9
1
mirror of https://github.com/QuantumPackage/qp2.git synced 2024-11-18 11:23:38 +01:00
qp2/src/davidson/davidson_parallel_nos2.irp.f

497 lines
14 KiB
Fortran

use bitmasks
use f77_zmq
subroutine davidson_nos2_slave_inproc(i)
implicit none
integer, intent(in) :: i
call davidson_nos2_run_slave(1,i)
end
subroutine davidson_nos2_slave_tcp(i)
implicit none
integer, intent(in) :: i
call davidson_nos2_run_slave(0,i)
end
subroutine davidson_nos2_run_slave(thread,iproc)
use f77_zmq
implicit none
BEGIN_DOC
! Slave routine for Davidson's diagonalization.
END_DOC
integer, intent(in) :: thread, iproc
integer :: worker_id, task_id, blockb
integer(ZMQ_PTR),external :: new_zmq_to_qp_run_socket
integer(ZMQ_PTR) :: zmq_to_qp_run_socket
integer(ZMQ_PTR), external :: new_zmq_push_socket
integer(ZMQ_PTR) :: zmq_socket_push
integer, external :: connect_to_taskserver
integer :: doexit, send, receive
zmq_to_qp_run_socket = new_zmq_to_qp_run_socket()
doexit = 0
if (connect_to_taskserver(zmq_to_qp_run_socket,worker_id,thread) == -1) then
doexit=1
endif
IRP_IF MPI
include 'mpif.h'
integer :: ierr
send = doexit
call MPI_AllReduce(send, receive, 1, MPI_INTEGER, MPI_SUM, MPI_COMM_WORLD, ierr)
if (ierr /= MPI_SUCCESS) then
doexit=1
endif
doexit = receive
IRP_ENDIF
if (doexit>0) then
call end_zmq_to_qp_run_socket(zmq_to_qp_run_socket)
return
endif
zmq_socket_push = new_zmq_push_socket(thread)
call davidson_nos2_slave_work(zmq_to_qp_run_socket, zmq_socket_push, N_states_diag, N_det, worker_id)
integer, external :: disconnect_from_taskserver
if (disconnect_from_taskserver(zmq_to_qp_run_socket,worker_id) == -1) then
call sleep(1)
if (disconnect_from_taskserver(zmq_to_qp_run_socket,worker_id) == -1) then
print *, irp_here, ': disconnect failed'
continue
endif
endif
call end_zmq_to_qp_run_socket(zmq_to_qp_run_socket)
call end_zmq_push_socket(zmq_socket_push)
end subroutine
subroutine davidson_nos2_slave_work(zmq_to_qp_run_socket, zmq_socket_push, N_st, sze, worker_id)
use f77_zmq
implicit none
integer(ZMQ_PTR),intent(in) :: zmq_to_qp_run_socket
integer(ZMQ_PTR),intent(in) :: zmq_socket_push
integer,intent(in) :: worker_id, N_st, sze
integer :: task_id
character*(512) :: msg
integer :: imin, imax, ishift, istep
integer, allocatable :: psi_det_read(:,:,:)
double precision, allocatable :: v_t(:,:), u_t(:,:)
!DIR$ ATTRIBUTES ALIGN : $IRP_ALIGN :: u_t, v_t
! Get wave function (u_t)
! -----------------------
integer :: rc, ni, nj
integer*8 :: rc8
integer :: N_states_read, N_det_read, psi_det_size_read
integer :: N_det_selectors_read, N_det_generators_read
integer, external :: zmq_get_dvector
integer, external :: zmq_get_dmatrix
PROVIDE psi_det_beta_unique psi_bilinear_matrix_order_transp_reverse psi_det_alpha_unique
PROVIDE psi_bilinear_matrix_transp_values psi_bilinear_matrix_values psi_bilinear_matrix_columns_loc
PROVIDE ref_bitmask_energy nproc
PROVIDE mpi_initialized
allocate(u_t(N_st,N_det))
! Warning : dimensions are modified for efficiency, It is OK since we get the
! full matrix
if (size(u_t,kind=8) < 8388608_8) then
ni = size(u_t)
nj = 1
else
ni = 8388608
nj = int(size(u_t,kind=8)/8388608_8,4) + 1
endif
do while (zmq_get_dmatrix(zmq_to_qp_run_socket, worker_id, 'u_t', u_t, ni, nj, size(u_t,kind=8)) == -1)
print *, 'mpi_rank, N_states_diag, N_det'
print *, mpi_rank, N_states_diag, N_det
stop 'u_t'
enddo
IRP_IF MPI
include 'mpif.h'
integer :: ierr
call broadcast_chunks_double(u_t,size(u_t,kind=8))
IRP_ENDIF
! Run tasks
! ---------
logical :: sending
sending=.False.
allocate(v_t(N_st,N_det))
do
integer, external :: get_task_from_taskserver
integer, external :: task_done_to_taskserver
if (get_task_from_taskserver(zmq_to_qp_run_socket,worker_id, task_id, msg) == -1) then
exit
endif
if(task_id == 0) exit
read (msg,*) imin, imax, ishift, istep
integer :: k
do k=imin,imax
v_t(:,k) = 0.d0
enddo
call H_u_0_nstates_openmp_work(v_t,u_t,N_st,N_det,imin,imax,ishift,istep)
if (task_done_to_taskserver(zmq_to_qp_run_socket,worker_id,task_id) == -1) then
print *, irp_here, 'Unable to send task_done'
endif
call davidson_push_results_async_recv(zmq_socket_push, sending)
call davidson_nos2_push_results_async_send(zmq_socket_push, v_t, imin, imax, task_id, sending)
end do
deallocate(u_t,v_t)
call davidson_push_results_async_recv(zmq_socket_push, sending)
end subroutine
subroutine davidson_nos2_push_results(zmq_socket_push, v_t, imin, imax, task_id)
use f77_zmq
implicit none
BEGIN_DOC
! Push the results of $H | U \rangle$ from a worker to the master.
END_DOC
integer(ZMQ_PTR) ,intent(in) :: zmq_socket_push
integer ,intent(in) :: task_id, imin, imax
double precision ,intent(in) :: v_t(N_states_diag,N_det)
integer :: rc, sz
integer*8 :: rc8
sz = (imax-imin+1)*N_states_diag
rc = f77_zmq_send( zmq_socket_push, task_id, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push task_id'
rc = f77_zmq_send( zmq_socket_push, imin, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push imin'
rc = f77_zmq_send( zmq_socket_push, imax, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push imax'
rc8 = f77_zmq_send8( zmq_socket_push, v_t(1,imin), 8_8*sz, 0)
if(rc8 /= 8_8*sz) stop 'davidson_nos2_push_results failed to push vt'
! Activate is zmq_socket_push is a REQ
IRP_IF ZMQ_PUSH
IRP_ELSE
character*(2) :: ok
rc = f77_zmq_recv( zmq_socket_push, ok, 2, 0)
if ((rc /= 2).and.(ok(1:2)/='ok')) then
print *, irp_here, ': f77_zmq_recv( zmq_socket_push, ok, 2, 0)'
stop -1
endif
IRP_ENDIF
end subroutine
subroutine davidson_nos2_push_results_async_send(zmq_socket_push, v_t, imin, imax, task_id,sending)
use f77_zmq
implicit none
BEGIN_DOC
! Push the results of $H | U \rangle$ from a worker to the master.
END_DOC
integer(ZMQ_PTR) ,intent(in) :: zmq_socket_push
integer ,intent(in) :: task_id, imin, imax
double precision ,intent(in) :: v_t(N_states_diag,N_det)
logical ,intent(inout) :: sending
integer :: rc, sz
integer*8 :: rc8
if (sending) then
print *, irp_here, ': sending=true'
stop -1
endif
sending = .True.
sz = (imax-imin+1)*N_states_diag
rc = f77_zmq_send( zmq_socket_push, task_id, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push task_id'
rc = f77_zmq_send( zmq_socket_push, imin, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push imin'
rc = f77_zmq_send( zmq_socket_push, imax, 4, ZMQ_SNDMORE)
if(rc /= 4) stop 'davidson_nos2_push_results failed to push imax'
rc8 = f77_zmq_send8( zmq_socket_push, v_t(1,imin), 8_8*sz, 0)
if(rc8 /= 8_8*sz) stop 'davidson_nos2_push_results failed to push vt'
end subroutine
subroutine davidson_nos2_pull_results(zmq_socket_pull, v_t, imin, imax, task_id)
use f77_zmq
implicit none
BEGIN_DOC
! Pull the results of $H | U \rangle$ on the master.
END_DOC
integer(ZMQ_PTR) ,intent(in) :: zmq_socket_pull
integer ,intent(out) :: task_id, imin, imax
double precision ,intent(out) :: v_t(N_states_diag,N_det)
integer :: rc, sz
integer*8 :: rc8
rc = f77_zmq_recv( zmq_socket_pull, task_id, 4, 0)
if(rc /= 4) stop 'davidson_nos2_pull_results failed to pull task_id'
rc = f77_zmq_recv( zmq_socket_pull, imin, 4, 0)
if(rc /= 4) stop 'davidson_nos2_pull_results failed to pull imin'
rc = f77_zmq_recv( zmq_socket_pull, imax, 4, 0)
if(rc /= 4) stop 'davidson_nos2_pull_results failed to pull imax'
sz = (imax-imin+1)*N_states_diag
rc8 = f77_zmq_recv8( zmq_socket_pull, v_t(1,imin), 8_8*sz, 0)
if(rc8 /= 8*sz) stop 'davidson_nos2_pull_results failed to pull v_t'
! Activate if zmq_socket_pull is a REP
IRP_IF ZMQ_PUSH
IRP_ELSE
rc = f77_zmq_send( zmq_socket_pull, 'ok', 2, 0)
if (rc /= 2) then
print *, irp_here, ' : f77_zmq_send (zmq_socket_pull,...'
stop -1
endif
IRP_ENDIF
end subroutine
subroutine davidson_nos2_collector(zmq_to_qp_run_socket, zmq_socket_pull, v0, sze, N_st)
use f77_zmq
implicit none
BEGIN_DOC
! Routine collecting the results of the workers in Davidson's algorithm.
END_DOC
integer(ZMQ_PTR), intent(in) :: zmq_socket_pull
integer, intent(in) :: sze, N_st
integer(ZMQ_PTR), intent(in) :: zmq_to_qp_run_socket
double precision ,intent(inout) :: v0(sze, N_st)
integer :: more, task_id, imin, imax
double precision, allocatable :: v_t(:,:)
logical :: sending
integer :: i,j
integer, external :: zmq_delete_task_async_send
integer, external :: zmq_delete_task_async_recv
allocate(v_t(N_st,N_det))
v0 = 0.d0
more = 1
sending = .False.
do while (more == 1)
call davidson_nos2_pull_results(zmq_socket_pull, v_t, imin, imax, task_id)
if (zmq_delete_task_async_send(zmq_to_qp_run_socket,task_id,sending) == -1) then
stop 'davidson: Unable to delete task (send)'
endif
do j=1,N_st
do i=imin,imax
v0(i,j) = v0(i,j) + v_t(j,i)
enddo
enddo
if (zmq_delete_task_async_recv(zmq_to_qp_run_socket,more,sending) == -1) then
stop 'davidson: Unable to delete task (recv)'
endif
end do
deallocate(v_t)
end subroutine
subroutine H_u_0_nstates_zmq(v_0,u_0,N_st,sze)
use omp_lib
use bitmasks
use f77_zmq
implicit none
BEGIN_DOC
! Computes $v_0 = H | u_0\rangle$
!
! n : number of determinants
!
! H_jj : array of $\langle j | H | j \rangle$
END_DOC
integer, intent(in) :: N_st, sze
double precision, intent(out) :: v_0(sze,N_st)
double precision, intent(inout):: u_0(sze,N_st)
integer :: i,j,k
integer :: ithread
double precision, allocatable :: u_t(:,:)
!DIR$ ATTRIBUTES ALIGN : $IRP_ALIGN :: u_t
integer(ZMQ_PTR) :: zmq_to_qp_run_socket, zmq_socket_pull
PROVIDE psi_det_beta_unique psi_bilinear_matrix_order_transp_reverse psi_det_alpha_unique
PROVIDE psi_bilinear_matrix_transp_values psi_bilinear_matrix_values psi_bilinear_matrix_columns_loc
PROVIDE ref_bitmask_energy nproc
PROVIDE mpi_initialized
call new_parallel_job(zmq_to_qp_run_socket,zmq_socket_pull,'davidson')
! integer :: N_states_diag_save
! N_states_diag_save = N_states_diag
! N_states_diag = N_st
if (zmq_put_N_states_diag(zmq_to_qp_run_socket, 1) == -1) then
stop 'Unable to put N_states_diag on ZMQ server'
endif
if (zmq_put_psi(zmq_to_qp_run_socket,1) == -1) then
stop 'Unable to put psi on ZMQ server'
endif
energy = 0.d0
if (zmq_put_dvector(zmq_to_qp_run_socket,1,'energy',energy,size(energy)) == -1) then
stop 'Unable to put energy on ZMQ server'
endif
! Create tasks
! ============
integer :: istep, imin, imax, ishift, ipos
integer, external :: add_task_to_taskserver
integer, parameter :: tasksize=20000
character*(100000) :: task
istep=1
ishift=0
imin=1
ipos=1
do imin=1,N_det,tasksize
imax = min(N_det,imin-1+tasksize)
if (imin==1) then
istep = 2
else
istep = 1
endif
do ishift=0,istep-1
write(task(ipos:ipos+50),'(4(I11,1X),1X,1A)') imin, imax, ishift, istep, '|'
ipos = ipos+50
if (ipos > 100000-50) then
if (add_task_to_taskserver(zmq_to_qp_run_socket,trim(task(1:ipos))) == -1) then
stop 'Unable to add task'
endif
ipos=1
endif
enddo
enddo
if (ipos > 1) then
if (add_task_to_taskserver(zmq_to_qp_run_socket,trim(task(1:ipos))) == -1) then
stop 'Unable to add task'
endif
ipos=1
endif
allocate(u_t(N_st,N_det))
do k=1,N_st
call dset_order(u_0(1,k),psi_bilinear_matrix_order,N_det)
enddo
call dtranspose( &
u_0, &
size(u_0, 1), &
u_t, &
size(u_t, 1), &
N_det, N_st)
ASSERT (N_st == N_states_diag)
ASSERT (sze >= N_det)
integer :: rc, ni, nj
integer*8 :: rc8
double precision :: energy(N_st)
integer, external :: zmq_put_dvector, zmq_put_psi, zmq_put_N_states_diag
integer, external :: zmq_put_dmatrix
if (size(u_t) < 8388608) then
ni = size(u_t)
nj = 1
else
ni = 8388608
nj = size(u_t)/8388608 + 1
endif
! Warning : dimensions are modified for efficiency, It is OK since we get the
! full matrix
if (zmq_put_dmatrix(zmq_to_qp_run_socket, 1, 'u_t', u_t, ni, nj, size(u_t,kind=8)) == -1) then
stop 'Unable to put u_t on ZMQ server'
endif
deallocate(u_t)
integer, external :: zmq_set_running
if (zmq_set_running(zmq_to_qp_run_socket) == -1) then
print *, irp_here, ': Failed in zmq_set_running'
endif
call set_multiple_levels_omp(.True.)
!$OMP PARALLEL DEFAULT(shared) NUM_THREADS(2) PRIVATE(ithread)
ithread = omp_get_thread_num()
if (ithread == 0 ) then
call davidson_nos2_collector(zmq_to_qp_run_socket, zmq_socket_pull, v_0, N_det, N_st)
else
call davidson_nos2_slave_inproc(1)
endif
!$OMP END PARALLEL
call end_parallel_job(zmq_to_qp_run_socket, zmq_socket_pull, 'davidson')
!$OMP PARALLEL
!$OMP SINGLE
do k=1,N_st
!$OMP TASK DEFAULT(SHARED) FIRSTPRIVATE(k,N_det)
call dset_order(v_0(1,k),psi_bilinear_matrix_order_reverse,N_det)
!$OMP END TASK
!$OMP TASK DEFAULT(SHARED) FIRSTPRIVATE(k,N_det)
call dset_order(u_0(1,k),psi_bilinear_matrix_order_reverse,N_det)
!$OMP END TASK
enddo
!$OMP END SINGLE
!$OMP TASKWAIT
!$OMP END PARALLEL
! N_states_diag = N_states_diag_save
! SOFT_TOUCH N_states_diag
end