LCOV - code coverage report
Current view: top level - src/67_common - m_dft_energy.F90 (source / functions) Coverage Total Hit
Test: coverage.info Lines: 74.1 % 405 300
Test Date: 2026-09-21 22:40:37 Functions: 100.0 % 3 3

            Line data    Source code
       1              : !!****m* ABINIT/m_dft_energy
       2              : !! NAME
       3              : !!  m_dft_energy
       4              : !!
       5              : !! FUNCTION
       6              : !!
       7              : !!
       8              : !! COPYRIGHT
       9              : !!  Copyright (C) 1998-2026 ABINIT group (DCA, XG, GMR, AR, MB, MT, EB)
      10              : !!  This file is distributed under the terms of the
      11              : !!  GNU General Public License, see ~abinit/COPYING
      12              : !!  or http://www.gnu.org/copyleft/gpl.txt .
      13              : !!
      14              : !! SOURCE
      15              : 
      16              : #if defined HAVE_CONFIG_H
      17              : #include "config.h"
      18              : #endif
      19              : 
      20              : #include "abi_common.h"
      21              : 
      22              : module m_dft_energy
      23              : 
      24              :  use defs_basis
      25              :  use defs_wvltypes
      26              :  use m_abicore
      27              :  use m_hamiltonian
      28              :  use m_errors
      29              :  use m_xmpi
      30              :  use m_xcdata
      31              :  use m_cgtools
      32              :  use m_dtset
      33              :  use m_extfpmd
      34              :  use m_ompgpu_utils
      35              : 
      36              :  use defs_datatypes, only : pseudopotential_type
      37              :  use defs_abitypes,      only : MPI_type
      38              :  use m_time,             only : timab
      39              :  use m_geometry,         only : metric
      40              :  use m_kg,               only : mkkin
      41              :  use m_energies,         only : energies_type
      42              :  use m_electronpositron, only : electronpositron_type, electronpositron_calctype, rhohxcpositron
      43              :  use m_bandfft_kpt,      only : bandfft_kpt, bandfft_kpt_type, prep_bandfft_tabs, &
      44              :                                 bandfft_kpt_savetabs, bandfft_kpt_restoretabs
      45              :  use m_pawang,           only : pawang_type
      46              :  use m_pawtab,           only : pawtab_type
      47              :  use m_paw_ij,           only : paw_ij_type
      48              :  use m_pawfgrtab,        only : pawfgrtab_type
      49              :  use m_pawrhoij,         only : pawrhoij_type, pawrhoij_alloc, pawrhoij_free, pawrhoij_init_unpacked, &
      50              :                                 pawrhoij_mpisum_unpacked, pawrhoij_free_unpacked, pawrhoij_inquire_dim, &
      51              : &                               pawrhoij_symrhoij
      52              :  use m_pawcprj,          only : pawcprj_type,pawcprj_alloc,pawcprj_free,pawcprj_gather_spin
      53              :  use m_pawfgr,           only : pawfgr_type
      54              :  use m_paw_dmft,         only : paw_dmft_type
      55              :  use m_paw_nhat,         only : pawmknhat
      56              :  use m_paw_mkrho,        only : pawmkrho
      57              :  use m_paw_occupancies,  only : pawaccrhoij
      58              :  use m_rcpaw,            only : rcpaw_type
      59              :  use m_fft,              only : fftpac, fourdp
      60              :  use m_spacepar,         only : meanvalue_g, hartre
      61              :  use m_dens,             only : constrained_dft_t,mag_penalty
      62              :  use m_mkrho,            only : mkrho
      63              :  use m_mkffnl,           only : mkffnl
      64              :  use m_getghc,           only : getghc
      65              :  use m_rhotoxc,          only : rhotoxc
      66              :  use m_mpinfo,           only : proc_distrb_cycle
      67              :  use m_nonlop,           only : nonlop
      68              :  use m_fourier_interpol, only : transgrid
      69              :  use m_prep_kgb,         only : prep_getghc, prep_nonlop
      70              :  use m_psolver,          only : psolver_rhohxc
      71              :  use m_gemm_nonlop_projectors, only : set_gemm_nonlop_ikpt, gemm_nonlop_use_gemm
      72              : 
      73              : #ifdef HAVE_FC_ISO_C_BINDING
      74              :  use, intrinsic :: iso_c_binding, only : c_int64_t
      75              : #endif
      76              : 
      77              : #if defined HAVE_GPU_CUDA
      78              :  use m_manage_cuda
      79              : #endif
      80              : 
      81              :  implicit none
      82              : 
      83              :  private
      84              : !!***
      85              : 
      86              :  public :: energy,entropy
      87              : !!***
      88              : 
      89              : contains
      90              : !!***
      91              : 
      92              : !!****f* ABINIT/energy
      93              : !! NAME
      94              : !!  energy
      95              : !!
      96              : !! FUNCTION
      97              : !!  Compute electronic energy terms
      98              : !!  energies%e_eigenvalues, ek and enl from arbitrary (orthonormal) provided wf,
      99              : !!  ehart, enxc, and eei from provided density and potential,
     100              : !!  energies%e_eigenvalues=Sum of the eigenvalues - Band energy (Hartree)
     101              : !!  energies%e_hspinfield=Spin magnetic energy from applied magnetic field -m.B
     102              : !!  ek=kinetic energy, ehart=Hartree electron-electron energy,
     103              : !!  enxc,enxcdc=exchange-correlation energies, eei=local pseudopotential energy,
     104              : !!  enl=nonlocal pseudopotential energy
     105              : !!  Also, compute new density from provided wfs, after the evaluation
     106              : !!  of ehart, enxc, and eei.
     107              : !!  WARNING XG180913 : At present, Fock energy not computed !
     108              : !!
     109              : !! NOTE that this routine is callned in m_scfcv_core only when nstep == 0
     110              : !!
     111              : !! INPUTS
     112              : !!  [add_tfw]=flag controling the addition of Weiszacker gradient correction to Thomas-Fermi kin energy
     113              : !!  cg(2,mpw*nspinor*mband*mkmem*nsppol)=<G|Cnk>=Fourier coefficients of wavefunction
     114              : !!  dtset <type(dataset_type)>=all input variables for this dataset
     115              : !!   | mband=maximum number of bands
     116              : !!   | mgfft=maximum size of 1D FFTs
     117              : !!   | mkmem=number of k points treated by this node.
     118              : !!   | mpw=maximum dimension for number of planewaves
     119              : !!   | natom=number of atoms in unit cell
     120              : !!   | nfft=(effective) number of FFT grid points (for this processor)
     121              : !!   | nkpt=number of k points
     122              : !!   | nspden=number of spin-density components
     123              : !!   | nsppol=1 for unpolarized, 2 for polarized
     124              : !!   | nspinor=number of spinorial components
     125              : !!   | nsym=number of symmetry elements in space group (at least 1)
     126              : !!   | occopt=option for occupancies
     127              : !!   | tsmear=smearing energy or temperature (if metal)
     128              : !!  eigen(mband*nkpt*nsppol)=array for holding eigenvalues (hartree)
     129              : !!  extfpmd <type(extfpmd_type)>=extended first-principles molecular dynamics type
     130              : !!  gsqcut=G^2 cutoff from gsqcut=ecut/(2 Pi^2)
     131              : !!  indsym(4,nsym,natom)=indirect indexing array for atom labels
     132              : !!  irrzon(nfft**(1-1/nsym),2,(nspden/nsppol)-3*(nspden/4))=irreducible zone data
     133              : !!  kg(3,mpw*mkmem)=work array for coordinates of G vectors in basis
     134              : !!  mcg=size of wave-functions array (cg) =mpw*nspinor*mband*mkmem*nsppol
     135              : !!  mpi_enreg=information about MPI parallelization
     136              : !!  my_natom=number of atoms treated by current processor
     137              : !!  nfftf= -PAW ONLY- number of FFT grid points for the fine grid
     138              : !!         (nfftf=nfft for norm-conserving potential runs)
     139              : !!  ngfftf(18)= -PAW ONLY- contain all needed information about 3D FFT for the fine grid
     140              : !!              (ngfftf=ngfft for norm-conserving potential runs)
     141              : !!  nhatgr(nfft,nspden,3*nhatgrdim)= -PAW only- cartesian gradients of compensation density
     142              : !!  nhatgrdim= -PAW only- 0 if nhatgr array is not used ; 1 otherwise
     143              : !!  npwarr(nkpt)=number of planewaves at each k point, and boundary
     144              : !!  n3xccc=dimension of the xccc3d array (0 or nfftf).
     145              : !!  occ(mband*nkpt*nsppol)=occupation numbers of bands (usually 2) at each k point
     146              : !!  optene=option for the computation of total energy (direct scheme or double-counting scheme)
     147              : !!  paw_ij(natom*usepaw) <type(paw_ij_type)>=paw arrays given on (i,j) channels
     148              : !!  pawang <type(pawang_type)>=paw angular mesh and related data
     149              : !!  pawfgr(natom) <type(pawfgr_type)>=fine grid parameters and related data
     150              : !!  pawfgrtab(natom) <type(pawfgrtab_type)>=atomic data given on fine rectangular grid
     151              : !!  pawtab(ntypat*usepaw) <type(pawtab_type)>=paw tabulated starting data
     152              : !!  phnons(2,nfft**(1-1/nsym),(nspden/nsppol)-3*(nspden/4))=nonsymmorphic translation phases
     153              : !!  ph1d(2,3*(2*mgfft+1)*natom)=phase information related to structure factor
     154              : !!  psps <type(pseudopotential_type)>=variables related to pseudopotentials
     155              : !!   | mpsang= 1+maximum angular momentum for nonlocal pseudopotentials
     156              : !!   | ntypat=number of types of atoms in cell
     157              : !!  rprimd(3,3)=dimensional real space primitive translations (bohr)
     158              : !!  symrec(3,3,nsym)=symmetry operations in reciprocal space
     159              : !!  usexcnhat= -PAW only- flag controling use of compensation density in Vxc
     160              : !!  vpsp(nfftf)=local pseudopotential in real space (hartree)
     161              : !!  wfs <type(wvl_projector_type)>=wavefunctions information for wavelets.
     162              : !!  wvl <type(wvl_internal_type)>=wavelets internal data
     163              : !!  xccc3d(n3xccc)=3D core electron density for XC core correction (bohr^-3)
     164              : !!  xcctau3d(n3xccc)=3D core electron kinetic energy density for XC core correction (bohr^-3)
     165              : !!  xred(3,natom)=reduced coordinates of atoms (dimensionless)
     166              : !!  ylm(mpw*mkmem,mpsang*mpsang*useylm)= real spherical harmonics for each G and k point
     167              : !!
     168              : !! OUTPUT
     169              : !!  compch_fft=-PAW only- compensation charge inside spheres computed over fine fft grid
     170              : !!  etotal=total energy (hartree):
     171              : !!    - computed by direct scheme if optene=0 or 2
     172              : !!    - computed by double-counting scheme if optene=1 or 3
     173              : !!  resid(mband*nkpt*nsppol)=residuals for each band over all k points (hartree^2)
     174              : !!  strsxc(6)=exchange-correlation contribution to stress tensor
     175              : !!  vhartr(nfftf)=work space to hold Hartree potential in real space (hartree)
     176              : !!  vtrial(nfftf,nspden)=total local potential (hartree)
     177              : !!  vxc(nfftf,nspden)=work space to hold Vxc(r) in real space (hartree)
     178              : !!  [vxctau(nfftf,nspden,4*usevxctau)]=(only for meta-GGA): derivative of XC energy density
     179              : !!    with respect to kinetic energy density (depsxcdtau). The arrays vxctau contains also
     180              : !!    the gradient of vxctau (gvxctau) in vxctau(:,:,2:4)
     181              : !!
     182              : !! SIDE EFFECTS
     183              : !!  electronpositron <type(electronpositron_type)>=quantities for the electron-positron annihilation (optional argument)
     184              : !!  energies <type(energies_type)>=all part of total energy.
     185              : !!   | entropy(IN)=entropy due to the occupation number smearing (if metal)
     186              : !!   | e_chempot(IN)=energy from spatially varying chemical potential (hartree)
     187              : !!   | e_ewald(IN)=Ewald energy (hartree)
     188              : !!   | e_vdw_dftd(IN)=VdW DFT-D energy
     189              : !!   | e_corepsp(IN)=psp core-core energy
     190              : !!   | paw%epaw(IN)=PAW spherical part energy
     191              : !!   | paw%epaw_dc(IN)=PAW spherical part double-counting energy
     192              : !!   | e_eigenvalues(OUT)=Sum of the eigenvalues - Band energy (Hartree)
     193              : !!   | e_hartree(OUT)=Hartree part of total energy (hartree units)
     194              : !!   | e_kinetic(OUT)=kinetic energy part of total energy.
     195              : !!   | e_nlpsp_vfock(OUT)=nonlocal psp + potential Fock ACE part of total energy.
     196              : !!   | e_xc(OUT)=exchange-correlation energy (hartree)
     197              : !!  ==== if optene==0, 2 or 3
     198              : !!   | e_localpsp(OUT)=local psp energy (hartree)
     199              : !!  ==== if optene==1, 2 or 3
     200              : !!   | e_xcdc(OUT)=exchange-correlation double-counting energy (hartree)
     201              : !!  rhog(2,nfftf)=work space for rho(G); save intact on return (? MT 08-12-2008: is that true now ?)
     202              : !!  rhor(nfftf,nspden)=work space for rho(r); save intact on return (? MT 08-12-2008: is that true now ?)
     203              : !!  taug(2,nfftf*dtset%usekden)=array for Fourier transform of kinetic energy density
     204              : !!  taur(nfftf,nspden*dtset%usekden)=array for kinetic energy density
     205              : !!  nspinor should not be modified in the call of rdnpw
     206              : !!  === if psps%usepaw==1 ===
     207              : !!    nhat(nfftf,nspden*usepaw)= compensation charge density
     208              : !!    pawrhoij(natom) <type(pawrhoij_type)>= paw rhoij occupancies and related data
     209              : !!
     210              : !! NOTES
     211              : !!  Be careful to the meaning of nfft (size of FFT grids):
     212              : !!   - In case of norm-conserving calculations the FFT grid is the usual FFT grid.
     213              : !!   - In case of PAW calculations:
     214              : !!     Two FFT grids are used; one with nfft points (coarse grid) for
     215              : !!     the computation of wave functions ; one with nfftf points
     216              : !!     (fine grid) for the computation of total density.
     217              : !!
     218              : !!  There is a large amount of overhead in the way this routine do the computation of the energy !
     219              : !!  For example, the density has already been precomputed, so why to compute it again here ??
     220              : !!
     221              : !! SOURCE
     222              : 
     223           70 : subroutine energy(cg,compch_fft,constrained_dft,dtset,electronpositron,&
     224           35 : & energies,eigen,etotal,gsqcut,extfpmd,indsym,irrzon,kg,mcg,mpi_enreg,my_natom,nfftf,ngfftf,nhat,&
     225           35 : & nhatgr,nhatgrdim,npwarr,n3xccc,occ,optene,paw_dmft,paw_ij,pawang,pawfgr,&
     226           35 : & pawfgrtab,pawrhoij,pawtab,phnons,ph1d,psps,resid,rhog,rhor,rprimd,strsxc,symrec,&
     227           35 : & taug,taur,usexcnhat,vhartr,vtrial,vpsp,vxc,wfs,wvl,wvl_den,wvl_e,xccc3d,xred,ylm,&
     228           35 : & add_tfw,vxctau,xcctau3d,rcpaw) ! optional argument
     229              : 
     230              : !Arguments ------------------------------------
     231              : !scalars
     232              :  integer,intent(in) :: mcg,my_natom,n3xccc,nfftf,nhatgrdim,optene,usexcnhat
     233              :  logical,intent(in),optional :: add_tfw
     234              :  real(dp),intent(in) :: gsqcut
     235              :  real(dp),intent(out) :: compch_fft,etotal
     236              :  type(MPI_type),intent(inout) :: mpi_enreg
     237              :  type(constrained_dft_t),intent(in) :: constrained_dft
     238              :  type(dataset_type),intent(in) :: dtset
     239              :  type(electronpositron_type),pointer :: electronpositron
     240              :  type(energies_type),intent(inout) :: energies
     241              :  type(extfpmd_type),pointer,intent(inout) :: extfpmd
     242              :  type(paw_dmft_type), intent(inout) :: paw_dmft
     243              :  type(pawang_type),intent(in) :: pawang
     244              :  type(pawfgr_type),intent(in) :: pawfgr
     245              :  type(pseudopotential_type),intent(in) :: psps
     246              :  type(wvl_internal_type), intent(in) :: wvl
     247              :  type(wvl_wf_type),intent(inout) :: wfs
     248              :  type(wvl_denspot_type), intent(inout) :: wvl_den
     249              :  type(wvl_energy_terms),intent(inout) ::wvl_e
     250              :  type(rcpaw_type),pointer,intent(in),optional :: rcpaw
     251              : !arrays
     252              : !nfft**(1-1/nsym) is 1 if nsym==1, and nfft otherwise
     253              :  integer, intent(in) :: indsym(4,dtset%nsym,dtset%natom)
     254              :  integer :: irrzon(dtset%nfft**(1-1/dtset%nsym),2,(dtset%nspden/dtset%nsppol)-3*(dtset%nspden/4)),kg(3,dtset%mpw*dtset%mkmem)
     255              :  integer, intent(in) :: ngfftf(18),npwarr(dtset%nkpt),symrec(3,3,dtset%nsym)
     256              :  real(dp), intent(in) :: cg(2,mcg),eigen(dtset%mband*dtset%nkpt*dtset%nsppol)
     257              :  real(dp), intent(in) :: occ(dtset%mband*dtset%nkpt*dtset%nsppol),ph1d(2,3*(2*dtset%mgfft+1)*dtset%natom)
     258              :  real(dp), intent(inout) :: nhat(nfftf,dtset%nspden*psps%usepaw)
     259              :  real(dp),intent(in) :: nhatgr(nfftf,dtset%nspden,3*nhatgrdim)
     260              : !nfft**(1-1/nsym) is 1 if nsym==1, and nfft otherwise
     261              :  real(dp), intent(in) :: phnons(2,dtset%nfft**(1-1/dtset%nsym),(dtset%nspden/dtset%nsppol)-3*(dtset%nspden/4))
     262              :  real(dp), intent(inout) :: resid(dtset%mband*dtset%nkpt*dtset%nsppol)
     263              :  real(dp), intent(inout) :: rhog(2,nfftf),rhor(nfftf,dtset%nspden)
     264              :  real(dp), intent(inout) :: taug(2,nfftf*dtset%usekden),taur(nfftf,dtset%nspden*dtset%usekden)
     265              :  real(dp), intent(out) :: strsxc(6)
     266              :  real(dp), intent(in) :: rprimd(3,3),vpsp(nfftf),xccc3d(n3xccc),xred(3,dtset%natom)
     267              :  real(dp), intent(in) :: xcctau3d(nfftf*dtset%usekden)
     268              :  real(dp), intent(out) :: vhartr(nfftf),vtrial(nfftf,dtset%nspden),vxc(nfftf,dtset%nspden)
     269              :  real(dp),intent(out),optional,target :: vxctau(:,:,:) ! vxctau(nfftf,dtset%nspden,4*usevxctau)
     270              :  real(dp), intent(in) :: ylm(dtset%mpw*dtset%mkmem,psps%mpsang*psps%mpsang*psps%useylm)
     271              :  type(paw_ij_type), intent(in) :: paw_ij(my_natom*psps%usepaw)
     272              :  type(pawfgrtab_type),intent(inout) :: pawfgrtab(my_natom)
     273              :  type(pawrhoij_type),target,intent(inout) :: pawrhoij(my_natom*psps%usepaw)
     274              :  type(pawtab_type),intent(in) :: pawtab(psps%ntypat*psps%usepaw)
     275              : 
     276              : !Local variables-------------------------------
     277              : !scalars
     278              :  integer :: bdtot_index,blocksize,choice,cplex,cplex_rhoij,cpopt,dimffnl
     279              :  integer :: iband,iband_last,iblock,iblocksize,icg,ider,idir,ierr,ifft,ikg,ikpt,ilm
     280              :  integer :: ipert,ipositron,iresid,ispden,isppol,istwf_k,itypat,izero
     281              :  integer :: me_distrb,mpi_comm_sphgrid,my_ikpt,my_nspinor,n1,n2,n3,n4,n5,n6
     282              :  integer :: nband_k,nblockbd,nfftotf,nkpg,nkxc,nk3xc,nnlout,npw_k,nspden_rhoij,option
     283              :  integer :: option_rhoij,paw_opt,signs,spaceComm,tim_mkrho,tim_nonlop
     284              :  logical :: add_tfw_,paral_atom,use_timerev,use_zeromag,with_vxctau
     285              :  logical :: non_magnetic_xc,wvlbigdft=.false.
     286              :  real(dp) :: dotr,doti,eeigk,ekk,enlk,evxc,e_xcdc_vxctau,ucvol,ucvol_local,vxcavg
     287              :  !character(len=500) :: message
     288           35 :  type(gs_hamiltonian_type) :: gs_hamk
     289              :  type(xcdata_type) :: xcdata
     290              : !arrays
     291           35 :  integer,allocatable :: kg_k(:,:)
     292              :  real(dp) :: gmet(3,3),gprimd(3,3),kpg_dum(0,0),kpoint(3),nonlop_out(1,1)
     293              :  real(dp) :: qpt(3),rhodum(1),rmet(3,3),tsec(2),ylmgr_dum(1,1,1),vhspinfield(4)
     294           70 :  real(dp) :: magvec(dtset%nspden)
     295              :  real(dp),target :: vxctau_dum(0,0,0)
     296           35 :  real(dp),allocatable :: buffer(:)
     297           35 :  real(dp),allocatable :: cwavef(:,:),eig_k(:),enlout(:),ffnl(:,:,:,:),ffnl_sav(:,:,:,:)
     298           35 :  real(dp),allocatable :: kinpw(:),kinpw_sav(:),kxc(:,:),occ_k(:),occblock(:)
     299           35 :  real(dp),allocatable :: ph3d(:,:,:),ph3d_sav(:,:,:)
     300           35 :  real(dp),allocatable :: resid_k(:),rhowfg(:,:),rhowfr(:,:),vlocal(:,:,:,:)
     301           35 :  real(dp),allocatable :: vxctaulocal(:,:,:,:,:),ylm_k(:,:),v_constr_dft_r(:,:)
     302           35 :  real(dp),pointer :: vxctau_(:,:,:)
     303              :  type(bandfft_kpt_type),pointer :: my_bandfft_kpt => null()
     304           35 :  type(pawcprj_type),target,allocatable :: cwaveprj(:,:)
     305           35 :  type(pawcprj_type),pointer :: cwaveprj_gat(:,:)
     306           35 :  type(pawrhoij_type),pointer :: pawrhoij_unsym(:)
     307              : ! *************************************************************************
     308              : 
     309              :  DBG_ENTER("COLL")
     310              : 
     311              : !Test size of kinetic energy potential Vxctau
     312           35 :  with_vxctau = (present(vxctau))
     313          140 :  if (with_vxctau) with_vxctau = (size(vxctau)>0.and.dtset%usekden/=0)
     314              :  if (with_vxctau) then
     315            0 :    if (size(vxctau)/=nfftf*dtset%nspden*4) then
     316            0 :      ABI_BUG("Wrong size for vxctau!")
     317              :    end if
     318              :  end if
     319           35 :  vxctau_ => vxctau_dum ; if (with_vxctau) vxctau_ => vxctau
     320              : 
     321              : !Test size of FFT grids (1 grid in norm-conserving, 2 grids in PAW)
     322          140 :  nfftotf=PRODUCT(ngfftf(1:3))
     323           35 :  if ((psps%usepaw==1.and.pawfgr%nfft/=nfftf).or.(psps%usepaw==0.and.dtset%nfft/=nfftf)) then
     324            0 :    ABI_BUG('wrong values for nfft, nfftf!')
     325              :  end if
     326              : 
     327              : !If usewvl: wvlbigdft indicates that the BigDFT workflow will be followed
     328           35 :  wvlbigdft=(dtset%usewvl==1 .and. dtset%wvl_bigdft_comp==1)
     329              : 
     330           35 :  call timab(59,1,tsec)
     331              : 
     332              : !Data for parallelism
     333           35 :  spaceComm=mpi_enreg%comm_cell
     334           35 :  if(mpi_enreg%paral_kgb==1) spaceComm=mpi_enreg%comm_kpt
     335           35 :  if(mpi_enreg%paral_hf==1) spaceComm=mpi_enreg%comm_kpt
     336           35 :  mpi_comm_sphgrid=mpi_enreg%comm_fft
     337           35 :  if(dtset%usewvl==1) then
     338            0 :    spaceComm=mpi_enreg%comm_wvl
     339            0 :    mpi_comm_sphgrid=mpi_enreg%comm_wvl
     340              :  end if
     341           35 :  me_distrb=mpi_enreg%me_kpt
     342           35 :  my_nspinor=max(1,dtset%nspinor/mpi_enreg%nproc_spinor)
     343           35 :  paral_atom=(my_natom/=dtset%natom)
     344              : 
     345              : !Compute gmet, gprimd and ucvol from rprimd
     346           35 :  call metric(gmet,gprimd,-1,rmet,rprimd,ucvol)
     347           35 :  if (dtset%usewvl == 0) then
     348           35 :    ucvol_local = ucvol
     349              : #if defined HAVE_BIGDFT
     350              :  else
     351              : !  We need to tune the volume when wavelets are used because, not
     352              : !  all FFT points are used.
     353              : !  ucvol_local = (half * dtset%wvl_hgrid) ** 3 * ngfft(1)*ngfft(2)*ngfft(3)
     354              :    ucvol_local = product(wvl_den%denspot%dpbox%hgrids) * real(product(wvl_den%denspot%dpbox%ndims), dp)
     355              : #endif
     356              :  end if
     357              : 
     358              : !Compute Hxc potential from density
     359           35 :  option=1;nkxc=0
     360           35 :  ipositron=electronpositron_calctype(electronpositron)
     361           35 :  add_tfw_=.false.;if (present(add_tfw)) add_tfw_=add_tfw
     362           35 :  non_magnetic_xc=(dtset%usepaw==1.and.mod(abs(dtset%usepawu),10)==4)
     363           35 :  if (ipositron/=1) then
     364              : 
     365           35 :    if (dtset%icoulomb == 0) then
     366              : !    Use the periodic solver to compute Hxc.
     367              :      call hartre(1,gsqcut,dtset%icutcoul,psps%usepaw,mpi_enreg,nfftf,ngfftf,&
     368           35 :                  &dtset%nkpt,dtset%rcut,rhog,rprimd,dtset%vcutgeo,vhartr)
     369           35 :      call xcdata_init(xcdata,dtset=dtset)
     370           35 :      ABI_MALLOC(kxc,(1,nkxc))
     371              : !    to be adjusted for the call to rhotoxc
     372           35 :      nk3xc=1
     373           35 :      if (ipositron==0) then
     374              :        call rhotoxc(energies%e_xc,energies%entropy_xc,kxc, &
     375              : &       mpi_enreg,nfftf,ngfftf,nhat,psps%usepaw,nhatgr,nhatgrdim, &
     376              : &       nkxc,nk3xc,non_magnetic_xc,n3xccc,option,rhor,rprimd, &
     377              : &       usexcnhat,vxc,vxcavg,xccc3d,xcdata,taur=taur,vhartr=vhartr, &
     378              : &       vxctau=vxctau_,exc_vdw_out=energies%e_xc_vdw,add_tfw=add_tfw_, &
     379           35 : &       xcctau3d=xcctau3d,strsxc=strsxc)
     380              :      else
     381              :        call rhotoxc(energies%e_xc,energies%entropy_xc,kxc, &
     382              : &       mpi_enreg,nfftf,ngfftf,nhat,psps%usepaw,nhatgr,nhatgrdim, &
     383              : &       nkxc,nk3xc,non_magnetic_xc,n3xccc,option,rhor,rprimd, &
     384              : &       usexcnhat,vxc,vxcavg,xccc3d,xcdata, &
     385              : &       electronpositron=electronpositron,taur=taur,vhartr=vhartr, &
     386              : &       vxctau=vxctau_,exc_vdw_out=energies%e_xc_vdw,add_tfw=add_tfw_, &
     387            0 : &       xcctau3d=xcctau3d,strsxc=strsxc)
     388              :      end if
     389           70 :      ABI_FREE(kxc)
     390            0 :    else if (dtset%usewvl == 0) then
     391              : !    Use the free boundary solver.
     392              :      call psolver_rhohxc(energies%e_hartree, energies%e_xc, evxc, &
     393              : &     dtset%icoulomb, dtset%ixc, mpi_enreg, nfftf, &
     394              : &     ngfftf,nhat,psps%usepaw,&
     395              : &     dtset%nscforder,dtset%nspden,n3xccc,rhor,rprimd, &
     396              : &     usexcnhat,psps%usepaw,dtset%usewvl,&
     397              : &     vhartr, vxc, vxcavg, wvl,wvl_den,wvl_e,&
     398            0 : &     xccc3d,dtset%xclevel,dtset%xc_denpos)
     399              :    end if
     400              :  else
     401            0 :    energies%e_xc=zero
     402              :    call rhohxcpositron(electronpositron,gprimd,kxc,mpi_enreg,nfftf,ngfftf,nhat,nkxc,dtset%nspden,n3xccc,&
     403            0 : &   dtset%paral_kgb,rhor,strsxc,ucvol,usexcnhat,psps%usepaw,vhartr,vxc,vxcavg,xccc3d,dtset%xc_denpos)
     404              :  end if
     405           35 :  if (ipositron/=0) then
     406              :    call dotprod_vn(1,rhor,electronpositron%e_hartree,doti,nfftf,nfftotf,1,1,electronpositron%vha_ep,&
     407            0 : &   ucvol,mpi_comm_sphgrid=mpi_comm_sphgrid)
     408            0 :    vhartr=vhartr+electronpositron%vha_ep
     409              :  end if
     410              : 
     411              : !Total local potential (for either spin channel) is
     412              : !Hartree + local psp + Vxc(spin), minus its mean
     413              : !(Note : this potential should agree with the input vtrial)
     414           70 :  do ispden=1,min(dtset%nspden,2)
     415       535080 :    do ifft=1,nfftf
     416       535045 :      vtrial(ifft,ispden)=vhartr(ifft)+vpsp(ifft)+vxc(ifft,ispden)
     417              :    end do
     418              :  end do
     419           35 :  if (dtset%nspden==4) then
     420            0 :    do ifft=1,nfftf
     421            0 :      vtrial(ifft,3:4)=vxc(ifft,3:4)
     422              :    end do
     423              :  end if
     424              : 
     425              : !Add the vhspinfield pot in the trial pot
     426              : !Vhspinfield might have to be allocated correctly --> to be checked
     427          140 :  if (any(abs(dtset%hspinfield(:))>tol8)) then
     428            0 :    vhspinfield(:) = zero
     429            0 :    if(dtset%nspden==2)then
     430              : !TODO: check this against rhotov and setvtr, where the potential is -1/2 and +1/2 for the 2 spin components.
     431              : ! see comment by SPr in rhotov
     432              : ! TODO: check this 1/2 factor is for the electron spin magnetic moment.
     433            0 :      vhspinfield(1) = -half*dtset%hspinfield(3) ! For collinear ispden=1 potential is v_upup
     434            0 :      vhspinfield(2) = +half*dtset%hspinfield(3) ! For collinear ispden=2 potential is v_dndn
     435              :    end if
     436            0 :    if(dtset%nspden==4)then
     437            0 :      vhspinfield(1)=-half*dtset%hspinfield(3)
     438            0 :      vhspinfield(2)= half*dtset%hspinfield(3)
     439            0 :      vhspinfield(3)=-half*dtset%hspinfield(1)
     440            0 :      vhspinfield(4)= half*dtset%hspinfield(2)
     441              :    end if
     442            0 :    magvec = zero
     443            0 :    do ispden=1,dtset%nspden
     444            0 :      do ifft=1,nfftf
     445              : !TODO: the full cell magnetization will need extra PAW terms, and is certainly calculated elsewhere.
     446              : !The calculation of the spin magnetic energy can be moved there
     447            0 :        magvec(ispden) = magvec(ispden) + rhor(ifft,ispden)
     448            0 :        vtrial(ifft,ispden)=vtrial(ifft,ispden)+vhspinfield(ispden)
     449              :      end do
     450              :    end do
     451            0 :    if(dtset%nspden==2)then
     452            0 :      energies%e_hspinfield = -half*dtset%hspinfield(3)*(two*magvec(2)-magvec(1)) !  diff rho = rhoup-rhodown = 2 rhoup - rho
     453            0 :    else if(dtset%nspden==4)then
     454              :      energies%e_hspinfield = -half * (dtset%hspinfield(1)*magvec(2)& ! x
     455              : &                                +dtset%hspinfield(2)*magvec(3)& ! y
     456            0 : &                                +dtset%hspinfield(3)*magvec(4)) ! z
     457              :    end if
     458              :  end if
     459              : 
     460              : !Compute the constrained potential for the magnetic moments
     461              : !NOTE: here in energy.F90 rhor and vtrial are given on nfftf grid
     462              : !the values coming from mag_penalty may be different from those calculated
     463              : !calling mag_penalty with nfft in setvtr and rhotov
     464           35 :  if (dtset%magconon==1.or.dtset%magconon==2) then
     465            0 :    ABI_MALLOC(v_constr_dft_r, (nfftf,dtset%nspden))
     466            0 :    v_constr_dft_r = zero
     467            0 :    call mag_penalty(constrained_dft,mpi_enreg,rhor,v_constr_dft_r,xred,dtset%qgbt,dtset%use_gbt)
     468              : !   call mag_penalty(dtset%natom, dtset%spinat, dtset%nspden, dtset%magconon, dtset%magcon_lambda, rprimd, &
     469              : !&   mpi_enreg, nfftf, dtset%ngfft, dtset%ntypat, dtset%ratsph, rhor, &
     470              : !&   dtset%typat, v_constr_dft_r, xred)
     471            0 :    do ispden=1,dtset%nspden
     472            0 :      do ifft=1,nfftf
     473            0 :        vtrial(ifft,ispden)=vtrial(ifft,ispden)+v_constr_dft_r(ifft,ispden)
     474              :      end do
     475              :    end do
     476            0 :    ABI_FREE(v_constr_dft_r)
     477              :  end if
     478              : 
     479              : !Compute Hartree energy - use up+down rhor
     480           35 :  if (ipositron/=1) then
     481              :    call dotprod_vn(1,rhor,energies%e_hartree ,doti,nfftf,nfftotf,1,1,vhartr,&
     482           35 : &   ucvol_local,mpi_comm_sphgrid=mpi_comm_sphgrid)
     483           35 :    if (ipositron==0) energies%e_hartree=half*energies%e_hartree
     484           35 :    if (ipositron==2) energies%e_hartree = half *(energies%e_hartree-electronpositron%e_hartree)
     485              :  else
     486            0 :    energies%e_hartree=zero
     487              :  end if
     488              : 
     489              : !Compute local psp energy - use up+down rhor
     490           35 :  if (optene/=1) then
     491              :    call dotprod_vn(1,rhor,energies%e_localpsp,doti,nfftf,nfftotf,1,1,vpsp,&
     492           35 : &   ucvol_local,mpi_comm_sphgrid=mpi_comm_sphgrid)
     493              :  end if
     494              : 
     495              : !Compute DC-xc energy - use up+down rhor
     496           35 :  if (optene>0) then
     497            3 :    if (ipositron/=1) then
     498            3 :      if (psps%usepaw==0.or.usexcnhat/=0) then
     499              :        call dotprod_vn(1,rhor,energies%e_xcdc,doti,nfftf,nfftotf,dtset%nspden,1,vxc,&
     500            1 : &       ucvol_local,mpi_comm_sphgrid=mpi_comm_sphgrid)
     501            1 :        if (with_vxctau)then
     502              :          call dotprod_vn(1,taur,e_xcdc_vxctau,doti,nfftf,nfftotf,dtset%nspden,1,vxctau(:,:,1),&
     503            0 : &         ucvol_local,mpi_comm_sphgrid=mpi_comm_sphgrid)
     504            0 :          energies%e_xcdc=energies%e_xcdc+e_xcdc_vxctau
     505              :        end if
     506              :      else
     507            8 :        ABI_MALLOC(rhowfr,(nfftf,dtset%nspden))
     508        54006 :        rhowfr=rhor-nhat
     509              :        call dotprod_vn(1,rhowfr,energies%e_xcdc,doti,nfftf,nfftotf,dtset%nspden,1,vxc,&
     510            2 : &       ucvol_local,mpi_comm_sphgrid=mpi_comm_sphgrid)
     511            2 :        ABI_FREE(rhowfr)
     512              :      end if
     513            3 :      if (ipositron==2) energies%e_xcdc=energies%e_xcdc-electronpositron%e_xcdc
     514              :    else
     515            0 :      energies%e_xcdc=zero
     516              :    end if
     517              :  end if
     518              : 
     519           35 :  energies%e_eigenvalues=zero
     520           35 :  energies%e_kinetic=zero
     521           35 :  energies%e_nlpsp_vfock=zero
     522           35 :  energies%e_fock0=zero
     523           35 :  bdtot_index=0
     524           35 :  icg=0
     525              : 
     526           35 :  n1=dtset%ngfft(1) ; n2=dtset%ngfft(2) ; n3=dtset%ngfft(3)
     527           35 :  n4=dtset%ngfft(4) ; n5=dtset%ngfft(5) ; n6=dtset%ngfft(6)
     528              : 
     529              : !============================================
     530              : !==== Initialize most of the Hamiltonian ====
     531              : !============================================
     532              : !1) Allocate all arrays and initialize quantities that do not depend on k and spin.
     533              : !2) Perform the setup needed for the non-local factors:
     534              : !* Norm-conserving: Constant kleimann-Bylander energies are copied from psps to gs_hamk.
     535              : !* PAW: Initialize the overlap coefficients and allocate the Dij coefficients.
     536              : 
     537              :  call gs_hamk%init(psps,pawtab,dtset%nspinor,dtset%nsppol,dtset%nspden,&
     538              : & dtset%natom,dtset%typat,xred,dtset%nfft,dtset%mgfft,dtset%ngfft,rprimd,dtset%nloalg,&
     539              : & comm_atom=mpi_enreg%comm_atom,mpi_atmtab=mpi_enreg%my_atmtab,mpi_spintab=mpi_enreg%my_isppoltab,&
     540              : & paw_ij=paw_ij,ph1d=ph1d,electronpositron=electronpositron,&
     541           35 : & nucdipmom=dtset%nucdipmom,gpu_option=dtset%gpu_option)
     542              : 
     543          210 :  ABI_MALLOC(vlocal,(n4,n5,n6,gs_hamk%nvloc))
     544           35 :  if (with_vxctau) then
     545            0 :    ABI_MALLOC(vxctaulocal,(n4,n5,n6,gs_hamk%nvloc,4))
     546              :  end if
     547              : 
     548              : !PAW: additional initializations
     549           35 :  if (psps%usepaw==1) then
     550           59 :    ABI_MALLOC(cwaveprj,(dtset%natom,my_nspinor))
     551            3 :    call pawcprj_alloc(cwaveprj,0,gs_hamk%dimcprj)
     552            3 :    if (mpi_enreg%paral_spinor==1) then
     553            0 :      ABI_MALLOC(cwaveprj_gat,(dtset%natom,dtset%nspinor))
     554            0 :      call pawcprj_alloc(cwaveprj_gat,0,gs_hamk%dimcprj)
     555              :    else
     556            3 :      cwaveprj_gat => cwaveprj
     557              :    end if
     558            3 :    if (paral_atom) then
     559            0 :      ABI_MALLOC(pawrhoij_unsym,(dtset%natom))
     560              :      call pawrhoij_inquire_dim(cplex_rhoij=cplex_rhoij,nspden_rhoij=nspden_rhoij,&
     561            0 : &                nspden=dtset%nspden,spnorb=dtset%pawspnorb,cpxocc=dtset%pawcpxocc)
     562              :      call pawrhoij_alloc(pawrhoij_unsym,cplex_rhoij,nspden_rhoij,dtset%nspinor,&
     563            0 : &     dtset%nsppol,dtset%typat,pawtab=pawtab,use_rhoijp=0,use_rhoij_=1)
     564              :    else
     565            3 :      pawrhoij_unsym => pawrhoij
     566            3 :      call pawrhoij_init_unpacked(pawrhoij_unsym)
     567              :    end if
     568            3 :    option_rhoij=1
     569            3 :    use_timerev=(dtset%kptopt>0.and.dtset%kptopt<3)
     570            6 :    use_zeromag=(pawrhoij_unsym(1)%nspden==4.and.dtset%nspden==1)
     571              :  else
     572           32 :    ABI_MALLOC(cwaveprj,(0,0))
     573              :  end if
     574              : 
     575              : !LOOP OVER SPINS
     576           70 :  do isppol=1,dtset%nsppol
     577           35 :    ikg=0
     578              : 
     579              :    ! Set up local potential vlocal on the coarse FFT mesh from vtrial taking into account the spin.
     580              :    ! Also take into account the spin.
     581              : 
     582              :    call gspot_transgrid_and_pack(isppol, psps%usepaw, dtset%paral_kgb, dtset%nfft, dtset%ngfft, nfftf, &
     583           35 :                                  dtset%nspden, gs_hamk%nvloc, 1, pawfgr, mpi_enreg, vtrial, vlocal)
     584           35 :    call gs_hamk%load_spin(isppol,vlocal=vlocal,with_nonlocal=.true.)
     585              : 
     586           35 :    if (with_vxctau) then
     587              :      call gspot_transgrid_and_pack(isppol, psps%usepaw, dtset%paral_kgb, dtset%nfft, dtset%ngfft, nfftf, &
     588            0 :                                    dtset%nspden, gs_hamk%nvloc, 4, pawfgr, mpi_enreg, vxctau, vxctaulocal)
     589            0 :      call gs_hamk%load_spin(isppol, vxctaulocal=vxctaulocal)
     590              :    end if
     591              : 
     592              : !  Loop over k points
     593          280 :    do ikpt=1,dtset%nkpt
     594          210 :      nband_k=dtset%nband(ikpt+(isppol-1)*dtset%nkpt)
     595          210 :      istwf_k=dtset%istwfk(ikpt)
     596          210 :      npw_k=npwarr(ikpt)
     597              : 
     598              : !    Skip this k-point if not the proper processor
     599          210 :      if(proc_distrb_cycle(mpi_enreg%proc_distrb,ikpt,1,nband_k,isppol,me_distrb)) then
     600            0 :        resid(1+bdtot_index : nband_k+bdtot_index) = zero
     601              :        bdtot_index=bdtot_index+nband_k
     602              :        cycle
     603              :      end if
     604              : 
     605              : !    Parallelism over FFT and/or bands: define sizes and tabs
     606          210 :      if (mpi_enreg%paral_kgb==1) then
     607            0 :        my_ikpt=mpi_enreg%my_kpttab(ikpt)
     608            0 :        nblockbd=nband_k/(mpi_enreg%nproc_band*mpi_enreg%bandpp)
     609            0 :        my_bandfft_kpt => bandfft_kpt(my_ikpt)
     610              :      else
     611          210 :        my_ikpt=ikpt
     612          210 :        nblockbd=nband_k
     613              :        !if (nband_k/=nblockbd*mpi_enreg%nproc_fft) nblockbd=nblockbd+1
     614              :      end if
     615          210 :      blocksize=nband_k/nblockbd
     616              : 
     617          630 :      ABI_MALLOC(eig_k,(nband_k))
     618          420 :      ABI_MALLOC(occ_k,(nband_k))
     619          420 :      ABI_MALLOC(resid_k,(nband_k))
     620          630 :      ABI_MALLOC(cwavef,(2,npw_k*my_nspinor*blocksize))
     621         2198 :      resid_k(:)=zero
     622          840 :      kpoint(:)=dtset%kptns(:,ikpt)
     623         2198 :      occ_k(:)=occ(1+bdtot_index:nband_k+bdtot_index)
     624         2198 :      eig_k(:)=eigen(1+bdtot_index:nband_k+bdtot_index)
     625         2577 :      if (minval(eig_k)>1.d100) eig_k=zero
     626          210 :      eeigk=zero ; ekk=zero ; enlk=zero
     627              : 
     628          630 :      ABI_MALLOC(kg_k,(3,npw_k))
     629       346246 :      kg_k(:,1:npw_k)=kg(:,1+ikg:npw_k+ikg)
     630              : 
     631          840 :      ABI_MALLOC(ylm_k,(npw_k,psps%mpsang*psps%mpsang*psps%useylm))
     632          210 :      if (psps%useylm==1) then
     633          370 :        do ilm=1,psps%mpsang*psps%mpsang
     634        61934 :          ylm_k(1:npw_k,ilm)=ylm(1+ikg:npw_k+ikg,ilm)
     635              :        end do
     636              :      end if
     637              : 
     638              : !    Compute kinetic energy
     639          630 :      ABI_MALLOC(kinpw,(npw_k))
     640          210 :      call mkkin(dtset%ecut,dtset%ecutsm,dtset%effmass_free,gmet,kg_k,kinpw,kpoint,npw_k,0,0)
     641              : 
     642              : !    Compute kinetic energy of each band
     643         2198 :      do iblock=1,nblockbd
     644         4186 :        do iblocksize=1,blocksize
     645         1988 :          iband=(iblock-1)*blocksize+iblocksize
     646         3976 :          if (abs(occ_k(iband))>tol8) then
     647              :            cwavef(1:2,1:npw_k*my_nspinor)= &
     648      2115366 : &           cg(:,1+(iband-1)*npw_k*my_nspinor+icg:iband*npw_k*my_nspinor+icg)
     649          906 :            call meanvalue_g(dotr,kinpw,0,istwf_k,mpi_enreg,npw_k,my_nspinor,cwavef,cwavef,0)
     650          906 :            energies%e_kinetic=energies%e_kinetic+dtset%wtk(ikpt)*occ_k(iband)*dotr
     651              :          end if
     652              :        end do
     653              :      end do
     654              : 
     655              : !    Compute nonlocal form factors ffnl at all (k+G):
     656          210 :      ider=0;dimffnl=1;nkpg=0
     657         1050 :      ABI_MALLOC(ffnl,(npw_k,dimffnl,psps%lmnmax,psps%ntypat))
     658              :      call mkffnl(psps%dimekb,dimffnl,psps%ekb,ffnl,psps%ffspl,&
     659              : &     gmet,gprimd,ider,ider,psps%indlmn,kg_k,kpg_dum,kpoint,psps%lmnmax,&
     660              : &     psps%lnmax,psps%mpsang,psps%mqgrid_ff,nkpg,&
     661              : &     npw_k,psps%ntypat,psps%pspso,psps%qgrid_ff,rmet,&
     662          210 : &     psps%usepaw,psps%useylm,ylm_k,ylmgr_dum)
     663              : 
     664              : !    Load k-dependent part in the Hamiltonian datastructure
     665              : !     - Compute 3D phase factors
     666              : !     - Prepare various tabs in case of band-FFT parallelism
     667              : !     - Load k-dependent quantities in the Hamiltonian
     668          840 :      ABI_MALLOC(ph3d,(2,npw_k,gs_hamk%matblk))
     669              :      call gs_hamk%load_k(kpt_k=dtset%kptns(:,ikpt),istwf_k=istwf_k,npw_k=npw_k,&
     670              : &     kinpw_k=kinpw,kg_k=kg_k,ffnl_k=ffnl,ph3d_k=ph3d,&
     671          210 : &     compute_ph3d=.true.,compute_gbound=(mpi_enreg%paral_kgb/=1))
     672              : 
     673              : !    Load band-FFT tabs (transposed k-dependent arrays)
     674          210 :      if (mpi_enreg%paral_kgb==1) then
     675            0 :        call bandfft_kpt_savetabs(my_bandfft_kpt,ffnl=ffnl_sav,ph3d=ph3d_sav,kinpw=kinpw_sav)
     676            0 :        call prep_bandfft_tabs(gs_hamk,ikpt,dtset%mkmem,mpi_enreg)
     677              :        call gs_hamk%load_k(npw_fft_k=my_bandfft_kpt%ndatarecv, &
     678              : &       gbound_k =my_bandfft_kpt%gbound, &
     679              : &       kinpw_k  =my_bandfft_kpt%kinpw_gather, &
     680              : &       kg_k     =my_bandfft_kpt%kg_k_gather, &
     681              : &       ffnl_k   =my_bandfft_kpt%ffnl_gather, &
     682            0 : &       ph3d_k   =my_bandfft_kpt%ph3d_gather)
     683              :      end if
     684              : 
     685              : !    If OpenMP GPU, load "hamiltonian" on GPU device
     686          210 :      if (dtset%gpu_option == ABI_GPU_OPENMP) then
     687            0 :        if(dtset%paral_kgb==0) then
     688            0 :          call ompgpu_load_hamilt_buffers(gs_hamk%kg_k,gs_hamk%kg_kp,gs_hamk%ffnl_k,gs_hamk%ph3d_k)
     689            0 :        else if(istwf_k==1) then
     690            0 :          call ompgpu_load_hamilt_buffers(gs_hamk%kg_k,gs_hamk%kg_kp,gs_hamk%ffnl_k,gs_hamk%ph3d_k,kg_k_gather=my_bandfft_kpt%kg_k_gather)
     691            0 :        else if(istwf_k==2) then
     692            0 :          call ompgpu_load_hamilt_buffers(gs_hamk%kg_k,gs_hamk%kg_kp,gs_hamk%ffnl_k,gs_hamk%ph3d_k,kg_k_gather=my_bandfft_kpt%kg_k_gather_sym)
     693              :        else
     694            0 :          ABI_ERROR("istwfk > 2 is not handled with OpenMP GPU offload mode !")
     695              :        end if
     696              :      end if
     697              : 
     698          210 :      choice=1-gs_hamk%usepaw ; signs=1 ; idir=0 ; nnlout=blocksize
     699              : 
     700              : !    Setup gemm_nonlop
     701          210 :      if (gemm_nonlop_use_gemm) then
     702              :        call set_gemm_nonlop_ikpt(my_ikpt,gs_hamk%npw_fft_k,gs_hamk%istwf_k,gs_hamk%indlmn,&
     703            0 :        &    gs_hamk%ntypat,gs_hamk%nattyp,gs_hamk%gpu_option)
     704              :      end if
     705              : 
     706              : #if defined HAVE_GPU_CUDA
     707              :      if (dtset%gpu_option==ABI_GPU_LEGACY .or. dtset%gpu_option==ABI_GPU_KOKKOS) then
     708              :        call gpu_update_ffnl_ph3d( &
     709              :          & ph3d, INT(size(ph3d,dim=1),c_int64_t)*size(ph3d,dim=2)*size(ph3d,dim=3), &
     710              :          & ffnl, INT(size(ffnl),c_int64_t) )
     711              :      end if
     712              : #endif
     713              : 
     714              : !    Compute nonlocal psp energy (NCPP) or Rhoij (PAW)
     715          630 :      ABI_MALLOC(enlout,(blocksize))
     716          420 :      ABI_MALLOC(occblock,(blocksize))
     717         2198 :      do iblock=1,nblockbd
     718         1988 :        iband=(iblock-1)*blocksize+1;iband_last=min(iband+blocksize-1,nband_k)
     719         1988 :        if(proc_distrb_cycle(mpi_enreg%proc_distrb,ikpt,iband,iband_last,isppol,me_distrb)) cycle
     720              : 
     721              : !      Select occupied bands
     722         3976 :        occblock(:)=occ(1+(iblock-1)*blocksize+bdtot_index:iblock*blocksize+bdtot_index)
     723         6174 :        if(abs(maxval(occblock))>=tol8 ) then
     724              :          cwavef(:,1:npw_k*my_nspinor*blocksize)=&
     725      2115366 : &         cg(:,1+(iblock-1)*npw_k*my_nspinor*blocksize+icg:iblock*npw_k*my_nspinor*blocksize+icg)
     726              : 
     727          906 :          paw_opt=gs_hamk%usepaw;cpopt=gs_hamk%usepaw-1
     728              : 
     729          906 :          if (mpi_enreg%paral_kgb/=1) then
     730          906 :            tim_nonlop=3
     731              :            call nonlop(choice,cpopt,cwaveprj,enlout,gs_hamk,idir,(/zero/),mpi_enreg,blocksize,nnlout,&
     732          906 : &           paw_opt,signs,nonlop_out,tim_nonlop,cwavef,cwavef)
     733              :          else
     734            0 :            tim_nonlop=14
     735              :            call prep_nonlop(choice,cpopt,cwaveprj,enlout,gs_hamk,idir,(/zero/),blocksize,&
     736            0 : &           mpi_enreg,nnlout,paw_opt,signs,nonlop_out,tim_nonlop,cwavef,cwavef)
     737              :          end if
     738              : 
     739         1812 :          do iblocksize=1,blocksize
     740          906 :            iband=(iblock-1)*blocksize+iblocksize
     741          906 :            energies%e_eigenvalues=energies%e_eigenvalues+dtset%wtk(ikpt)*occ_k(iband)*eig_k(iband)
     742              : !          WARNING : the Fock contribution is NOT computed !!!
     743         1812 :            energies%e_nlpsp_vfock=energies%e_nlpsp_vfock+dtset%wtk(ikpt)*occ_k(iband)*enlout(iblocksize)
     744              :          end do
     745              : 
     746              : !        PAW: accumulate rhoij
     747          906 :          if (psps%usepaw==1) then
     748          452 :            cplex=merge(1,2,istwf_k>1)
     749          452 :            if (mpi_enreg%paral_spinor==1) then
     750              :              call pawcprj_gather_spin(cwaveprj,cwaveprj_gat,dtset%natom,1,my_nspinor,dtset%nspinor,&
     751            0 : &             mpi_enreg%comm_spinor,ierr)
     752              :              call pawaccrhoij(gs_hamk%atindx,cplex,cwaveprj_gat,cwaveprj_gat,0,isppol,dtset%natom,dtset%natom,&
     753            0 : &             dtset%nspinor,occ_k(iband),option_rhoij,pawrhoij_unsym,use_timerev,use_zeromag,dtset%wtk(ikpt))
     754              :            else
     755              :              call pawaccrhoij(gs_hamk%atindx,cplex,cwaveprj,cwaveprj,0,isppol,dtset%natom,dtset%natom,&
     756          452 : &             dtset%nspinor,occ_k(iband),option_rhoij,pawrhoij_unsym,use_timerev,use_zeromag,dtset%wtk(ikpt))
     757              :            end if
     758              :          end if
     759              : 
     760              : !        End loop on bands
     761              :        end if
     762              :      end do
     763              : 
     764              : !    Compute residual of each band (for informative purposes)
     765          210 :      call mkresi(cg,eig_k,gs_hamk,icg,ikpt,isppol,mcg,mpi_enreg,nband_k,dtset%prtvol,resid_k)
     766         2198 :      resid(1+bdtot_index : nband_k+bdtot_index) = resid_k(:)
     767              : 
     768              : !    Restore the bandfft tabs
     769          210 :      if (mpi_enreg%paral_kgb==1) then
     770            0 :        call bandfft_kpt_restoretabs(my_bandfft_kpt,ffnl=ffnl_sav,ph3d=ph3d_sav,kinpw=kinpw_sav)
     771              :      end if
     772              : 
     773              : !    Incremente indexes
     774          210 :      bdtot_index=bdtot_index+nband_k
     775          210 :      if (dtset%mkmem/=0) then
     776          210 :        icg=icg+npw_k*my_nspinor*nband_k
     777          210 :        ikg=ikg+npw_k
     778              :      end if
     779              : 
     780              : #if defined HAVE_GPU_CUDA
     781              :      if(dtset%gpu_option==ABI_GPU_LEGACY .or. dtset%gpu_option==ABI_GPU_KOKKOS) then
     782              :        call gpu_finalize_ffnl_ph3d()
     783              :      end if
     784              : #endif
     785              : 
     786          210 :      ABI_FREE(eig_k)
     787          210 :      ABI_FREE(occ_k)
     788          210 :      ABI_FREE(resid_k)
     789          210 :      ABI_FREE(enlout)
     790          210 :      ABI_FREE(occblock)
     791          210 :      ABI_FREE(ffnl)
     792          210 :      ABI_FREE(kinpw)
     793          210 :      ABI_FREE(ph3d)
     794          210 :      ABI_FREE(cwavef)
     795          210 :      ABI_FREE(kg_k)
     796          245 :      ABI_FREE(ylm_k)
     797              : 
     798              : !    End loops on isppol and ikpt
     799              :    end do
     800              :  end do
     801              : 
     802           35 :  call gs_hamk%free()
     803           35 :  if ( dtset%gpu_option == ABI_GPU_OPENMP) then
     804            0 :    call ompgpu_free_hamilt_buffers()
     805              :  end if
     806              : 
     807              :  if(xmpi_paral==1)then
     808              : !  Accumulate enl eeig and ek on all proc.
     809          105 :    ABI_MALLOC(buffer,(3+dtset%mband*dtset%nkpt*dtset%nsppol))
     810           35 :    buffer(1)=energies%e_nlpsp_vfock ; buffer(2)=energies%e_kinetic ; buffer(3)=energies%e_eigenvalues
     811         2023 :    do iresid=1,dtset%mband*dtset%nkpt*dtset%nsppol
     812         2023 :      buffer(iresid+3)=resid(iresid)
     813              :    end do
     814           35 :    call timab(48,1,tsec)
     815           35 :    call xmpi_sum(buffer,spaceComm,ierr)
     816           35 :    call timab(48,2,tsec)
     817           35 :    energies%e_nlpsp_vfock=buffer(1) ; energies%e_kinetic=buffer(2) ; energies%e_eigenvalues=buffer(3)
     818         2023 :    do iresid=1,dtset%mband*dtset%nkpt*dtset%nsppol
     819         2023 :      resid(iresid)=buffer(iresid+3)
     820              :    end do
     821           35 :    ABI_FREE(buffer)
     822              : !  Accumulate rhoij_
     823           70 :    if (psps%usepaw==1) then
     824            3 :      call pawrhoij_mpisum_unpacked(pawrhoij_unsym,spaceComm,comm2=mpi_enreg%comm_band)
     825              :    end if
     826              :  end if
     827              : 
     828              : !Compute total (free) energy
     829           35 :  if (optene==0.or.optene==2) then
     830              :    etotal = energies%e_kinetic + energies%e_hartree + energies%e_xc + &
     831              : !&   energies%e_nlpsp_vfock - energies%e_fock0 +
     832              : !   Should compute the e_fock0 energy !! Also, the Fock contribution to e_nlpsp_vfock
     833           34 : &   energies%e_nlpsp_vfock + energies%e_localpsp + energies%e_corepsp
     834           34 :    if (psps%usepaw==1) etotal=etotal + energies%paw%epaw
     835            1 :  else if (optene==1.or.optene==3) then
     836              :    etotal = energies%e_eigenvalues - energies%e_hartree + energies%e_xc - &
     837            1 : &   energies%e_xcdc + energies%e_corepsp - energies%e_corepspdc
     838            1 :    if (psps%usepaw==1) etotal=etotal + energies%paw%epaw_dc
     839              :  end if
     840           35 :  etotal = etotal + energies%e_ewald + energies%e_chempot + energies%e_vdw_dftd
     841              : 
     842              : !Add the contribution of extfpmd to the entropy
     843           35 :  if(associated(extfpmd)) then
     844            0 :    energies%e_extfpmd=extfpmd%e_kinetic
     845            0 :    energies%edc_extfpmd=extfpmd%edc_kinetic
     846            0 :    if(optene==0.or.optene==2) etotal=etotal+energies%e_extfpmd
     847            0 :    if(optene==1.or.optene==3) etotal=etotal+energies%edc_extfpmd
     848              :  end if
     849              : 
     850              :  ! Add the contribution from cores
     851           35 :  if(present(rcpaw)) then
     852           35 :    if(associated(rcpaw)) then
     853            0 :      energies%paw%epaw_core=rcpaw%ehnzc+rcpaw%ekinc
     854            0 :      energies%paw%epaw_core_dc=rcpaw%eeigc-rcpaw%edcc+rcpaw%ehnzc
     855            0 :      if (ipositron/=1) then
     856            0 :        do itypat=1,dtset%ntypat
     857            0 :          energies%paw%epaw_core=energies%paw%epaw_core+pawtab(itypat)%exccore*rcpaw%atm(itypat)%mult
     858            0 :          energies%paw%epaw_core_dc=energies%paw%epaw_core_dc+pawtab(itypat)%exccore*rcpaw%atm(itypat)%mult
     859              :        enddo
     860              :      endif
     861            0 :      if(optene==0.or.optene==2) etotal=etotal+energies%paw%epaw_core
     862            0 :      if(optene==1.or.optene==3) etotal=etotal+energies%paw%epaw_core_dc
     863              :    endif
     864              :  endif
     865              : 
     866           35 :  call entropy(dtset,energies)
     867           35 :  etotal=etotal+energies%e_entropy
     868              : 
     869              : !Additional stuff for electron-positron
     870           35 :  if (dtset%positron/=0) then
     871            0 :    if (ipositron==0) then
     872            0 :      energies%e_electronpositron  =zero
     873            0 :      energies%edc_electronpositron=zero
     874              :    else
     875            0 :      energies%e_electronpositron  =electronpositron%e_hartree+electronpositron%e_xc
     876            0 :      energies%edc_electronpositron=electronpositron%e_hartree+electronpositron%e_xcdc
     877            0 :      if (psps%usepaw==1) then
     878            0 :        energies%e_electronpositron  =energies%e_electronpositron  +electronpositron%e_paw
     879            0 :        energies%edc_electronpositron=energies%edc_electronpositron+electronpositron%e_pawdc
     880              :      end if
     881              :    end if
     882            0 :    if (optene==0.or.optene==2) electronpositron%e0=etotal
     883            0 :    if (optene==1.or.optene==3) electronpositron%e0=etotal-energies%edc_electronpositron
     884            0 :    etotal=electronpositron%e0+energies%e0_electronpositron+energies%e_electronpositron
     885              :  end if
     886              : 
     887              : !Compute new charge density based on incoming wf
     888              : !Keep rhor and rhog intact for later use e.g. in stress. (? MT 08-12-2008: is that true now ?)
     889              : !=== Norm-conserving psps: simply compute rho from WFs
     890              :  !paw_dmft%use_dmft=0 ! dmft not used here
     891              :  !paw_dmft%use_sc_dmft=0 ! dmft not used here
     892           35 :  if (psps%usepaw==0) then
     893           32 :    tim_mkrho=3
     894              :    call mkrho(cg,dtset,gprimd,irrzon,kg,mcg,mpi_enreg,&
     895              : &   npwarr,occ,paw_dmft,phnons,rhog,rhor,rprimd,tim_mkrho,ucvol,wvl_den,wfs,&
     896           32 : &   extfpmd=extfpmd)
     897           32 :    if(dtset%usekden==1)then
     898              :      call mkrho(cg,dtset,gprimd,irrzon,kg,mcg,mpi_enreg,&
     899            0 : &     npwarr,occ,paw_dmft,phnons,taug,taur,rprimd,tim_mkrho,ucvol,wvl_den,wfs,option=1)
     900              :    end if
     901              :  else
     902              : !  === PAW case: symmetrize rhoij and add compensation charge density
     903            3 :    tim_mkrho=3;option=1;choice=1
     904              :    call pawrhoij_symrhoij(pawrhoij,pawrhoij_unsym,choice,gprimd,indsym,0,dtset%natom,dtset%nsym,&
     905              : &   dtset%ntypat,option,pawang,dtset%pawprtvol,pawtab,rprimd,dtset%symafm,symrec,dtset%typat,&
     906            3 : &   comm_atom=mpi_enreg%comm_atom,mpi_atmtab=mpi_enreg%my_atmtab)
     907            3 :    ider=0;izero=0;cplex=1;ipert=0;idir=0;qpt(:)=zero
     908              :    call pawmknhat(compch_fft,cplex,ider,idir,ipert,izero,gprimd,&
     909              : &   my_natom,dtset%natom,nfftf,ngfftf,&
     910              : &   0,dtset%nspden,dtset%ntypat,pawang,pawfgrtab,rhodum,nhat,pawrhoij,pawrhoij,&
     911              : &   pawtab,qpt,rprimd,ucvol_local,dtset%usewvl,xred,&
     912              : &   comm_fft=mpi_enreg%comm_fft,paral_kgb=dtset%paral_kgb,me_g0=mpi_enreg%me_g0,&
     913              : &   comm_atom=mpi_enreg%comm_atom,mpi_atmtab=mpi_enreg%my_atmtab,&
     914            3 : &   distribfft=mpi_enreg%distribfft,mpi_comm_wvl=mpi_enreg%comm_wvl)
     915              : 
     916           12 :    ABI_MALLOC(rhowfr,(dtset%nfft,dtset%nspden))
     917            9 :    ABI_MALLOC(rhowfg,(2,dtset%nfft))
     918        98602 :    rhowfr(:,:)=zero
     919              :    call mkrho(cg,dtset,gprimd,irrzon,kg,mcg,mpi_enreg,&
     920            3 : &   npwarr,occ,paw_dmft,phnons,rhowfg,rhowfr,rprimd,tim_mkrho,ucvol_local,wvl_den,wfs)
     921              : 
     922              :    call pawmkrho(1,compch_fft,cplex,gprimd,idir,indsym,ipert,mpi_enreg,&
     923              : &   my_natom,dtset%natom,dtset%nspden,dtset%nsym,dtset%ntypat,dtset%paral_kgb,pawang,pawfgr,pawfgrtab,&
     924              : &   dtset%pawprtvol,pawrhoij,pawrhoij_unsym,pawtab,qpt,rhowfg,rhowfr,rhor,rprimd,dtset%symafm,&
     925            3 : &   symrec,dtset%typat,ucvol,dtset%usewvl,xred,pawnhat=nhat,rhog=rhog,extfpmd=extfpmd)
     926            3 :    call pawrhoij_free_unpacked(pawrhoij_unsym)
     927            3 :    if (paral_atom) then
     928            0 :      call pawrhoij_free(pawrhoij_unsym)
     929            0 :      ABI_FREE(pawrhoij_unsym)
     930              :    end if
     931            3 :    if(dtset%usekden==1)then
     932              :      call mkrho(cg,dtset,gprimd,irrzon,kg,mcg,mpi_enreg,npwarr,occ,paw_dmft,phnons,&
     933            0 : &               rhowfg,rhowfr,rprimd,tim_mkrho,ucvol,wvl_den,wfs,option=1)
     934            0 :      call transgrid(1,mpi_enreg,dtset%nspden,+1,1,1,dtset%paral_kgb,pawfgr,rhowfg,taug,rhowfr,taur)
     935              :    end if
     936            3 :    ABI_FREE(rhowfr)
     937            3 :    ABI_FREE(rhowfg)
     938              :  end if
     939              : 
     940           35 :  ABI_COMMENT('New density rho(r) made from input wfs')
     941              : 
     942           35 :  call timab(59,2,tsec)
     943              : 
     944           35 :  ABI_FREE(vlocal)
     945           35 :  if (with_vxctau) then
     946            0 :    ABI_FREE(vxctaulocal)
     947              :  end if
     948              : 
     949           35 :  if (psps%usepaw==1) then
     950            3 :    call pawcprj_free(cwaveprj)
     951           47 :    ABI_FREE(cwaveprj)
     952            3 :    if (mpi_enreg%paral_spinor==1) then
     953            0 :      call pawcprj_free(cwaveprj_gat)
     954            0 :      ABI_FREE(cwaveprj_gat)
     955              :    else
     956            3 :      nullify(cwaveprj_gat)
     957              :    end if
     958              :  end if
     959              : 
     960              :  DBG_EXIT("COLL")
     961              : 
     962           70 : end subroutine energy
     963              : !!***
     964              : 
     965              : !!****f* ABINIT/mkresi
     966              : !! NAME
     967              : !!  mkresi
     968              : !!
     969              : !! FUNCTION
     970              : !!  Make residuals from knowledge of wf in G space and application of Hamiltonian.
     971              : !!
     972              : !! INPUTS
     973              : !!  cg(2,mcg)=<G|Cnk>=Fourier coefficients of wavefunction
     974              : !!  gs_hamk <type(gs_hamiltonian_type)>=all data for the Hamiltonian at k
     975              : !!  icg=shift to be applied on the location of data in the array cg
     976              : !!  ikpt=index of k-point
     977              : !!  isppol=index of spin
     978              : !!  mcg=second dimension of the cg array
     979              : !!  mpi_enreg=information about MPI parallelization
     980              : !!  nband=number of bands involved in subspace matrix.
     981              : !!  npw=number of planewaves in basis sphere at this k point.
     982              : !!  prtvol=control print volume and debugging output
     983              : !!  usepaw= 0 for non paw calculation; =1 for paw calculation
     984              : !!
     985              : !! OUTPUT
     986              : !!  eig_k(nband)$= \langle C_n \mid H \mid C_n \rangle $ for each band.
     987              : !!  resid_k(nband)=residual for each band
     988              : !!   $= \langle C_n \mid H H \mid C_n \rangle- \langle C_n \mid H \mid C_n \rangle^2 $.
     989              : !!
     990              : !! SOURCE
     991              : 
     992          210 : subroutine mkresi(cg,eig_k,gs_hamk,icg,ikpt,isppol,mcg,mpi_enreg,nband,prtvol,resid_k)
     993              : 
     994              : !Arguments ------------------------------------
     995              : !scalars
     996              :  integer,intent(in) :: icg,ikpt,isppol,mcg,nband,prtvol
     997              :  type(MPI_type),intent(inout) :: mpi_enreg
     998              :  type(gs_hamiltonian_type),intent(inout) :: gs_hamk
     999              : !arrays
    1000              :  real(dp),intent(in) :: cg(2,mcg)
    1001              :  real(dp),intent(out) :: eig_k(nband),resid_k(nband)
    1002              : 
    1003              : !Local variables-------------------------------
    1004              : !scalars
    1005              :  integer,parameter :: tim_getghc=3
    1006              :  integer :: blocksize,cpopt,iband,iband_last,iblock,iblocksize,ipw,ipw_shift
    1007              :  integer :: my_nspinor,nblockbd,npw_k
    1008              :  real(dp) :: doti,dotr
    1009              : !arrays
    1010              :  real(dp) :: tsec(2)
    1011          420 :  real(dp),allocatable,target :: cwavef(:,:),ghc(:,:),gsc(:,:),gvnlxc(:,:)
    1012          420 :  real(dp), contiguous, pointer :: cwavef_ptr(:,:),ghc_ptr(:,:),gsc_ptr(:,:)
    1013         1470 :  type(pawcprj_type) :: cwaveprj(1,1)
    1014              : ! *************************************************************************
    1015              : 
    1016              : !Keep track of total time spent in mkresi
    1017          210 :  call timab(13,1,tsec)
    1018              : 
    1019              : !Parallelism over FFT and/or bands: define sizes and tabs
    1020          210 :  my_nspinor=max(1,gs_hamk%nspinor/mpi_enreg%nproc_spinor)
    1021          210 :  if (mpi_enreg%paral_kgb==1) then
    1022            0 :    nblockbd=nband/(mpi_enreg%nproc_band*mpi_enreg%bandpp)
    1023              :  else
    1024          210 :    nblockbd=nband/mpi_enreg%nproc_fft
    1025          210 :    if (nband/=nblockbd*mpi_enreg%nproc_fft) nblockbd=nblockbd+1
    1026              :  end if
    1027          210 :  blocksize=nband/nblockbd
    1028              : 
    1029          210 :  npw_k=gs_hamk%npw_k
    1030          630 :  ABI_MALLOC(cwavef,(2,npw_k*my_nspinor))
    1031          420 :  ABI_MALLOC(ghc,(2,npw_k*my_nspinor))
    1032          420 :  ABI_MALLOC(gvnlxc,(2,npw_k*my_nspinor))
    1033          210 :  if (gs_hamk%usepaw==1)  then
    1034          146 :    ABI_MALLOC(gsc,(2,npw_k*my_nspinor))
    1035              :  else
    1036          137 :    ABI_MALLOC(gsc,(0,0))
    1037              :  end if
    1038              : 
    1039              : !Loop over (blocks of) bands
    1040         2198 :  do iblock=1,nblockbd
    1041         1988 :    iband=(iblock-1)*blocksize+1;iband_last=min(iband+blocksize-1,nband)
    1042         1988 :    if(proc_distrb_cycle(mpi_enreg%proc_distrb,ikpt,iband,iband_last,isppol,mpi_enreg%me_kpt)) cycle
    1043              : 
    1044              : !  Load |Cn>
    1045         1988 :    ipw_shift=(iblock-1)*npw_k*my_nspinor*blocksize+icg
    1046              : !$OMP PARALLEL DO
    1047       892572 :    do ipw=1,npw_k*my_nspinor*blocksize
    1048       890584 :      cwavef(1,ipw)=cg(1,ipw+ipw_shift)
    1049       892572 :      cwavef(2,ipw)=cg(2,ipw+ipw_shift)
    1050              :    end do
    1051              : 
    1052              : !  Compute H|Cn>
    1053         1988 :    cpopt=-1
    1054         1988 :    if (mpi_enreg%paral_kgb==0) then
    1055              :      call getghc(cpopt,cwavef,cwaveprj,ghc,gsc,gs_hamk,gvnlxc,zero,mpi_enreg,1,&
    1056         1988 : &     prtvol,gs_hamk%usepaw,tim_getghc,0)
    1057              :    else
    1058              :      call prep_getghc(cwavef,gs_hamk,gvnlxc,ghc,gsc,zero,nband,mpi_enreg,&
    1059              : &     prtvol,gs_hamk%usepaw,cpopt,cwaveprj,&
    1060            0 : &     already_transposed=.false.)
    1061              :    end if
    1062              : 
    1063              :    !call cg_get_eigens(usepaw, istwf_k, npwsp, nband, cg, ghc, gsc, eig, me_g0, comm_bsf)
    1064              :    !call cg_get_residvecs(usepaw, npwsp, nband, eig, cg, ghc, gsc, gwork)
    1065              :    !call cg_norm2g(istwf_k, npwsp, nband, gwork, resid, me_g0, comm_bsf)
    1066              :    ! MG: Communicators are wrongi if paral_kgb. One should use mpi_enreg%comm_bandspinorfft
    1067              : 
    1068              : !  Compute the residual, <Cn|(H-<Cn|H|Cn>)**2|Cn>:
    1069         4186 :    do iblocksize=1,blocksize
    1070         1988 :      iband=(iblock-1)*blocksize+iblocksize
    1071         1988 :      ipw_shift=(iblocksize-1)*npw_k*my_nspinor
    1072         1988 :      cwavef_ptr => cwavef(:,1+ipw_shift:npw_k*my_nspinor+ipw_shift)
    1073         1988 :      ghc_ptr    => ghc   (:,1+ipw_shift:npw_k*my_nspinor+ipw_shift)
    1074              : 
    1075              : !    First get eigenvalue <Cn|H|Cn>:
    1076              :      call dotprod_g(dotr,doti,gs_hamk%istwf_k,npw_k*my_nspinor,1,cwavef_ptr,ghc_ptr,&
    1077         1988 : &     mpi_enreg%me_g0,mpi_enreg%comm_spinorfft)
    1078         1988 :      eig_k(iband)=dotr
    1079              : 
    1080              : !    Next need <G|(H-S<Cn|H|Cn>)|Cn> (in ghc):
    1081         1988 :      if (gs_hamk%usepaw==0) then
    1082              : !$OMP PARALLEL DO PRIVATE(ipw) SHARED(cwavef_ptr,ghc_ptr,eig_k,iband,npw_k,my_nspinor)
    1083       362941 :        do ipw=1,npw_k*my_nspinor
    1084       362050 :          ghc_ptr(1,ipw)=ghc_ptr(1,ipw)-eig_k(iband)*cwavef_ptr(1,ipw)
    1085       362941 :          ghc_ptr(2,ipw)=ghc_ptr(2,ipw)-eig_k(iband)*cwavef_ptr(2,ipw)
    1086              :        end do
    1087              :      else
    1088         1097 :        gsc_ptr => gsc(:,1+ipw_shift:npw_k*my_nspinor+ipw_shift)
    1089              : !$OMP PARALLEL DO PRIVATE(ipw) SHARED(gsc_ptr,ghc_ptr,eig_k,iband,npw_k,my_nspinor)
    1090       529631 :        do ipw=1,npw_k*my_nspinor
    1091       528534 :          ghc_ptr(1,ipw)=ghc_ptr(1,ipw)-eig_k(iband)*gsc_ptr(1,ipw)
    1092       529631 :          ghc_ptr(2,ipw)=ghc_ptr(2,ipw)-eig_k(iband)*gsc_ptr(2,ipw)
    1093              :        end do
    1094              :      end if
    1095              : 
    1096              : !    Then simply square the result:
    1097              :      call sqnorm_g(dotr,gs_hamk%istwf_k,npw_k*my_nspinor,ghc_ptr,&
    1098         1988 : &     mpi_enreg%me_g0,mpi_enreg%comm_fft)
    1099         3976 :      resid_k(iband)=dotr
    1100              : 
    1101              :    end do ! iblocksize
    1102              : 
    1103              :  end do ! iblock
    1104              : 
    1105          210 :  ABI_FREE(cwavef)
    1106          210 :  ABI_FREE(ghc)
    1107          210 :  ABI_FREE(gvnlxc)
    1108          210 :  ABI_FREE(gsc)
    1109              : 
    1110          210 :  call timab(13,2,tsec)
    1111              : 
    1112          840 : end subroutine mkresi
    1113              : !!***
    1114              : 
    1115              : !!****f* ABINIT/entropy
    1116              : !! NAME
    1117              : !!  entropy
    1118              : !!
    1119              : !! FUNCTION
    1120              : !!  Compute electronic entropy terms
    1121              : !!  This subroutine returns the total entropy and entropy energy. In the most
    1122              : !!  common case, at finite temperature, the electronic entropy is mainly constitued
    1123              : !!  of the non-interacting entropy (entropy_ks). Finite-temperature exchange-correlation
    1124              : !!  functionals or other methods may introduce additional entropy terms.
    1125              : !!
    1126              : !! NOTE
    1127              : !!
    1128              : !! INPUTS
    1129              : !!  dtset <type(dataset_type)>=all input variables for this dataset
    1130              : !!   | occopt=option for occupancies
    1131              : !!   | tsmear=smearing energy or temperature (if metal)
    1132              : !!   | tphysel=electornic temperature for particular values of occopt
    1133              : !!  energies <type(energies_type)>=all part of total energy.
    1134              : !!   | entropy_ks(IN)=non-interacting entropy of the kohn-sham states
    1135              : !!   | entropy_paw(IN)=entropy due to paw corrections (for finite-temperature xc functionals)
    1136              : !!   | entropy_xc(IN)=exchange-correlation entropy (for finite-temperature xc functionals)
    1137              : !!   | entropy_extfpmd(IN)=entropy of extfpmd model
    1138              : !!   | entropy_imp(IN)=entropy of impurity electrons (for DFT+DMFT)
    1139              : !!
    1140              : !! OUTPUT
    1141              : !!  energies <type(energies_type)>=all part of total energy.
    1142              : !!   | entropy(OUT)=total entropy
    1143              : !!   | e_entropy(OUT)=total entropy energy (hartree units)
    1144              : !!
    1145              : !! SOURCE
    1146        53849 : subroutine entropy(dtset,energies)
    1147              : !Arguments ------------------------------------
    1148              : !scalars
    1149              :  type(dataset_type),intent(in) :: dtset
    1150              :  type(energies_type),intent(inout) :: energies
    1151              : ! *************************************************************************
    1152              : 
    1153              : !In case we have other sources of entropy than kohn-sham states occupation,
    1154              : !we sum all entropy terms. %entropy is now total entropy.
    1155              : !Examples of other sources of entropy: finite-temperature xc functionals, extfpmd, ...
    1156        53849 :  energies%entropy=energies%entropy_ks
    1157        53849 :  if(abs(energies%paw%entropy_paw)>tiny(zero))     energies%entropy=energies%entropy+energies%paw%entropy_paw
    1158        53849 :  if(abs(energies%entropy_xc)>tiny(zero))      energies%entropy=energies%entropy+energies%entropy_xc
    1159        53849 :  if(abs(energies%entropy_extfpmd)>tiny(zero)) energies%entropy=energies%entropy+energies%entropy_extfpmd
    1160        53849 :  if(abs(energies%entropy_imp)>tiny(zero))     energies%entropy=energies%entropy+energies%entropy_imp
    1161              : 
    1162              : !When the finite-temperature VG broadening scheme is used,
    1163              : !the total entropy contribution "tsmear*entropy" has a meaning,
    1164              : !and gather the two last terms of Eq.8 of VG paper
    1165              : !Warning : might have to be changed for fixed moment calculations
    1166        53849 :  if(dtset%occopt>=3 .and. dtset%occopt<=8) then
    1167        14965 :    if (abs(dtset%tphysel) < tol10) then
    1168        14919 :      energies%e_entropy = - dtset%tsmear * energies%entropy
    1169              :    else
    1170           46 :      energies%e_entropy = - dtset%tphysel * energies%entropy
    1171              :    end if
    1172              :  else
    1173        38884 :    energies%e_entropy = zero
    1174              :  end if
    1175              : 
    1176        53849 : end subroutine entropy
    1177              : 
    1178              : end module m_dft_energy
    1179              : !!***
        

Generated by: LCOV version 2.3-1