quantum-espresso/GWW/gww/do_self_lanczos_time.f90

1024 lines
36 KiB
Fortran

!
! Copyright (C) 2001-2013 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 .
!
!
subroutine do_reducible_pola(tf ,options)
!this subroutine calculates and writes on disk the reducible polarizability from the screen interaction
USE kinds, ONLY : DP
USE io_global, ONLY : stdout, ionode, ionode_id
USE input_gw, ONLY : input_options
USE basic_structures, ONLY : v_pot,wannier_u,free_memory, initialize_memory,lanczos_chain, vt_mat_lanczos,tt_mat_lanczos,&
& semicore
USE green_function, ONLY : green, read_green, free_memory_green, initialize_green
USE polarization, ONLY : polaw, free_memory_polaw, read_polaw, write_polaw,invert_v_pot, initialize_polaw, &
& read_polaw_global
USE mp, ONLY : mp_sum, mp_bcast
USE mp_world, ONLY : nproc,mpime,world_comm
USE times_gw, ONLY : times_freqs
USE self_energy_storage, ONLY : self_storage,write_self_storage_ondisk,free_memory_self_storage
USE lanczos
USE constants, ONLY : tpi,pi
USE start_end ! debug
USE parallel_include
implicit none
TYPE(times_freqs), INTENT(in) :: tf!for time frequency grids
TYPE(input_options) :: options
TYPE(self_storage) :: ss
TYPE(v_pot) :: vp,vpi
TYPE(polaw) :: ww!dressed polarization
INTEGER :: l_blk, nbegin,nend, nsize
REAL(kind=DP), ALLOCATABLE:: wtemp(:,:)
INTEGER :: iw
REAL(kind=DP) :: v_head
nullify(vp%vmat)
nullify(vpi%vmat)
call initialize_polaw(ww)
write(stdout,*) 'Trasform W to Pgreek'
FLUSH(stdout)
if(options%w_divergence == 2) then
call read_data_pw_v(vp,options%prefix,options%debug,0,.true.)
else
call read_data_pw_v(vp,options%prefix,options%debug,0,.false.)
endif
v_head=vp%vmat(vp%numpw,vp%numpw)
call invert_v_pot(vp,vpi)
call free_memory(vp)
l_blk= (tf%n+1)/nproc
if(l_blk*nproc < (tf%n+1)) l_blk = l_blk+1
nbegin=mpime*l_blk
nend=nbegin+l_blk-1
!loop on imaginary frequency i\omega
do iw=nbegin,nbegin+l_blk-1
if(iw <= tf%n) then
call read_polaw(iw,ww,options%debug,options%l_verbose)
allocate(wtemp(ww%numpw,ww%numpw))
call dgemm('N','N',ww%numpw,ww%numpw,ww%numpw,1.d0,vpi%vmat,ww%numpw,ww%pw,ww%numpw,&
&0.d0, wtemp,ww%numpw)
call dgemm('N','N',ww%numpw,ww%numpw,ww%numpw,1.d0,wtemp,ww%numpw,vpi%vmat,ww%numpw,&
&0.d0,ww%pw,ww%numpw)
deallocate(wtemp)
call write_polaw(ww,options%debug)
endif
enddo
call free_memory(vpi)
call free_memory_polaw(ww)
write(stdout,*) 'Done'
FLUSH(stdout)
return
end subroutine do_reducible_pola
subroutine do_self_lanczos_time(ss, tf ,options,l_real_axis,energy)
!this subroutine calculte the self-energy on time using fourier trasfrom using the lanczos scheme
USE kinds, ONLY : DP
USE io_global, ONLY : stdout, ionode, ionode_id
USE input_gw, ONLY : input_options
USE basic_structures, ONLY : v_pot,wannier_u,free_memory, initialize_memory,lanczos_chain, vt_mat_lanczos,tt_mat_lanczos,&
& semicore
USE green_function, ONLY : green, read_green, free_memory_green, initialize_green
USE polarization, ONLY : polaw, free_memory_polaw, read_polaw, write_polaw,invert_v_pot, initialize_polaw, &
& read_polaw_global
USE mp, ONLY : mp_sum, mp_bcast
USE mp_world, ONLY : nproc,mpime,world_comm
USE times_gw, ONLY : times_freqs
USE self_energy_storage, ONLY : self_storage,write_self_storage_ondisk,free_memory_self_storage
USE lanczos
USE constants, ONLY : tpi,pi
USE start_end ! debug
USE parallel_include
USE io_files, ONLY : prefix, tmp_dir
implicit none
TYPE(times_freqs), INTENT(in) :: tf!for time frequency grids
TYPE(input_options) :: options
TYPE(self_storage) :: ss
LOGICAL, INTENT(in) :: l_real_axis
!if true calculates on real frequency axis at given energy
REAL(kind=DP), INTENT(in) :: energy!energy on real axis at which calculating the self-energy (or part of it)
!only if l_real_axis == true
TYPE(v_pot) :: vp,vpi
TYPE(polaw) :: ww!dressed polarization
REAL(kind=DP) :: inv_epsi,v_head
INTEGER :: l_blk, nbegin,nend, nsize, l_blk_freq, nbegin_freq,nend_freq
REAL(kind=DP), ALLOCATABLE:: wtemp(:,:)
INTEGER :: iw
TYPE(wannier_u) :: uu
REAL(kind=DP) :: offset
TYPE(lanczos_chain) :: lc
REAL(kind=DP), ALLOCATABLE :: re_e_mat(:,:,:),im_e_mat(:,:,:)
COMPLEX(kind=DP), ALLOCATABLE :: e_mat_tmp(:,:,:)
COMPLEX(kind=DP) :: af(1)
REAL(kind=DP), ALLOCATABLE :: pw_mat(:,:,:),pw_mat_t(:,:,:)
INTEGER :: numpw,numt,numl
INTEGER :: it, ii,jj
REAL(kind=DP) :: time,factor
COMPLEX(kind=DP) :: cfactor
REAL(kind=DP), ALLOCATABLE :: pw_tmp(:,:),pw_dumm(:,:)
INTEGER :: iproc_time,ierr
TYPE(vt_mat_lanczos) :: vtl,vtl_j
TYPE(tt_mat_lanczos) :: ttl,ttl_j
REAL(kind=DP), ALLOCATABLE :: re_g_mat(:,:,:), im_g_mat(:,:,:),re_g_mat_t(:,:,:), im_g_mat_t(:,:,:)
REAL(kind=DP), ALLOCATABLE :: tmp_mat(:,:),tmp_mat1(:,:),tmp_mat2(:,:)
REAL(kind=DP), ALLOCATABLE :: g_tmp(:,:), g_dumm(:,:), re_h_mat(:,:),im_h_mat(:,:)
REAL(kind=DP), EXTERNAL :: DDOT
LOGICAL :: l_single=.true.!if true e_mat is saved in single precision
REAL(kind=4), ALLOCATABLE :: re_e_mat_single(:,:,:),im_e_mat_single(:,:,:)
REAL(kind=DP), ALLOCATABLE :: e_mat_double(:,:)
INTEGER :: l_blk_t, nbegin_t,nend_t, nsize_t,in
INTEGER, PARAMETER :: ndivt=1!10
INTEGER :: l_blk_g, nbegin_g,nend_g, nsize_g!paremter for optional dedicated frequency grid for G
INTEGER :: j_min, j_max, is
TYPE(semicore) :: sc
REAL(kind=DP), ALLOCATABLE :: tmp_vec_sc(:)
INTEGER :: iv_sc
LOGICAL, PARAMETER :: l_distribute_sm=.true.!if true the S matrices are distributed among mpi tasks instead of being read from disk
!it requires the parameter l_single==false
INTEGER :: l_blk_sm, nbegin_sm,nend_sm, nsize_sm,iproc
REAL(kind=DP), ALLOCATABLE :: st_save(:,:,:), sl_save(:,:,:)
COMPLEX(kind=DP), ALLOCATABLE :: exp_table(:,:)
REAL(kind=DP), ALLOCATABLE :: re_e_mat_t(:,:,:),im_e_mat_t(:,:,:)!in time for fourier trasform
REAL(kind=DP), ALLOCATABLE :: re_e_mat_part(:,:,:),im_e_mat_part(:,:,:)!in time for storing partial calculations
REAL(kind=DP), ALLOCATABLE :: pw_part_t(:,:,:)!in time for storing partial calculations
INTEGER :: n_cycles, i_cycles,i_min_cycles,i_max_cycles
INTEGER :: n_list(2),iun,iun2
INTEGER, ALLOCATABLE :: i_list(:,:)
INTEGER, EXTERNAL :: find_free_unit
if(options%whole_s) then
l_single=.false.
endif
write(stdout,*) 'Routine do_self_lanczos_time'
FLUSH(stdout)
if(options%l_big_system) then
n_cycles=options%i_max-options%i_min+1
else
n_cycles=1
endif
if(options%l_list) then
if(ionode) then
iun = find_free_unit()
open( unit=iun, file=trim(tmp_dir)//trim(prefix)//'-'//'list_1.dat', status='old')
read(iun,*) n_list(1)
if(uu%nspin==2) then
iun2 = find_free_unit()
open( unit=iun2, file=trim(tmp_dir)//trim(prefix)//'-'//'list_2.dat', status='old')
read(iun,*) n_list(2)
else
n_list(2)=0
endif
endif
call mp_bcast(n_list,ionode_id,world_comm)
allocate(i_list(max(n_list(1),n_list(2)),2))
i_list=0
if(ionode) then
do ii=1,n_list(1)
read(iun,*) i_list(ii,1)
enddo
close(iun)
if(uu%nspin==2) then
do ii=1,n_list(2)
read(iun2,*) i_list(ii,2)
enddo
close(iun2)
endif
endif
call mp_bcast(i_list,ionode_id,world_comm)
n_cycles=n_list(1)
endif
!keeps in memory G, P(i\tau)
nullify(vp%vmat)
nullify(vpi%vmat)
call initialize_polaw(ww)
call initialize_memory(sc)
!calculate offset
!read in DFT energies
call read_data_pw_u(uu,options%prefix)
ss%ontime=.true.
ss%max_i=options%max_i
ss%i_min=options%i_min
ss%i_max=options%i_max
ss%n=tf%n
ss%tau=options%tau
ss%whole_s=options%whole_s
ss%n_grid_fit=tf%n_grid_fit
if(ss%whole_s) then
ss%i_min_whole=options%i_min_whole
ss%i_max_whole=options%i_max_whole
endif
ss%nspin=uu%nspin
if(ss%whole_s) then
allocate(ss%whole(ss%i_min_whole:ss%i_max_whole,ss%max_i,2*ss%n+1,ss%nspin))
ss%whole(:,:,:,:)=(0.d0,0.d0)
allocate(ss%whole_freq_fit(ss%i_min_whole:ss%i_max_whole,ss%max_i,2*ss%n_grid_fit+1,ss%nspin))
ss%whole_freq_fit(:,:,:,:)=(0.d0,0.d0)
allocate(ss%diag(ss%max_i,2*ss%n+1,ss%nspin))
ss%diag(:,:,:)=(0.d0,0.d0)
allocate(ss%diag_freq_fit(ss%max_i,2*ss%n_grid_fit+1,ss%nspin))
ss%diag_freq_fit(:,:,:)=(0.d0,0.d0)
else
allocate(ss%diag(ss%max_i,2*ss%n+1,ss%nspin))
ss%diag(:,:,:)=(0.d0,0.d0)
nullify(ss%whole)
allocate(ss%diag_freq_fit(ss%max_i,2*ss%n_grid_fit+1,ss%nspin))
ss%diag_freq_fit(:,:,:)=(0.d0,0.d0)
nullify(ss%whole_freq_fit)
endif
!for compatibility
allocate(ss%ene_remainder(ss%max_i,1))
ss%ene_remainder(:,1)=0.d0
!loop on spin
do is=1,ss%nspin
!NOT_TO_BE_INCLUDED_START
if(options%l_semicore) call read_data_pw_semicore(sc, options%prefix, is)
!NOT_TO_BE_INCLUDED_END
if(.not.l_real_axis) then
if(uu%nums > uu%nums_occ(is)) then
offset=-(uu%ene(uu%nums_occ(is)+1,is)+uu%ene(uu%nums_occ(is),is))/2.d0! CUSSI XE GIUSTO DEBUG
!offset=-(uu%ene(uu%nums_occ(2)+1,2)+uu%ene(uu%nums_occ(1),1))/2.d0
else
offset=-uu%ene(uu%nums_occ(is),is)
endif
else
offset=-energy
endif
!!!!!!!!
l_blk= (tf%n+1)/nproc
if(l_blk*nproc < (tf%n+1)) l_blk = l_blk+1
nbegin=mpime*l_blk
nend=nbegin+l_blk-1
if(nend > tf%n) nend=tf%n
nsize=nend-nbegin+1
!read polarizability matrices
if( options%l_verbose) write(stdout,*) 'Read Pgreek'
FLUSH(stdout)
do iw=nbegin,nend
call read_polaw(iw,ww,options%debug,options%l_verbose)
if(iw==nbegin) allocate(pw_mat(ww%numpw,ww%numpw,l_blk))
pw_mat(:,:,iw-nbegin+1)=ww%pw(:,:)
call free_memory_polaw(ww)
enddo
numpw=ww%numpw
call mp_bcast(numpw, ionode_id,world_comm)
if(nbegin > tf%n) allocate(pw_mat(numpw,numpw,l_blk))
!Fourier trasform reducible polarizability matrices to imaginary time
write(stdout,*) 'Fourier trasform Pgreek'
FLUSH(stdout)
allocate(pw_tmp(numpw,numpw))
allocate(pw_dumm(numpw,numpw))
allocate(pw_mat_t(numpw,numpw,l_blk))
!loop on time
do it=0,tf%n
!each procs sums up its matrices in frequency with opportune factor
time=tf%times(it)
pw_tmp(:,:)=0.d0
do iw=nbegin,nend
factor=2.d0*dble(tf%weights_freq(iw)*exp((0.d0,1.d0)*tf%times(it)*tf%freqs_eff(iw)))/(2.d0*pi)
pw_tmp(:,:)=pw_tmp(:,:)+pw_mat(:,:,iw-nbegin+1)*factor
enddo
#if defined(__MPI)
!the distribution of times on procs is the same of that for frequecies
iproc_time=it/l_blk
!all processors sums to iproc_time
if(iproc_time==mpime) then
call MPI_REDUCE(pw_tmp,pw_mat_t(1,1,it-nbegin+1),numpw*numpw,MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
else
call MPI_REDUCE(pw_tmp,pw_dumm,numpw*numpw,MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
endif
#else
pw_mat_t(:,:,it+1)=pw_tmp
#endif
!mp_sum to processer owing that time
!this processor put on opportune array
enddo
deallocate(pw_tmp)
deallocate(pw_dumm)
deallocate(pw_mat)
l_blk_g= (tf%n_g+1)/nproc
if(l_blk_g*nproc < (tf%n_g+1)) l_blk_g = l_blk_g+1
nbegin_g=mpime*l_blk_g
nend_g=nbegin_g+l_blk_g-1
if(nend_g > tf%n_g) nend_g=tf%n_g
nsize_g=nend_g-nbegin_g+1
!allocate and compute table for Fourier trasform
allocate(exp_table(tf%n+1,l_blk_g))
do it=0,tf%n
do iw=nbegin_g,nend_g
exp_table(it+1,iw-nbegin_g+1)=exp((0.d0,1.d0)*tf%times(it)*tf%freqs_g_eff(iw))
enddo
enddo
do i_cycles=1,n_cycles
!calculates G
call initialize_memory(lc)
if(.not.options%l_list) then
call read_data_pw_lanczos_chain(lc, i_cycles, options%prefix, .false.,is)
else
call read_data_pw_lanczos_chain(lc, i_list(i_cycles,is), options%prefix, .false.,is)
endif
write(stdout,*) 'Lanczos dimensions', lc%numt,lc%num_steps,is
FLUSH(stdout)
if(.not.l_single) then
allocate(re_e_mat(lc%numt,lc%numt,l_blk_g))
allocate(im_e_mat(lc%numt,lc%numt,l_blk_g))
allocate(e_mat_tmp(lc%numt,lc%numt,1))
do iw=nbegin_g,nbegin_g+l_blk_g-1
if(iw <= tf%n_g) then
af(1)=dcmplx(offset,-tf%freqs_g(iw))
call solve_lanczos_complex(1,af,e_mat_tmp,lc)
re_e_mat(:,:,iw-nbegin_g+1)=dble(e_mat_tmp(:,:,1))
im_e_mat(:,:,iw-nbegin_g+1)=dimag(e_mat_tmp(:,:,1))
else
call solve_lanczos_fake_complex(lc)
endif
end do
deallocate(e_mat_tmp)
!for entire self-energy store
!!!!!!!!!!!!!!
if( ss%whole_s .and.l_distribute_sm ) then
if( options%l_verbose) write(stdout,*) 'before allocate',lc%numt,l_blk
FLUSH(stdout)
allocate(re_e_mat_t(lc%numt,lc%numt,l_blk), im_e_mat_t(lc%numt,lc%numt,l_blk))
if( options%l_verbose) write(stdout,*) 'after '
allocate(g_tmp(lc%numt,lc%numt))
allocate(g_dumm(lc%numt,lc%numt))
if( options%l_verbose) write(stdout,*) 'ATT1'
FLUSH(stdout)
numt=lc%numt
!loop on time
do it=0,tf%n
!each procs sums up its matrices in frequency with opportune factor
time=tf%times(it)
g_tmp(:,:)=0.d0
do iw=nbegin_g,nend_g
factor=2.d0*dble(tf%weights_freq_g(iw)*exp_table(it+1,iw-nbegin_g+1))/(2.d0*pi)
g_tmp(1:numt,1:numt)=g_tmp(1:numt,1:numt)+re_e_mat(1:numt,1:numt,iw-nbegin_g+1)*factor
enddo
#if defined(__MPI)
!the distribution of times on procs is the same of that for frequecies
iproc_time=it/l_blk
!all processors sums to iproc_time
if(iproc_time==mpime) then
call MPI_REDUCE(g_tmp,re_e_mat_t(1,1,it-nbegin+1),numt*numt,&
&MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
else
call MPI_REDUCE(g_tmp,g_dumm,numt*numt,MPI_DOUBLE_PRECISION,&
&MPI_SUM,iproc_time,world_comm,ierr)
endif
#else
re_e_mat_t(1:numt,1:numt,it+1)=g_tmp(1:numt,1:numt)
#endif
enddo
!loop on time
do it=0,tf%n
!each procs sums up its matrices in frequency with opportune factor
time=tf%times(it)
g_tmp(:,:)=0.d0
do iw=nbegin_g,nend_g
factor=-2.d0*dimag(tf%weights_freq_g(iw)*exp_table(it+1,iw-nbegin_g+1))/(2.d0*pi)
g_tmp(1:numt,1:numt)=g_tmp(1:numt,1:numt)+im_e_mat(1:numt,1:numt,iw-nbegin_g+1)*factor
enddo
#if defined(__MPI)
!the distribution of times on procs is the same of that for frequecies
iproc_time=it/l_blk
!all processors sums to iproc_time
if(iproc_time==mpime) then
call MPI_REDUCE(g_tmp,im_e_mat_t(1,1,it-nbegin+1),numt*numt,&
&MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
else
call MPI_REDUCE(g_tmp,g_dumm,numt*numt,MPI_DOUBLE_PRECISION,&
&MPI_SUM,iproc_time,world_comm,ierr)
endif
#else
im_e_mat_t(1:numt,1:numt,it+1)=g_tmp(1:numt,1:numt)
#endif
enddo
deallocate(g_tmp,g_dumm)
endif
if( options%l_verbose) write(stdout,*) 'ATT2' !DEBUG
FLUSH(stdout)
else
if(i_cycles==1) then
allocate(re_e_mat_single(lc%numt,lc%numt,l_blk_g))
allocate(im_e_mat_single(lc%numt,lc%numt,l_blk_g))
endif
do iw=nbegin_g,nbegin_g+l_blk_g-1
if(iw <= tf%n_g) then
af(1)=dcmplx(offset,-tf%freqs_g(iw))
call solve_lanczos_single(af(1),re_e_mat_single(1,1,iw-nbegin_g+1),im_e_mat_single(1,1,iw-nbegin_g+1),lc)
else
call solve_lanczos_fake_single(lc)
endif
end do
l_blk_t= (lc%numt)/ndivt
if(l_blk_t*ndivt < (lc%numt)) l_blk_t = l_blk_t+1
if(i_cycles==1) allocate(e_mat_double(lc%numt,l_blk_t))
endif
if( options%l_verbose) write(stdout,*) 'Done'
FLUSH(stdout)
call initialize_memory(vtl)
call initialize_memory(ttl)
call initialize_memory(vtl_j)
call initialize_memory(ttl_j)
!if required read all S matrices and distribute among mpi tasks
if( ss%whole_s .and.l_distribute_sm ) then
l_blk_sm= (ss%max_i)/nproc
if(l_blk_sm*nproc < (ss%max_i)) l_blk_sm = l_blk_sm+1
nbegin_sm=mpime*l_blk_sm+1
nend_sm=nbegin_sm+l_blk_sm-1
if(nend_sm > ss%max_i) nend_sm=ss%max_i
nsize_sm=nend_sm-nbegin_sm+1
do jj=1,ss%max_i
call read_data_pw_vt_mat_lanczos(vtl_j, jj, options%prefix, .false.,is)
call read_data_pw_tt_mat_lanczos(ttl_j, jj, options%prefix, .false.,is)
if(jj==1) then
numt=ttl_j%numt
numl=ttl_j%numl
allocate(sl_save(numpw,numl,l_blk_sm))
allocate(st_save(numt,numl,l_blk_sm))
endif
if(jj>=nbegin_sm .and. jj<=nend_sm) then
st_save(1:numt,1:numl,jj-nbegin_sm+1)=ttl_j%tt_mat(1:numt,1:numl)
sl_save(1:numpw,1:numl,jj-nbegin_sm+1)=vtl_j%vt_mat(1:numpw,1:numl)
endif
call free_memory(vtl_j)
call free_memory(ttl_j)
enddo
allocate(re_e_mat_part(numt,numl,l_blk),im_e_mat_part(numt,numl,l_blk))
allocate(pw_part_t(numpw,numl,l_blk))
endif
!loop on KS states
if(options%l_big_system) then
if(.not.options%l_list) then
i_min_cycles=ss%i_min+i_cycles-1
i_max_cycles=i_min_cycles
else
i_min_cycles=i_list(i_cycles,is)
i_max_cycles=i_min_cycles
endif
else
i_min_cycles=ss%i_min
i_max_cycles=ss%i_max
endif
do ii=i_min_cycles,i_max_cycles
write(stdout,*) 'Loop on KS:',ii, is
FLUSH(stdout)
!calculates G on partial basis
!read vtl ttl
call read_data_pw_vt_mat_lanczos(vtl, ii, options%prefix, .false.,is)
call read_data_pw_tt_mat_lanczos(ttl, ii, options%prefix, .false.,is)
if(ii==i_min_cycles.and.i_cycles==1) then
numt=ttl%numt
numl=ttl%numl
allocate(re_g_mat(numl,numl,l_blk_g),im_g_mat(numl,numl,l_blk_g))
allocate(re_g_mat_t(numl,numl,l_blk),im_g_mat_t(numl,numl,l_blk))
endif
!multiply and put on array
!loop on frequency
if(.not.ss%whole_s) then
j_min=ii
j_max=ii
else
j_min=ss%i_min_whole
j_max=ss%i_max_whole
endif
if( options%l_verbose) write(stdout,*) 'Doing dgemms',numl,numt,numl,numpw
if(ss%whole_s .and. l_distribute_sm) then!if required calculates partial products
do it=nbegin,nend
call dgemm('N','N',numt,numl,numt,1.d0,re_e_mat_t(1,1,it-nbegin+1),numt,&
&ttl%tt_mat,numt,0.d0,re_e_mat_part(1,1,it-nbegin+1),numt)
call dgemm('N','N',numt,numl,numt,1.d0,im_e_mat_t(1,1,it-nbegin+1),numt,&
&ttl%tt_mat,numt,0.d0,im_e_mat_part(1,1,it-nbegin+1),numt)
call dgemm('N','N',numpw,numl,numpw,1.d0,pw_mat_t(1,1,it-nbegin+1) ,numpw, &
& vtl%vt_mat,numpw,0.d0,pw_part_t(1,1,it-nbegin+1),numpw)
enddo
endif
FLUSH(stdout)
do jj=j_min,j_max
if(ss%whole_s .and. l_distribute_sm) then
allocate(ttl_j%tt_mat(numt,numl))
allocate(vtl_j%vt_mat(numpw,numl))
if(jj>=nbegin_sm .and. jj<=nend_sm) then
vtl_j%vt_mat(1:numpw,1:numl)=sl_save(1:numpw,1:numl,jj-nbegin_sm+1)
ttl_j%tt_mat(1:numt,1:numl)=st_save(1:numt,1:numl,jj-nbegin_sm+1)
endif
iproc=(jj-1)/l_blk_sm
call mp_bcast(vtl_j%vt_mat, iproc,world_comm)
call mp_bcast(ttl_j%tt_mat, iproc,world_comm)
else
call read_data_pw_vt_mat_lanczos(vtl_j, jj, options%prefix, .false.,is)
call read_data_pw_tt_mat_lanczos(ttl_j, jj, options%prefix, .false.,is)
endif
if(.not.(ss%whole_s .and. l_distribute_sm)) then
allocate(tmp_mat(numl,numt))
do iw=nbegin_g,nend_g
if( options%l_verbose) write(stdout,*) 'Doing dgemms',numl,numt,numpw,l_blk,iw
FLUSH(stdout)
if(.not.l_single) then
call dgemm('T','N',numl,numt,numt,1.d0,ttl_j%tt_mat,numt,re_e_mat(1,1,iw-nbegin_g+1),numt,0.d0,tmp_mat,numl)
call dgemm('N','N',numl,numl,numt,1.d0,tmp_mat,numl,ttl%tt_mat,numt,0.d0,re_g_mat(1,1,iw-nbegin_g+1),numl)
call dgemm('T','N',numl,numt,numt,1.d0,ttl_j%tt_mat,numt,im_e_mat(1,1,iw-nbegin_g+1),numt,0.d0,tmp_mat,numl)
call dgemm('N','N',numl,numl,numt,1.d0,tmp_mat,numl,ttl%tt_mat,numt,0.d0,im_g_mat(1,1,iw-nbegin_g+1),numl)
else
do in=0,ndivt-1
nbegin_t=in*l_blk_t+1
nend_t=min(nbegin_t+l_blk_t-1,lc%numt)
nsize_t=nend_t-nbegin_t+1
if(nsize_t >= 1) then
e_mat_double(1:lc%numt, 1:nsize_t)=dble(re_e_mat_single(1:lc%numt, nbegin_t:nend_t, iw-nbegin_g+1))
call dgemm('T','N',numl,nsize_t,numt,1.d0,ttl_j%tt_mat,numt,e_mat_double,numt,0.d0,&
&tmp_mat(1,nbegin_t),numl)
endif
enddo
if( options%l_verbose) write(stdout,*) 'ATT1'!DEBUG
FLUSH(stdout)
call dgemm('N','N',numl,numl,numt,1.d0,tmp_mat,numl,ttl%tt_mat,numt,0.d0,&
&re_g_mat(1,1,iw-nbegin_g+1),numl)
if( options%l_verbose) write(stdout,*) 'ATT2'!DEBUG
FLUSH(stdout)
do in=0,ndivt-1
nbegin_t=in*l_blk_t+1
nend_t=min(nbegin_t+l_blk_t-1,lc%numt)
nsize_t=nend_t-nbegin_t+1
e_mat_double(1:lc%numt, 1:nsize_t)=dble(im_e_mat_single(1:lc%numt, nbegin_t:nend_t, iw-nbegin_g+1))
call dgemm('T','N',numl,nsize_t,numt,1.d0,ttl_j%tt_mat,numt,e_mat_double,numt,0.d0,tmp_mat(1,nbegin_t),numl)
enddo
if( options%l_verbose) write(stdout,*) 'ATT3'!DEBUG
FLUSH(stdout)
call dgemm('N','N',numl,numl,numt,1.d0,tmp_mat,numl,ttl%tt_mat,numt,0.d0,im_g_mat(1,1,iw-nbegin_g+1),numl)
endif
enddo
deallocate(tmp_mat)
write(stdout,*) 'Fourier trasform:'
FLUSH(stdout)
!Fourier trasform
allocate(g_tmp(numl,numl))
allocate(g_dumm(numl,numl))
if( options%l_verbose) write(stdout,*) 'ATT1'
FLUSH(stdout)
!loop on time
do it=0,tf%n
!each procs sums up its matrices in frequency with opportune factor
time=tf%times(it)
g_tmp(:,:)=0.d0
do iw=nbegin_g,nend_g
!factor=2.d0*dble(tf%weights_freq_g(iw)*exp((0.d0,1.d0)*tf%times(it)*tf%freqs_g_eff(iw)))/(2.d0*pi)
factor=2.d0*dble(tf%weights_freq_g(iw)*exp_table(it+1,iw-nbegin_g+1))/(2.d0*pi)
g_tmp(:,:)=g_tmp(:,:)+re_g_mat(:,:,iw-nbegin_g+1)*factor
enddo
#if defined(__MPI)
!the distribution of times on procs is the same of that for frequecies
iproc_time=it/l_blk
!all processors sums to iproc_time
if(iproc_time==mpime) then
call MPI_REDUCE(g_tmp,re_g_mat_t(1,1,it-nbegin+1),numl*numl,&
&MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
else
call MPI_REDUCE(g_tmp,g_dumm,numl*numl,MPI_DOUBLE_PRECISION,&
&MPI_SUM,iproc_time,world_comm,ierr)
endif
#else
re_g_mat_t(:,:,it+1)=g_tmp(:,:)
#endif
g_tmp(:,:)=0.d0
do iw=nbegin_g,nend_g
!factor=-2.d0*dimag(tf%weights_freq_g(iw)*exp((0.d0,1.d0)*tf%times(it)*tf%freqs_g_eff(iw)))/(2.d0*pi)
factor=-2.d0*dimag(tf%weights_freq_g(iw)*exp_table(it+1,iw-nbegin_g+1))/(2.d0*pi)
g_tmp(:,:)=g_tmp(:,:)+im_g_mat(:,:,iw-nbegin_g+1)*factor
enddo
#if defined(__MPI)
!the distribution of times on procs is the same of that for frequecies
iproc_time=it/l_blk
!all processors sums to iproc_time
if(iproc_time==mpime) then
call MPI_REDUCE(g_tmp,im_g_mat_t(1,1,it-nbegin+1),numl*numl,&
&MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
else
call MPI_REDUCE(g_tmp,g_dumm,numl*numl,MPI_DOUBLE_PRECISION,MPI_SUM,iproc_time,world_comm,ierr)
endif
#else
im_g_mat_t(:,:,it+1)=g_tmp(:,:)
#endif
enddo
deallocate(g_tmp,g_dumm)
else!for whole self-energy calculates directly in time domain
do it=nbegin,nend
call dgemm('T','N',numl,numl,numt,1.d0,ttl_j%tt_mat,numt,re_e_mat_part(1,1,it-nbegin+1),&
&numt,0.d0,re_g_mat_t(1,1,it-nbegin+1),numl)
call dgemm('T','N',numl,numl,numt,1.d0,ttl_j%tt_mat,numt,im_e_mat_part(1,1,it-nbegin+1),&
&numt,0.d0,im_g_mat_t(1,1,it-nbegin+1),numl)
enddo
endif
if( options%l_verbose) write(stdout,*) 'done'
FLUSH(stdout)
!loop on frequency
write(stdout,*) 'Products in imaginary time:'
FLUSH(stdout)
if( .not.(ss%whole_s .and.l_distribute_sm )) then
allocate(tmp_mat(numpw,numl),re_h_mat(numpw,numpw),im_h_mat(numpw,numpw))
do it=nbegin,nend
!matrix multiply
call dgemm('N','N',numpw,numl,numl,1.d0,vtl_j%vt_mat,numpw,re_g_mat_t(1,1,it-nbegin+1),numl,0.d0,tmp_mat,numpw)
call dgemm('N','T',numpw,numpw,numl,1.d0,tmp_mat,numpw, vtl%vt_mat,numpw,0.d0,re_h_mat,numpw)
call dgemm('N','N',numpw,numl,numl,1.d0,vtl_j%vt_mat,numpw,im_g_mat_t(1,1,it-nbegin+1),numl,0.d0,tmp_mat,numpw)
call dgemm('N','T',numpw,numpw,numl,1.d0,tmp_mat,numpw, vtl%vt_mat,numpw,0.d0,im_h_mat,numpw)
!product
if(ii==jj) then
ss%diag(ii,it+ss%n+1,is)=DDOT(numpw*numpw,re_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)+&
&DDOT(numpw*numpw,im_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)
ss%diag(ii,ss%n+1-it,is)=DDOT(numpw*numpw,re_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)-&
&DDOT(numpw*numpw,im_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)
!if required add semicore terms
if(options%l_semicore) then
!NOT_TO_BE_INCLUDED_START
allocate(tmp_vec_sc(numpw))
do iv_sc=1,sc%n_semicore
if(it==0) write(stdout,*) 'SEMICORE PRIMA',iv_sc,ii, ss%diag(ii,it+ss%n+1,is)
call dgemv('N',numpw,numpw,1.d0,pw_mat_t(1,1,it-nbegin+1),numpw,&
&sc%ppw_mat(1,iv_sc,ii),1,0.d0,tmp_vec_sc,1)
ss%diag(ii,it+ss%n+1,is)= ss%diag(ii,it+ss%n+1,is)- &!ATTENZIONE ERA +
&DDOT(numpw,tmp_vec_sc,1,sc%ppw_mat(1,iv_sc,ii),1)*exp((offset+sc%en_sc(iv_sc))*tf%times(it))
if(it==0) write(stdout,*) 'SEMICORE',iv_sc,ii, ss%diag(ii,it+ss%n+1,is)
if((offset+sc%en_sc(iv_sc))*tf%times(it)>0) write(stdout,*) 'OCIO!!!'!DEBUG ATTENZIONE
enddo
deallocate(tmp_vec_sc)
!NOT_TO_BE_INCLUDED_END
endif
endif
if(ss%whole_s) then
ss%whole(jj,ii,it+ss%n+1,is)=DDOT(numpw*numpw,re_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)+&
&DDOT(numpw*numpw,im_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)
ss%whole(jj,ii,ss%n+1-it,is)=DDOT(numpw*numpw,re_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)-&
&DDOT(numpw*numpw,im_h_mat,1,pw_mat_t(1,1,it-nbegin+1),1)
endif
enddo
deallocate(re_h_mat,im_h_mat)
deallocate(tmp_mat)
else
allocate(tmp_mat(numl,numl))
do it=nbegin,nend
call dgemm('T','N',numl,numl,numpw,1.d0,vtl_j%vt_mat,numpw,pw_part_t(1,1,it-nbegin+1),numpw,0.d0,tmp_mat,numl)
if(ii==jj) then
ss%diag(ii,it+ss%n+1,is)=DDOT(numl*numl,re_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)+&
&DDOT(numl*numl,im_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)
ss%diag(ii,ss%n+1-it,is)=DDOT(numl*numl,re_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)-&
&DDOT(numl*numl,im_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)
endif
ss%whole(jj,ii,it+ss%n+1,is)=DDOT(numl*numl,re_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)+&
&DDOT(numl*numl,im_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)
ss%whole(jj,ii,ss%n+1-it,is)=DDOT(numl*numl,re_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)-&
&DDOT(numl*numl,im_g_mat_t(1,1,it-nbegin+1),1,tmp_mat,1)
enddo
deallocate(tmp_mat)
endif
if( options%l_verbose) write(stdout,*) 'done'
FLUSH(stdout)
!mp_sum for distributing on all processors
if(ii==jj) then
call mp_sum(ss%diag(ii,1:2*ss%n+1,is),world_comm)
endif
if(ss%whole_s) then
call mp_sum(ss%whole(jj,ii,1:2*ss%n+1,is),world_comm)
endif
!FT of diagonal part of self-nergy
!deallocate(re_g_mat_t, im_g_mat_t)
call free_memory(vtl_j)
call free_memory(ttl_j)
enddo !jj
call free_memory(vtl)
call free_memory(ttl)
enddo!on i_min, i_max
enddo!on i_cycles
!put factor due to FFT on imaginary axis
ss%diag(:,:,is)=ss%diag(:,:,is)*(0.d0,1.d0)
if(ss%whole_s) then
ss%whole(:,:,:,is)=ss%whole(:,:,:,is)*(0.d0,1.d0)
endif
deallocate(re_g_mat,im_g_mat)
deallocate(re_g_mat_t,im_g_mat_t)
deallocate(pw_mat_t)
if(.not.l_single) then
deallocate(re_e_mat,im_e_mat)
else
deallocate(re_e_mat_single, im_e_mat_single, e_mat_double)
endif
if( ss%whole_s .and.l_distribute_sm ) then
deallocate(st_save,sl_save)
if(.not.l_single) deallocate(re_e_mat_t,im_e_mat_t)
if(.not.l_single) deallocate(re_e_mat_part,im_e_mat_part,pw_part_t)
endif
deallocate(exp_table)
call free_memory(lc)
enddo!on spin
call free_memory(uu)
call free_memory(sc)
return
end subroutine do_self_lanczos_time
subroutine solve_lanczos_single(alpha,re_e_mat,im_e_mat,lc)
!this subroutine sums to the matrix E_{no}=<t_n|(H-alpha)^-1|t_o>
USE kinds, ONLY : DP
USE basic_structures, ONLY : lanczos_chain, initialize_memory,free_memory
USE io_global, ONLY : stdout
USE mp, ONLY : mp_sum,mp_bcast
USE mp_world, ONLY : nproc, mpime, world_comm
implicit none
COMPLEX(kind=DP) :: alpha!constant for Ev+iw
TYPE(lanczos_chain) :: lc!lanczos chain descriptor
REAL(kind=4) :: re_e_mat(lc%numt,lc%numt) !real part of matrix to be calculated
REAL(kind=4) :: im_e_mat(lc%numt,lc%numt) !imaginary part of matrix to be calculated
INTEGER :: io,info,ii,jj
COMPLEX(kind=DP), ALLOCATABLE :: dl(:),du(:),d(:),t(:)
COMPLEX(kind=DP), ALLOCATABLE :: omat(:,:)
REAL(kind=DP), ALLOCATABLE :: tmp_mat(:,:)
INTEGER :: l_blk,nbegin,nend, iproc
COMPLEX(kind=DP), ALLOCATABLE :: e_mat(:)
l_blk= (lc%numt)/nproc
if(l_blk*nproc < (lc%numt)) l_blk = l_blk+1
nbegin=mpime*l_blk+1
nend=nbegin+l_blk-1
allocate(dl(lc%num_steps-1),du(lc%num_steps-1),d(lc%num_steps),t(lc%num_steps))
re_e_mat(:,:)=0.0
im_e_mat(:,:)=0.0
allocate(omat(lc%numt,lc%num_steps))
allocate(tmp_mat(lc%numt,lc%num_steps))
allocate(e_mat(lc%numt))
!loop on o
do io=1,lc%numt
!!set up vectors for lapack routine
!recover matrix from processor
tmp_mat(:,:)=0.d0
if(io >= nbegin .and. io <= nend) then
tmp_mat(1:lc%numt,1:lc%num_steps)=lc%o_mat(1:lc%numt,1:lc%num_steps,io-nbegin+1)
endif
iproc=(io-1)/l_blk
call mp_bcast(tmp_mat(:,:), iproc, world_comm)
omat(:,:)=dcmplx(tmp_mat(:,:),0.d0)
dl(1:lc%num_steps-1)=cmplx(lc%f(1:lc%num_steps-1,io),0.d0)
du(1:lc%num_steps-1)=cmplx(lc%f(1:lc%num_steps-1,io),0.d0)
d(1:lc%num_steps)=cmplx(lc%d(1:lc%num_steps,io),0.d0)+alpha
t(:)=(0.d0,0.d0)
t(1)=(1.d0,0.d0)
!!call lapack
call zgtsv(lc%num_steps,1,dl,d,du,t,lc%num_steps,info)
if(info /= 0) then
write(stdout,*) 'ZGTSV info:', info
FLUSH(stdout)
stop
endif
!!calculate term
call zgemm('N','N',lc%numt,1,lc%num_steps,(1.d0,0.d0),omat,lc%numt,t,lc%num_steps,(0.d0,0.d0),e_mat,lc%numt)
re_e_mat(1:lc%numt,io)=re_e_mat(1:lc%numt,io)+real(e_mat(1:lc%numt))
im_e_mat(1:lc%numt,io)=im_e_mat(1:lc%numt,io)+imag(e_mat(1:lc%numt))
enddo
do ii=1,lc%numt
do jj=ii+1,lc%numt
re_e_mat(ii,jj)=0.5*(re_e_mat(ii,jj)+re_e_mat(jj,ii))
re_e_mat(jj,ii)=re_e_mat(ii,jj)
im_e_mat(ii,jj)=0.5*(im_e_mat(ii,jj)+im_e_mat(jj,ii))
im_e_mat(jj,ii)=im_e_mat(ii,jj)
enddo
enddo
deallocate(dl,du,d,t)
deallocate(omat,tmp_mat)
deallocate(e_mat)
return
end subroutine solve_lanczos_single
subroutine solve_lanczos_fake_single(lc)
!this subroutine is a parallel fake routine for the solve lanczos routine
USE kinds, ONLY : DP
USE basic_structures, ONLY : lanczos_chain, initialize_memory,free_memory
USE io_global, ONLY : stdout
USE mp, ONLY : mp_sum,mp_bcast
USE mp_world, ONLY : nproc,mpime,world_comm
implicit none
TYPE(lanczos_chain) :: lc!lanczos chain descriptor
INTEGER :: l_blk,nbegin,nend, iproc
REAL(kind=DP), ALLOCATABLE :: o_mat(:,:)
INTEGER :: io
allocate(o_mat(lc%numt,lc%num_steps))
l_blk= (lc%numt)/nproc
if(l_blk*nproc < (lc%numt)) l_blk = l_blk+1
nbegin=mpime*l_blk+1
nend=nbegin+l_blk-1
!loop on io
do io=1,lc%numt
!recover matrix from processor
o_mat(:,:)=0.d0
if(io >= nbegin .and. io <= nend) then
o_mat(:,:)=lc%o_mat(:,:,io-nbegin+1)
endif
iproc=(io-1)/l_blk
call mp_bcast(o_mat(:,:), iproc, world_comm)
enddo
deallocate(o_mat)
end subroutine solve_lanczos_fake_single