quantum-espresso/Modules/mp_world.f90

137 lines
4.1 KiB
Fortran

!
! Copyright (C) 2001-2015 Quantum ESPRESSO group
! This file is distributed under the terms of the
! GNU General Public License. See the file `License'
! in the root directory of the present distribution,
! or http://www.gnu.org/copyleft/gpl.txt .
!
!----------------------------------------------------------------------------
MODULE mp_world
!----------------------------------------------------------------------------
!! World group - all QE routines using \(\texttt{mp_world_start}\) to
!! start MPI will work in the communicator passed as input to
!! \(\texttt{mp_world_start}\)
!
USE mp, ONLY : mp_barrier, mp_start, mp_end, mp_stop, mp_count_nodes
USE io_global, ONLY : meta_ionode_id, meta_ionode
#if defined(__CUDA)
use cudafor, ONLY : cudaSetDevice, cudaGetDeviceCount, cudaDeviceSynchronize
#endif
!
USE parallel_include
!
IMPLICIT NONE
SAVE
!
! ... World group - all QE routines using mp_world_start to start MPI
! ... will work in the communicator passed as input to mp_world_start
!
INTEGER :: nnode = 1
!! number of nodes
INTEGER :: nproc = 1
!! number of processors
INTEGER :: mpime = 0
!! processor index (starts from 0 to nproc-1)
INTEGER :: root = 0
!! index of the root processor
INTEGER :: world_comm = 0
!! communicator
!
#if defined(__MPI)
LOGICAL :: library_mode = .FALSE.
!! library\_mode = TRUE if QE is called as a library by an external code.
!! If true:
!! - \(\texttt{MPI_Init()}\) is not called when starting MPI;
!! - \(\texttt{MPI_Finalize()}\) is not called when stopping MPI.
#endif
!
PRIVATE
PUBLIC :: nnode, nproc, mpime, root, world_comm, mp_world_start, mp_world_end
!
CONTAINS
!
!-----------------------------------------------------------------------
SUBROUTINE mp_world_start ( my_world_comm )
!-----------------------------------------------------------------------
!
IMPLICIT NONE
INTEGER, INTENT(IN) :: my_world_comm
INTEGER :: color, key, ndev
#if defined(__MPI) || defined(__CUDA)
INTEGER :: ierr
#endif
#if defined(_OPENMP)
INTEGER :: PROVIDED
#endif
!
world_comm = my_world_comm
!
! ... check if mpi is already initialized (library mode) or not
!
#if defined(__MPI)
CALL MPI_Initialized ( library_mode, ierr)
IF (ierr/=0) CALL mp_stop( 8000 )
IF (.NOT. library_mode ) THEN
#if defined(_OPENMP)
CALL MPI_Init_thread(MPI_THREAD_MULTIPLE, PROVIDED, ierr)
#else
CALL MPI_Init(ierr)
#endif
IF (ierr/=0) CALL mp_stop( 8001 )
END IF
#endif
!
CALL mp_count_nodes ( nnode, color, key, world_comm )
!
#if defined(__CUDA)
ierr = cudaGetDeviceCount( ndev )
IF (ierr/=0) CALL mp_stop( 9000 + ierr )
!
! WARNING: the OpenMP standard does not guarantee that the thread
! pool remains the same in different parallel regions.
! However, this is apparently what all major implementations do.
!$omp parallel firstprivate(key, ndev) reduction(max:ierr)
ierr = cudaSetDevice(mod(key, ndev))
!$omp end parallel
IF (ierr/=0) CALL mp_stop( 9100 + ierr )
ierr = cudaDeviceSynchronize()
IF (ierr/=0) CALL mp_stop( 9200 + ierr )
#if defined(__DEBUG)
write(*,*) "MPI ", key, " on node ", color, " is using GPU: ", mod(key, ndev)
#endif
#endif
!
CALL mp_start( nproc, mpime, world_comm )
!
!
! ... meta_ionode is true if this processor is the root processor
! ... of the world group - "ionode_world" would be a better name
! ... meta_ionode_id is the index of such processor
!
meta_ionode = ( mpime == root )
meta_ionode_id = root
!
RETURN
!
END SUBROUTINE mp_world_start
!
!-----------------------------------------------------------------------
SUBROUTINE mp_world_end ( )
!-----------------------------------------------------------------------
#if defined(__MPI)
INTEGER :: ierr
#endif
!
CALL mp_barrier( world_comm )
CALL mp_end ( world_comm )
#if defined(__MPI)
IF (.NOT. library_mode ) THEN
CALL mpi_finalize(ierr)
IF (ierr/=0) CALL mp_stop( 8002 )
END IF
#endif
!
END SUBROUTINE mp_world_end
!
END MODULE mp_world