Tuned pt2_F

ocaml/TaskServer.ml

@@ -50,7 +50,7 @@ let zmq_context =
   Zmq.Context.create ()
 
 let () =
-  Zmq.Context.set_io_threads zmq_context 8
+  Zmq.Context.set_io_threads zmq_context 16
 
 
 let bind_socket ~socket_type ~socket ~port =

src/cipsi/pt2_stoch_routines.irp.f

@@ -13,9 +13,9 @@ END_PROVIDER
   integer :: i
   integer :: e
   e = elec_num - n_core_orb * 2
-  pt2_n_tasks_max = 1+min((e*(e-1))/2, int(dsqrt(dble(N_det_selectors)))/4)
+  pt2_n_tasks_max = 1+min((e*(e-1)), int(dsqrt(dble(N_det_selectors)))/10)
   do i=1,N_det_generators
-    pt2_F(i) = 1 + int(dble(pt2_n_tasks_max)*dsqrt(dabs(psi_coef_sorted_gen(i,pt2_stoch_istate))))
+    pt2_F(i) = 1 + int(dble(pt2_n_tasks_max)*dabs(maxval(psi_coef_sorted_gen(i,:)))**(0.75d0))
   enddo
 END_PROVIDER
 
@@ -120,7 +120,11 @@ subroutine ZMQ_pt2(E, pt2,relative_error, error, variance, norm, N_in)
   PROVIDE psi_bilinear_matrix_rows psi_det_sorted_order psi_bilinear_matrix_order
   PROVIDE psi_bilinear_matrix_transp_rows_loc psi_bilinear_matrix_transp_columns
   PROVIDE psi_bilinear_matrix_transp_order psi_selectors_coef_transp psi_det_sorted
+  PROVIDE psi_det_hii
 
+  if (s2_eig) then
+    PROVIDE psi_occ_pattern_hii det_to_occ_pattern
+  endif
 
   if (N_det < max(10,N_states)) then
     pt2=0.d0

src/cipsi/run_pt2_slave.irp.f

@@ -93,7 +93,7 @@ subroutine run_pt2_slave(thread,iproc,energy)
 !print *,  i_generator(1), time1-time2, n_tasks, pt2_F(i_generator(1))
     enddo
     call wall_time(time1)
-!print *,  i_generator(1), time1-time0, n_tasks
+!print *,  '-->', i_generator(1), time1-time0, n_tasks
 
     integer, external :: tasks_done_to_taskserver
     if (tasks_done_to_taskserver(zmq_to_qp_run_socket,worker_id,task_id,n_tasks) == -1) then

src/cipsi/selection.irp.f

@@ -1266,4 +1266,4 @@ subroutine bitstring_to_list_in_selection( string, list, n_elements, Nint)
   enddo
 
 end
-
+!

src/cipsi/slave_cipsi.irp.f

@@ -279,6 +279,11 @@ subroutine run_slave_main
           print *,  'pt2_stoch_istate', pt2_stoch_istate
           print *,  'state_average_weight', state_average_weight
           print *,  'Number of threads', nproc_target
+          PROVIDE psi_det_hii
+
+          if (s2_eig) then
+            PROVIDE psi_occ_pattern_hii det_to_occ_pattern
+          endif
         endif
 
         !$OMP PARALLEL PRIVATE(i) NUM_THREADS(nproc_target+1)

src/determinants/determinants.irp.f

@@ -877,9 +877,11 @@ BEGIN_PROVIDER [ double precision, psi_det_Hii, (N_det) ]
  END_DOC
  integer :: i,j
  double precision, external :: diag_H_mat_elem
+ !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i)
  do i=1,N_det
    psi_det_Hii(i) = diag_H_mat_elem(psi_det(1,1,i),N_int)
  enddo
+ !$OMP END PARALLEL DO
 END_PROVIDER
 
 

src/determinants/occ_pattern.irp.f

@@ -408,11 +408,13 @@ subroutine make_s2_eigenfunction
   integer                        :: N_det_new, ithread, omp_get_thread_num
   integer, parameter             :: bufsze = 1000
   logical, external              :: is_in_wavefunction
+  logical                        :: update
 
+  update=.False.
   call write_int(6,N_occ_pattern,'Number of occupation patterns')
 
   !$OMP PARALLEL DEFAULT(NONE) &
-  !$OMP  SHARED(N_occ_pattern, psi_occ_pattern, elec_alpha_num,N_int) &
+  !$OMP  SHARED(N_occ_pattern, psi_occ_pattern, elec_alpha_num,N_int,update) &
   !$OMP  PRIVATE(s,ithread, d, det_buffer, smax, N_det_new,i,j,k)
   N_det_new = 0
   call occ_pattern_to_dets_size(psi_occ_pattern(1,1,1),s,elec_alpha_num,N_int)
@@ -434,6 +436,7 @@ subroutine make_s2_eigenfunction
       if ( is_in_wavefunction(d(1,1,j), N_int) ) then
         cycle
       endif
+      update = .true.
       N_det_new += 1
       det_buffer(:,:,N_det_new) = d(:,:,j)
       if (N_det_new == bufsze) then
@@ -451,8 +454,10 @@ subroutine make_s2_eigenfunction
   deallocate(d,det_buffer)
   !$OMP END PARALLEL
 
+  if (update) then
     call copy_H_apply_buffer_to_wf
-  SOFT_TOUCH N_det psi_coef psi_det psi_occ_pattern N_occ_pattern
+    TOUCH N_det psi_coef psi_det psi_occ_pattern N_occ_pattern
+  endif
   call write_time(6)
 
 end