 1
  INTEGER, parameter :: timer_vanleer = 2
  INTEGER, parameter :: timer_dissip = 3
  INTEGER, parameter :: timer_physic = 4
  INTEGER, parameter :: stopped = 1
  INTEGER, parameter :: running = 2
  INTEGER, parameter :: suspended = 3
  
  INTEGER :: max_size
  REAL,    ALLOCATABLE, DIMENSION(:,:,:) :: timer_table
  REAL,    ALLOCATABLE, DIMENSION(:,:,:) :: timer_table_sqr
  INTEGER, ALLOCATABLE, DIMENSION(:,:,:) :: timer_iteration
  REAL,    ALLOCATABLE, DIMENSION(:,:,:) :: timer_average
  REAL,    ALLOCATABLE, DIMENSION(:,:,:) :: timer_delta
  REAL,    ALLOCATABLE,DIMENSION(:) :: timer_running, last_time
  INTEGER, ALLOCATABLE,DIMENSION(:) :: timer_state
  
  CONTAINS
  
  SUBROUTINE init_timer
    USE parallel_lmdz
    IMPLICIT NONE
    INCLUDE "dimensions.h"
    INCLUDE "paramet.h"
    
    max_size=jjm+1
    allocate(timer_table(max_size,nb_timer,0:mpi_size-1))
    allocate(timer_table_sqr(max_size,nb_timer,0:mpi_size-1))
    allocate(timer_iteration(max_size,nb_timer,0:mpi_size-1))
    allocate(timer_average(max_size,nb_timer,0:mpi_size-1))
    allocate(timer_delta(max_size,nb_timer,0:mpi_size-1))
    allocate(timer_running(nb_timer))
    allocate(timer_state(nb_timer))
    allocate(last_time(nb_timer))
    
    timer_table(:,:,:)=0
    timer_table_sqr(:,:,:)=0
    timer_iteration(:,:,:)=0
    timer_average(:,:,:)=0
    timer_delta(:,:,:)=0
    timer_state(:)=stopped      
  END SUBROUTINE  init_timer
  
  SUBROUTINE start_timer(no_timer)


    IMPLICIT NONE
    INTEGER :: no_timer
    
    IF (AllTimer_IsActive) THEN
      IF (timer_state(no_timer)/=stopped) THEN
        CALL abort_gcm("times","start_timer :: timer is already running or suspended",1)
      else
        timer_state(no_timer)=running
      endif
      
      timer_running(no_timer)=0
      CALL cpu_time(last_time(no_timer))
    
    endif
    
  END SUBROUTINE  start_timer
  
  SUBROUTINE suspend_timer(no_timer)
    IMPLICIT NONE
    INTEGER :: no_timer
     
    IF (AllTimer_IsActive) THEN
      IF (timer_state(no_timer)/=running) THEN
         CALL abort_gcm("times","suspend_timer :: timer is not running",1)
      else
        timer_state(no_timer)=suspended
      endif
    
      timer_running(no_timer)=timer_running(no_timer)-last_time(no_timer)
      CALL cpu_time(last_time(no_timer))
      timer_running(no_timer)=timer_running(no_timer)+last_time(no_timer)
    endif
  END SUBROUTINE  suspend_timer
  
  SUBROUTINE resume_timer(no_timer)
    IMPLICIT NONE
    INTEGER :: no_timer
     
    IF (AllTimer_IsActive) THEN
      IF (timer_state(no_timer)/=suspended) THEN
        CALL abort_gcm("times","resume_timer :: timer is not suspended",1)
      else
        timer_state(no_timer)=running
      endif
      
      CALL cpu_time(last_time(no_timer))
    endif
    
  END SUBROUTINE  resume_timer

  SUBROUTINE stop_timer(no_timer)
    USE parallel_lmdz
    IMPLICIT NONE
    INTEGER :: no_timer
    INTEGER :: N
    REAL :: V,V2
    
    IF (AllTimer_IsActive) THEN
      IF (timer_state(no_timer)/=running) THEN
        CALL abort_gcm("times","stop_timer :: timer is not running",1)
      else
        timer_state(no_timer)=stopped
      endif
    
      timer_running(no_timer)=timer_running(no_timer)-last_time(no_timer)
      CALL cpu_time(last_time(no_timer))
      timer_running(no_timer)=timer_running(no_timer)+last_time(no_timer)
    
      timer_table(jj_nb,no_timer,mpi_rank)=timer_table(jj_nb,no_timer,mpi_rank)+timer_running(no_timer)
      timer_table_sqr(jj_nb,no_timer,mpi_rank)=timer_table_sqr(jj_nb,no_timer,mpi_rank)+timer_running(no_timer)**2
      timer_iteration(jj_nb,no_timer,mpi_rank)=timer_iteration(jj_nb,no_timer,mpi_rank)+1
      timer_average(jj_nb,no_timer,mpi_rank)=timer_table(jj_nb,no_timer,mpi_rank)/timer_iteration(jj_nb,no_timer,mpi_rank)
      IF (timer_iteration(jj_nb,no_timer,mpi_rank)>=2) THEN
        N=timer_iteration(jj_nb,no_timer,mpi_rank)
	V2=timer_table_sqr(jj_nb,no_timer,mpi_rank)
	V=timer_table(jj_nb,no_timer,mpi_rank)
	timer_delta(jj_nb,no_timer,mpi_rank)=sqrt(ABS(V2-V*V/N)/(N-1)) 
      else
        timer_delta(jj_nb,no_timer,mpi_rank)=0
      endif
    endif
    
  END SUBROUTINE  stop_timer
   
  SUBROUTINE allgather_timer
    USE parallel_lmdz
    USE lmdz_mpi
    IMPLICIT NONE

    INTEGER :: ierr
    INTEGER :: data_size
    REAL, ALLOCATABLE,DIMENSION(:,:) :: tmp_table

    IF (using_mpi) THEN    
   
      IF (AllTimer_IsActive) THEN
      allocate(tmp_table(max_size,nb_timer))
    
      data_size=max_size*nb_timer
    
      tmp_table(:,:)=timer_table(:,:,mpi_rank)
      CALL mpi_allgather(tmp_table(1,1),data_size,MPI_REAL_LMDZ,timer_table(1,1,0),data_size,MPI_REAL_LMDZ,COMM_LMDZ,ierr)
      tmp_table(:,:)=timer_table_sqr(:,:,mpi_rank)
      CALL mpi_allgather(tmp_table(1,1),data_size,MPI_REAL_LMDZ,timer_table_sqr(1,1,0),data_size,MPI_REAL_LMDZ,COMM_LMDZ,ierr)
      deallocate(tmp_table)
    
      endif
      
    ENDIF ! using_mpi
    
  END SUBROUTINE  allgather_timer
  
  SUBROUTINE allgather_timer_average
    USE parallel_lmdz
    USE lmdz_mpi
    IMPLICIT NONE
    INTEGER :: ierr
    INTEGER :: data_size
    REAL, ALLOCATABLE,DIMENSION(:,:),target :: tmp_table
    INTEGER, ALLOCATABLE,DIMENSION(:,:),target :: tmp_iter
    INTEGER :: istats

    IF (using_mpi) THEN
        
      IF (AllTimer_IsActive) THEN
      allocate(tmp_table(max_size,nb_timer))
      allocate(tmp_iter(max_size,nb_timer))
   
      data_size=max_size*nb_timer

      tmp_table(:,:)=timer_average(:,:,mpi_rank)
      CALL mpi_allgather(tmp_table(1,1),data_size,MPI_REAL_LMDZ,timer_average(1,1,0),data_size,MPI_REAL_LMDZ,COMM_LMDZ,ierr)
      tmp_table(:,:)=timer_delta(:,:,mpi_rank)
      CALL mpi_allgather(tmp_table(1,1),data_size,MPI_REAL_LMDZ,timer_delta(1,1,0),data_size,MPI_REAL_LMDZ,COMM_LMDZ,ierr)
      tmp_iter(:,:)=timer_iteration(:,:,mpi_rank)
      CALL mpi_allgather(tmp_iter(1,1),data_size,MPI_INTEGER,timer_iteration(1,1,0),data_size,MPI_INTEGER,COMM_LMDZ,ierr)
      deallocate(tmp_table)
    
      endif
      
    ENDIF  ! using_mpi
  END SUBROUTINE  allgather_timer_average
  
  SUBROUTINE InitTime
  IMPLICIT NONE
    INTEGER :: count,count_rate,count_max
    
    AllTimer_IsActive=.TRUE.
    IF (AllTimer_IsActive) THEN
      CALL system_clock(count,count_rate,count_max)
      CALL cpu_time(Last_cpuCount)
      Last_Count=count
    endif
  END SUBROUTINE  InitTime
  
  function DiffTime()
  IMPLICIT NONE
    double precision :: DiffTime
    INTEGER :: count,count_rate,count_max
  
    CALL system_clock(count,count_rate,count_max)
    IF (Count>=Last_Count) THEN
      DiffTime=(1.*(Count-last_Count))/count_rate
    else
      DiffTime=(1.*(Count-last_Count+Count_max))/count_rate
    endif
    Last_Count=Count 
  END FUNCTION DiffTime
  
  function DiffCpuTime()
  IMPLICIT NONE
    REAL :: DiffCpuTime
    REAL :: Count
    
    CALL cpu_time(Count)
    DiffCpuTime=Count-Last_cpuCount
    Last_cpuCount=Count 
  END FUNCTION DiffCpuTime

END MODULE times