Homec4science
Diffusion FENNECS: Finite Element Non-Neutral Electron Cloud Simulator (breakdown) Browse

fennecs/src/mpihelper_mod.f90
81171805e064breakdown

Last Change
Raw File

mpihelper_mod.f90
View Options

!------------------------------------------------------------------------------
! EPFL/Swiss Plasma Center
!------------------------------------------------------------------------------
!
! MODULE: mpihelper
!
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!> Module responsible for setting up the MPI variables used in the communications.
!------------------------------------------------------------------------------
MODULE mpihelper
USE constants
use mpi
USE particletypes
IMPLICIT NONE
INTEGER, SAVE :: basicdata_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for communicating basicdata
INTEGER, SAVE :: particle_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for particles communication between nodes
!INTEGER, SAVE :: particles_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for particles gathering on node 0 and broadcast from node 0
INTEGER, SAVE :: rhsoverlap_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for the communication of a rhs column
INTEGER, SAVE :: db_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for the communication of a REAL(kind=db)
INTEGER, SAVE :: momentsoverlap_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for the communication of a column of a grid variable
INTEGER, SAVE :: rcvrhsoverlap_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for the receive communication of a rhs column
INTEGER, SAVE :: rcvmomentsoverlap_type=MPI_DATATYPE_NULL !< Stores the MPI data type used for the receive communication of a column of a grid variable
INTEGER, SAVE:: db_sum_op !< Store the MPI sum operation for db_type
REAL(kind=db), ALLOCATABLE, SAVE:: rhsoverlap_buffer(:) !< buffer used for storing the rhs ghost cells
!< received from the left or right MPI process
REAL(kind=db), ALLOCATABLE, SAVE:: momentsoverlap_buffer(:) !< buffer used for storing the moments ghost cells
!< received from the left or right MPI process
!INTEGER, SAVE:: momentsoverlap_requests(2) = MPI_REQUEST_NULL
!INTEGER, SAVE:: rhsoverlap_requests(2) = MPI_REQUEST_NULL
INTEGER:: rhsoverlap_tag= 200
INTEGER:: momentsoverlap_tag= 300
INTEGER:: partsgather_tag= 500
INTEGER:: partsexchange_tag=600
INTEGER:: nbpartsexchange_tag=700
CONTAINS
!---------------------------------------------------------------------------
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!>
!> @brief
!> Initialize the MPI types used for inter process communications
!
!---------------------------------------------------------------------------
SUBROUTINE mpitypes_init
IMPLICIT NONE
INTEGER:: ierr
! Initialize db_type to use real(kind=db) in MPI and the sum operator for reduce
CALL MPI_TYPE_CREATE_F90_REAL(dprequestedprec,MPI_UNDEFINED,db_type,ierr)
CALL MPI_Type_commit(db_type,ierr)
CALL MPI_Op_Create(DB_sum, .true., db_sum_op, ierr)
CALL init_particlempi
END SUBROUTINE mpitypes_init
!---------------------------------------------------------------------------
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!>
!> @brief
!> Computes the sum in MPI_Reduce operations involving Real(kinc=db)
!
!---------------------------------------------------------------------------
SUBROUTINE DB_sum(INVEC, INOUTVEC, LEN, TYPE)bind(c)
!REAL(kind=db):: INVEC(0:LEN-1), INOUTVEC(0:LEN-1)
use, intrinsic:: iso_c_binding, ONLY: c_ptr,c_f_pointer
use mpi
implicit none
TYPE(C_PTR), VALUE:: invec, inoutvec
real(kind=db),pointer:: ivec(:), iovec(:)
INTEGER:: LEN
INTEGER:: TYPE
INTEGER:: i
call c_f_pointer(INVEC,ivec, (/len/))
call c_f_pointer(inoutvec,iovec, (/len/))
Do i=1,LEN
IOVEC(i)=IVEC(i)+IOVEC(i)
END DO
END SUBROUTINE DB_sum
!--------------------------------------------------------------------------
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!>
!> @brief
!> Initialize the MPI communicators used for allreduce between neighbors
!
!> @param[in] nrank ranks of the FEM array in (1) z direction and (2) r direction
!> @param[in] femorder finite element method order in z and r direction
!> @param[in] zlimleft z index delimiting the mpi local left boundary
!> @param[in] zlimright z index delimiting the mpi local right boundary
!> @param[in] nbmoments number of moments calculated and stored.
!
!---------------------------------------------------------------------------
SUBROUTINE init_overlaps(nrank, femorder, zlimleft, zlimright, nbmoments)
INTEGER, INTENT(IN):: nrank(:), femorder(:), zlimright, zlimleft, nbmoments
IF(ALLOCATED(rhsoverlap_buffer)) DEALLOCATE(rhsoverlap_buffer)
IF(ALLOCATED(momentsoverlap_buffer)) DEALLOCATE(momentsoverlap_buffer)
ALLOCATE(rhsoverlap_buffer(nrank(1)*femorder(2)))
ALLOCATE(momentsoverlap_buffer(nbmoments*nrank(1)*femorder(2)))
! Initialize the MPI column overlap type for rhs
CALL init_coltypempi(nrank(2), zlimright-zlimleft+femorder(2), 1, 1, db_type, rhsoverlap_type)
! Initialize the MPI grid col type
CALL init_coltypempi(nrank(2), zlimright-zlimleft+femorder(2), nbmoments, 1, db_type, momentsoverlap_type)
! Initialize the MPI receive column overlap type for rhs
CALL init_coltypempi(nrank(1), nrank(2), 1, 1, db_type, rcvrhsoverlap_type)
! Initialize the MPI receive grid col type
CALL init_coltypempi(nrank(1), nrank(2), nbmoments, 1, db_type, rcvmomentsoverlap_type)
END SUBROUTINE init_overlaps
SUBROUTINE start_persistentcomm(requests, mpirank, leftproc, rightproc)
INTEGER,INTENT(INOUT):: requests(:)
INTEGER, INTENT(IN):: mpirank, leftproc, rightproc
INTEGER:: ierr
INTEGER:: stats(MPI_STATUS_SIZE,2)
LOGICAL:: completed=.false.
IF(leftproc .lt. mpirank) THEN
! Start to receive
CALL MPI_START(requests(2),ierr)
IF(IERR .ne. MPI_SUCCESS) WRITE(*,*) "error in recv_init"
END IF
IF(rightproc .gt. mpirank) THEN
! Start to send
CALL MPI_START(requests(1),ierr)
IF(IERR .ne. MPI_SUCCESS) WRITE(*,*) "error in send_init"
END IF
IF(leftproc .lt. mpirank) THEN
! Start to receive
completed=.FALSE.
DO WHILE(.not. completed)
CALL MPI_TEST(requests(2), completed,stats(:,2),ierr)
END DO
WRITE(*,*)"status 2", completed, stats(:,2)
!CALL MPI_WAIT(requests(2),stats(:,2),ierr)
!WRITE(*,*)"status 2", stats(:,2)
IF(IERR .ne. MPI_SUCCESS) WRITE(*,*) "error in recv_init"
END IF
IF(rightproc .gt. mpirank) THEN
! Start to send
completed=.FALSE.
DO WHILE(.not. completed)
CALL MPI_TEST(requests(1), completed,stats(:,1),ierr)
END DO
!CALL MPI_WAIT(requests(1),stats(:,1),ierr)
IF(IERR .ne. MPI_SUCCESS) WRITE(*,*) "error in send_init"
END IF
END SUBROUTINE start_persistentcomm
SUBROUTINE rhsoverlapcomm(mpirank, leftproc, rightproc, moments, nrank, femorder, zlimright)
INTEGER, INTENT(IN):: mpirank, leftproc, rightproc
REAL(kind=db), DIMENSION(:), INTENT(INOUT):: moments
INTEGER, INTENT(IN):: nrank(2), femorder(2), zlimright
INTEGER, SAVE:: rhsoverlap_requests(2) = MPI_REQUEST_NULL
INTEGER:: ierr
INTEGER:: stats(MPI_STATUS_SIZE,2)
rhsoverlap_requests=MPI_REQUEST_NULL
rhsoverlap_buffer=0
IF(rightproc .gt. mpirank .and. rightproc .ge. 0) THEN
CALL MPI_ISEND(moments(zlimright*nrank(1)+1), femorder(2)*nrank(1), db_type, rightproc, rhsoverlap_tag, &
& MPI_COMM_WORLD, rhsoverlap_requests(1), ierr )
END IF
! If the processor on the left has actually lower z positions
IF(leftproc .lt. mpirank .and. leftproc .ge. 0) THEN
CALL MPI_IRECV(rhsoverlap_buffer, nrank(1)*(femorder(2)), db_type, leftproc, rhsoverlap_tag, &
& MPI_COMM_WORLD, rhsoverlap_requests(2), ierr )
END IF
CALL MPI_WAITALL(2,rhsoverlap_requests,stats, ierr)
END SUBROUTINE rhsoverlapcomm
SUBROUTINE momentsoverlapcomm(mpirank, leftproc, rightproc, moments, nrank, femorder, zlimright)
INTEGER, INTENT(IN):: mpirank, leftproc, rightproc
REAL(kind=db), DIMENSION(:,:), INTENT(INOUT):: moments
INTEGER, INTENT(IN):: nrank(2), femorder(2), zlimright
INTEGER, SAVE:: momentsoverlap_requests(2) = MPI_REQUEST_NULL
INTEGER:: ierr
INTEGER:: stats(MPI_STATUS_SIZE,2)
momentsoverlap_requests=MPI_REQUEST_NULL
momentsoverlap_buffer=0
IF(rightproc .gt. mpirank .and. rightproc .ge. 0) THEN
CALL MPI_ISEND(moments(1,zlimright*nrank(1)+1), femorder(2)*nrank(1)*10, db_type, rightproc, momentsoverlap_tag, &
& MPI_COMM_WORLD, momentsoverlap_requests(1), ierr )
END IF
! If the processor on the left has actually lower z positions
IF(leftproc .lt. mpirank .and. leftproc .ge. 0) THEN
CALL MPI_IRECV(momentsoverlap_buffer, 10*nrank(1)*(femorder(2)), db_type, leftproc, momentsoverlap_tag, &
& MPI_COMM_WORLD, momentsoverlap_requests(2), ierr )
END IF
CALL MPI_WAITALL(2,momentsoverlap_requests,stats, ierr)
END SUBROUTINE momentsoverlapcomm
!---------------------------------------------------------------------------
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!>
!> @brief
!> Initialize the particle MPI type used for inter process communications and publish it to
!> the process in the communicator
!
!---------------------------------------------------------------------------
SUBROUTINE init_particlempi()
INTEGER :: nblock = 7
INTEGER:: blocklength(7)
INTEGER(kind=MPI_ADDRESS_KIND):: displs(7), displ0
INTEGER:: types(7)
TYPE(particle) :: part
INTEGER:: ierr
CALL mpi_get_address(part%partindex, displs(1), ierr)
CALL mpi_get_address(part%cellindex, displs(2), ierr)
types(1:2)=MPI_INTEGER
CALL mpi_get_address(part%pos, displs(3), ierr)
CALL mpi_get_address(part%U, displs(4), ierr)
CALL mpi_get_address(part%geomweight, displs(5), ierr)
CALL mpi_get_address(part%GAMMA, displs(6), ierr)
CALL mpi_get_address(part%pot, displs(7), ierr)
types(3:7)=db_type
blocklength(1:7) = 1
blocklength(2:5) =3
CALL mpi_get_address(part, displ0, ierr)
displs=displs-displ0
CALL MPI_Type_create_struct(nblock, blocklength, displs, types, particle_type, ierr)
CALL MPI_Type_commit(particle_type,ierr)
END SUBROUTINE init_particlempi
!---------------------------------------------------------------------------
!> @author
!> Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!>
!> @brief
!> Initialize the particles MPI type used for gathering particles to the root and broadcast them and publish it to
!> the process in the communicator
!
!---------------------------------------------------------------------------
SUBROUTINE init_particles_gather_mpi(p,idstart,nsend,mpi_particles_type)
INTEGER:: mpi_particles_type
INTEGER:: nsend
INTEGER:: idstart
INTEGER :: nblock = 8
INTEGER:: blocklength(8)
INTEGER(kind=MPI_ADDRESS_KIND):: displs(8), displ0
INTEGER:: types(8)
TYPE(particles), INTENT(INOUT):: p
INTEGER:: ierr
INTEGER:: temptype
IF(nsend .lt. 1) RETURN
temptype=MPI_DATATYPE_NULL
IF( mpi_particles_type .ne. MPI_DATATYPE_NULL) CALL MPI_TYPE_FREE(mpi_particles_type,ierr)
!CALL mpi_get_address(p%Rindex(idstart), displs(1), ierr)
!CALL mpi_get_address(p%Zindex(idstart), displs(2), ierr)
CALL mpi_get_address(p%cellindex(1,idstart), displs(1), ierr)
CALL mpi_get_address(p%partindex(idstart), displs(2), ierr)
types(1:2)=MPI_INTEGER
CALL mpi_get_address(p%pos(1,idstart), displs(3), ierr)
!CALL mpi_get_address(p%Z(idstart), displs(5), ierr)
!CALL mpi_get_address(p%THET(idstart), displs(6), ierr)
CALL mpi_get_address(p%pot(idstart), displs(4), ierr)
CALL mpi_get_address(p%U(1,idstart), displs(5), ierr)
CALL mpi_get_address(p%Uold(1,idstart), displs(6), ierr)
!CALL mpi_get_address(p%UTHET(idstart), displs(10), ierr)
!CALL mpi_get_address(p%UTHETold(idstart), displs(11), ierr)
!CALL mpi_get_address(p%UZ(idstart), displs(12), ierr)
!CALL mpi_get_address(p%UZold(idstart), displs(13), ierr)
CALL mpi_get_address(p%GAMMA(idstart), displs(7), ierr)
CALL mpi_get_address(p%GAMMAold(idstart), displs(8), ierr)
types(3:8)=db_type
blocklength = nsend
blocklength(1)=3*nsend
blocklength(3)=3*nsend
blocklength(5:6)=3*nsend
CALL mpi_get_address(p, displ0, ierr)
displs=displs-displ0
CALL MPI_Type_create_struct(nblock, blocklength, displs, types, mpi_particles_type, ierr)
CALL MPI_TYPE_COMMIT(mpi_particles_type, ierr)
END SUBROUTINE init_particles_gather_mpi
!---------------------------------------------------------------------------
!> @author Guillaume Le Bars EPFL/SPC
!
! DESCRIPTION:
!> @brief Initialize the column MPI type used for inter process communications and publish it to
!> the processes in the communicator (can be rhs or grid quantities)
!
!> @param[in] nr number of elements in the r direction
!> @param[in] nz number of elements in the z direction
!> @param[in] init_type MPI type of the initial data
!> @param[inout] mpi_coltype final type usable in communications
!---------------------------------------------------------------------------
SUBROUTINE init_coltypempi(nr, nz, block_size, stride, init_type, mpi_coltype)
INTEGER, INTENT(IN) :: nr
INTEGER, INTENT(IN) :: nz
INTEGER, INTENT(IN) :: block_size
INTEGER, INTENT(IN) :: stride
INTEGER, INTENT(IN) :: init_type
INTEGER, INTENT(OUT) :: mpi_coltype
INTEGER :: temp_mpi_coltype
INTEGER:: ierr
INTEGER(KIND=MPI_ADDRESS_KIND):: init_type_lb, init_type_extent
!(nrank(2), nrank(1), 1, 10, db_type, rhsoverlap_type)
! if mpi_coltype was used, we free it first
IF( mpi_coltype .ne. MPI_DATATYPE_NULL) CALL MPI_TYPE_FREE(mpi_coltype,ierr)
! Create vector type of length nx
CALL MPI_TYPE_VECTOR(nr, block_size, stride*block_size*nz, init_type, temp_mpi_coltype, ierr)
CALL MPI_TYPE_COMMIT(temp_mpi_coltype, ierr)
! Get the size in bytes of the initial type
CALL MPI_TYPE_GET_EXTENT(init_type, init_type_lb, init_type_extent, ierr)
if(mpi_coltype .ne. MPI_DATATYPE_NULL) CALL MPI_TYPE_FREE(mpi_coltype,ierr)
! Resize temp_mpi_coltype such that the next item to read is at j+1
CALL MPI_TYPE_CREATE_RESIZED(temp_mpi_coltype, init_type_lb, stride*block_size*init_type_extent ,&
& mpi_coltype, ierr)
CALL MPI_TYPE_COMMIT(mpi_coltype, ierr)
CALL MPI_TYPE_FREE(temp_mpi_coltype,ierr)
END SUBROUTINE init_coltypempi
END MODULE mpihelper
c4science · Help