mirror of https://gitlab.com/QEF/q-e.git
Parallelization choice made more fool-proof
This commit is contained in:
Paolo Giannozzi
parent
7d7a129d01
commit
125077923f
|
|
@ -58,8 +58,8 @@ SUBROUTINE set_para_diag( nbnd, use_para_diag )
|
|||
!
|
||||
IF ( ionode ) THEN
|
||||
!
|
||||
WRITE( stdout, '(/,5X,"Subspace diagonalization in iterative solution ",&
|
||||
& "of the eigenvalue problem:")' )
|
||||
WRITE( stdout, '(5X,"Subspace diagonalization in iterative solution ",&
|
||||
& "of the eigenvalue problem:")' )
|
||||
IF ( use_para_diag ) THEN
|
||||
IF (ortho_parent_comm == intra_pool_comm) THEN
|
||||
WRITE( stdout, '(5X,"one sub-group per k-point group (pool) will be used")' )
|
||||
|
|
|
|||
|
|
@ -97,18 +97,12 @@ SUBROUTINE setup()
|
|||
!
|
||||
IMPLICIT NONE
|
||||
!
|
||||
INTEGER :: na, is, ierr, ibnd, ik, nrot_, nr3
|
||||
INTEGER :: na, is, ierr, ibnd, ik, nrot_, nr3, nk_
|
||||
LOGICAL :: magnetic_sym, skip_equivalence=.FALSE.
|
||||
REAL(DP) :: iocc, ionic_charge, one
|
||||
!
|
||||
TYPE(output_type) :: output_obj
|
||||
!
|
||||
#if defined(__MPI)
|
||||
LOGICAL :: lpara = .true.
|
||||
#else
|
||||
LOGICAL :: lpara = .false.
|
||||
#endif
|
||||
!
|
||||
! ... okvan/okpaw = .TRUE. : at least one pseudopotential is US/PAW
|
||||
!
|
||||
okvan = ANY( upf(1:ntyp)%tvanp )
|
||||
|
|
@ -653,7 +647,14 @@ SUBROUTINE setup()
|
|||
!
|
||||
nr3 = int ( sqrt(gcutms)*sqrt (at(1, 3)**2 + at(2, 3)**2 + at(3, 3)**2) ) + 1
|
||||
nr3 = good_fft_order( 2*nr3, fft_fact(3) )
|
||||
CALL setup_para ( nr3, nkstot, nbnd )
|
||||
!
|
||||
! ... nk_ = number of k-points usable for k-point parallelization
|
||||
! (set to 1 if k-point parallelization is not implemented)
|
||||
nk_ = nkstot
|
||||
IF ( lberry .OR. lelfield .OR. lorbm .OR. &
|
||||
( xclib_dft_is('hybrid') .AND. tstress ) ) nk_ = 1
|
||||
!
|
||||
CALL setup_para ( nr3, nk_, nbnd )
|
||||
!
|
||||
! ... distribute k-points across processors of a pool
|
||||
!
|
||||
|
|
@ -684,10 +685,12 @@ END SUBROUTINE setup
|
|||
SUBROUTINE setup_para ( nr3, nkstot, nbnd )
|
||||
!----------------------------------------------------------------------------
|
||||
!
|
||||
! Initialize the various parallelization levels
|
||||
! Must be called after setup and before setup_exx only once
|
||||
! Initialize the various parallelization levels, trying to guess decent
|
||||
! parameters for npool, ndiag, ntg, if not specified in the command line
|
||||
! Must be called at the end of "setup" but before "setup_exx",
|
||||
! only once per run
|
||||
!
|
||||
USE control_flags, ONLY : use_para_diag, use_gpu
|
||||
USE control_flags, ONLY : use_para_diag, use_gpu
|
||||
USE io_global, ONLY : stdout
|
||||
USE mp_bands, ONLY : mp_start_bands, nbgrp, ntask_groups, nproc_bgrp,&
|
||||
nyfft
|
||||
|
|
@ -733,9 +736,10 @@ pool: do np = 2, nkstot
|
|||
END IF
|
||||
CALL mp_start_pools ( npool_, intra_image_comm )
|
||||
!
|
||||
! band parallelization - FIXME: why the following section?
|
||||
! band parallelization (nband only; ntg and nyfft below are just copied)
|
||||
!
|
||||
#if defined (__CUDA_OPTIMIZED)
|
||||
! band parallelization is effectively disabled in this case
|
||||
CALL mp_start_bands ( 1 , ntg_, nyfft_, intra_pool_comm )
|
||||
#else
|
||||
CALL mp_start_bands ( nband_, ntg_, nyfft_, intra_pool_comm )
|
||||
|
|
@ -745,13 +749,15 @@ pool: do np = 2, nkstot
|
|||
!
|
||||
use_gpu = check_gpu_support( )
|
||||
!
|
||||
! Set "task_groups" if still too many processors for PW parallelization
|
||||
! Set "task groups", max 16, if too many processors for PW parallelization
|
||||
!
|
||||
IF ( ntask_groups == 0 ) THEN
|
||||
ntask_groups = 1
|
||||
if ( nproc_bgrp > nr3 ) THEN
|
||||
maxtask = min (nbnd, 16)
|
||||
task: do np = 2, maxtask
|
||||
! ensure that ntask_group, that coincides with nyfft,
|
||||
! is commensurate with the number of procs for PW parallelization
|
||||
if ( mod(nproc_bgrp,np) /= 0 ) cycle
|
||||
if ( nproc_bgrp/np < nr3/4 ) then
|
||||
ntask_groups = np
|
||||
|
|
@ -762,17 +768,21 @@ task: do np = 2, maxtask
|
|||
END IF
|
||||
!
|
||||
! Note that "task_groups" require to set pencil_decomposition to .true.
|
||||
! Same if there are more processors than xy planes in the FFT
|
||||
!
|
||||
IF ( ntask_groups /= 1 ) pencil_decomposition_ = .true.
|
||||
IF ( ntask_groups /= 1 .OR. nproc_bgrp > nr3 ) pencil_decomposition_ = .true.
|
||||
!
|
||||
! printout - same as in environment.f90
|
||||
!
|
||||
WRITE( stdout, * )
|
||||
IF ( npool > 1 ) WRITE( stdout, &
|
||||
'(5X,"K-points division: npool = ",I7)' ) npool
|
||||
IF ( nbgrp > 1 ) WRITE( stdout, &
|
||||
'(5X,"band groups division: nbgrp = ",I7)' ) nbgrp
|
||||
IF ( nproc_bgrp > 1 ) WRITE( stdout, &
|
||||
'(5X,"R & G space division: proc/nbgrp/npool/nimage = ",I7)' ) nproc_bgrp
|
||||
IF ( nproc_bgrp > nr3 ) WRITE( stdout, &
|
||||
'(5X,"WARNING: too many processors for an effective parallelization!")' )
|
||||
IF ( nyfft > 1 ) WRITE( stdout, &
|
||||
'(5X,"wavefunctions fft division: Y-proc x Z-proc = ",2I7)' ) &
|
||||
nyfft, nproc_bgrp / nyfft
|
||||
|
|
@ -781,11 +791,16 @@ task: do np = 2, maxtask
|
|||
ntask_groups, nproc_bgrp / ntask_groups
|
||||
IF ( nmany_ > 1) WRITE( stdout, '(5X,"FFT bands division: nmany = ",I7)' ) nmany_
|
||||
!
|
||||
! linear algebra - for GPUs, ndiag_=1; otherwise, a value ensuring that
|
||||
! matrices nbnd*nbnd are distributed into blocks of size > 100x100 or so
|
||||
! linear algebra - for GPUs, ndiag = 1; otherwise, a value ensuring that
|
||||
! matrices nbnd*nbnd are distributed into blocks of size > 100x100
|
||||
!
|
||||
if ( use_gpu ) ndiag_ = 1
|
||||
if ( ndiag_ == 0 ) ndiag_ = max(1,nint(nbnd/100.)**2)
|
||||
if ( ndiag_ == 0 .AND. use_gpu ) ndiag_ = 1
|
||||
if ( ndiag_ == 0 ) then
|
||||
do np = nint(nbnd/100.), 1, -1
|
||||
if ( np**2 <= nproc_bgrp ) exit
|
||||
end do
|
||||
ndiag_ = max(1,np**2)
|
||||
end if
|
||||
!
|
||||
CALL set_para_diag( nbnd, use_para_diag )
|
||||
!
|
||||
|
|
|
|||
Loading…
Reference in New Issue