! ! Copyright (C) 2001-2009 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 solve_linter(dvbarein, iw, drhoscf) !----------------------------------------------------------------------------- ! HL ! Driver routine for the solution of the linear system which ! defines the change of the wavefunction due to a perturbing potential ! parameterized in r and iw. i.e. v_{r, iw} (r') ! It performs the following tasks: ! a) computes the bare potential term Delta V | psi > ! b) adds to it the screening term Delta V_{SCF} | psi > ! c) applies P_c^+ (orthogonalization to valence states) ! d) calls c_bi_cgsolve_all to solve the linear system ! e) computes Delta rho, Delta V_{SCF}. ! Currently symmetrized in terms of mode etc. Might need to strip this out ! and check PW for how it stores/symmetrizes charge densities. !---------------------------------------------------------------------------- !------------------------------------------------------------------------------ ! grep @10TION for all the instances where I've turned off the regular ! PH parallelism which always assumes we have split k-points across processors. ! Each processor has a full slice of k-points. USE kinds, ONLY : DP USE ions_base, ONLY : nat, ntyp => nsp, ityp USE io_global, ONLY : stdout, ionode USE io_files, ONLY : prefix, iunigk USE check_stop, ONLY : check_stop_now USE wavefunctions_module, ONLY : evc USE constants, ONLY : degspin USE cell_base, ONLY : tpiba2 USE ener, ONLY : ef USE klist, ONLY : lgauss, degauss, ngauss, xk, wk, nkstot USE gvect, ONLY : nrxx, g, nl, nr1, nr2, nr3, nrx1, nrx2, nrx3 USE gsmooth, ONLY : doublegrid, nrxxs, nr1s, nr2s, nr3s, nrx1s, nrx2s, nrx3s USE lsda_mod, ONLY : lsda, nspin, current_spin, isk USE spin_orb, ONLY : domag USE wvfct, ONLY : nbnd, npw, npwx, igk,g2kin, et USE scf, ONLY : rho USE uspp, ONLY : okvan, vkb USE uspp_param, ONLY : upf, nhm, nh USE noncollin_module, ONLY : noncolin, npol, nspin_mag USE paw_variables, ONLY : okpaw USE paw_onecenter, ONLY : paw_dpotential, paw_dusymmetrize, & paw_dumqsymmetrize USE control_gw, ONLY : rec_code, niter_gw, nmix_gw, tr2_gw, & alpha_pv, lgamma, lgamma_gamma, convt, & nbnd_occ, alpha_mix, ldisp, rec_code_read, & where_rec, flmixdpot, current_iq, & ext_recover, eta USE nlcc_gw, ONLY : nlcc_any USE units_gw, ONLY : iudrho, lrdrho, iudwf, lrdwf, iubar, lrbar, & iuwfc, lrwfc, iunrec, iudvscf, iudwfm, iudwfp USE output, ONLY : fildrho, fildvscf USE gwus, ONLY : int3_paw, becsumort ! USE eqv, ONLY : dvpsi, dpsi, evq, eprec, dpsim, dpsip USE eqv, ONLY : dvpsi, dpsi, evq, eprec USE qpoint, ONLY : xq, npwq, igkq, nksq, ikks, ikqs USE modes, ONLY : npertx, npert, u, t, irotmq, tmq, & minus_q, irgq, nsymq, rtau USE recover_mod, ONLY : read_rec, write_rec ! used oly to write the restart file USE mp_global, ONLY : inter_pool_comm, intra_pool_comm, mpime, mp_global_end,& intra_image_comm USE mp, ONLY : mp_sum, mp_barrier ! USE freq_gw, ONLY : fpol, fiu, nfs, nfsmax implicit none ! counter on frequencies. integer :: iw, ir integer :: irr, imode0, npe ! input: the irreducible representation ! input: the number of perturbation ! input: the position of the modes complex(DP) :: drhoscf (nrxx, nfs) ! output: the change of the scf charge complex(DP) :: dvbarein (nrxxs) ! HL prec ! HL careful now... complexifying preconditioner: real(DP) , allocatable :: h_diag (:,:) ! h_diag: diagonal part of the Hamiltonian real(DP) :: thresh, anorm, averlt, dr2 ! thresh: convergence threshold ! anorm : the norm of the error ! averlt: average number of iterations ! dr2 : self-consistency error real(DP) :: dos_ef, weight, aux_avg (2) ! Misc variables for metals ! dos_ef: density of states at Ef real(DP), external :: w0gauss, wgauss ! functions computing the delta and theta function complex(DP), allocatable, target :: dvscfin(:,:) ! change of the scf potential complex(DP), pointer :: dvscfins (:,:) ! change of the scf potential (smooth part only) complex(DP), allocatable :: drhoscfh (:,:), dvscfout (:,:) ! change of rho / scf potential (output) ! change of scf potential (output) complex(DP), allocatable :: ldos (:,:), ldoss (:,:), mixin(:), mixout(:), & dbecsum (:,:,:), dbecsum_nc(:,:,:,:), aux1 (:,:) complex(DP) :: cw complex(DP), allocatable :: etc(:,:) !HL dbecsum (:,:,:,:), dbecsum_nc(:,:,:,:,:), aux1 (:,:) ! Misc work space ! ldos : local density of states af Ef ! ldoss: as above, without augmentation charges ! dbecsum: the derivative of becsum ! becsum1 PAW array. REAL(DP), allocatable :: becsum1(:,:,:) !For approx, mixing scheme. real(kind=DP) :: DZNRM2 complex(kind=DP) :: ZDOTC external ZDOTC, DZNRM2 logical :: conv_root, & ! true if linear system is converged exst, & ! used to open the recover file lmetq0, & ! true if xq=(0,0,0) in a metal cgsolver integer :: kter, & ! counter on iterations iter0, & ! starting iteration ipert, & ! counter on perturbations ibnd, & ! counter on bands iter, & ! counter on iterations lter, & ! counter on iterations of linear system ltaver, & ! average counter lintercall, & ! average number of calls to cgsolve_all ik, ikk, & ! counter on k points ikq, & ! counter on k+q points ig, & ! counter on G vectors ndim, & ! integer actual row dimension of dpsi is, & ! counter on spin polarizations nt, & ! counter on types na, & ! counter on atoms nrec, nrec1,& ! the record number for dvpsi and dpsi ios, & ! integer variable for I/O control mode, & ! mode index igpert, & ! bare perturbation g vector. lmres ! number of gmres iterations to include when using bicgstabl. real(DP) :: tcpu, get_clock ! timing variables real(DP) :: meandvb !external ch_psi_all, cg_psi, ccg_psi, cch_psi_all_fix external ch_psi_all, cg_psi, cch_psi_all_fix COMPLEX(DP) :: dpsip(npwx*npol, nbnd), dpsim(npwx*npol, nbnd) allocate (dpsi(npwx*npol, nbnd)) IF (rec_code_read > 20 ) RETURN !HL- Allocate arrays for dV_scf (need to alter these from (nrxx, nspin_mag, npe) to just (nrxx, nspin_mag). npe = 1 imode0 = 1 irr = 1 ipert = 1 lter = 0 lmres = 1 !HLallocate (hpsi(npwx*npol, 4)) ! Test array for whether linear system is being properly solved call start_clock ('solve_linter') allocate (dvscfout ( nrxx , nfs)) allocate (drhoscfh ( nrxx , nfs)) allocate (dvscfin ( nrxx , nfs)) allocate (etc(nbnd, nkstot)) allocate (dbecsum ( (nhm * (nhm + 1))/2 , nat, nspin_mag)) if (doublegrid) then allocate (dvscfins ( nrxxs , nfs)) else dvscfins => dvscfin endif !Complex eigenvalues IF (noncolin) allocate (dbecsum_nc (nhm, nhm, nat , nspin)) allocate (aux1 ( nrxxs, npol)) allocate (h_diag ( npwx*npol, nbnd)) if (rec_code_read == 10.AND.ext_recover) then ! restart from GW calculation rec_code=0 else iter0 = 0 convt =.FALSE. where_rec='no_recover' endif IF (convt) GOTO 155 ! In this case it has recovered after computing the contribution ! to the dynamical matrix. This is a new iteration that has to ! start from the beginning. IF (iter0==-1000) iter0=0 ! The outside loop is over the iterations. ! niter_gw := maximum number of iterations do kter = 1, niter_gw iter = kter + iter0 ltaver = 0 lintercall = 0 drhoscf(:,:) = (0.d0, 0.d0) drhoscfh(:,:) = (0.d0, 0.d0) dbecsum(:,:,:) = (0.d0, 0.d0) IF (noncolin) dbecsum_nc = (0.d0, 0.d0) if (nksq.gt.1) rewind (unit = iunigk) !start kpoints loop do ik = 1, nksq if (nksq.gt.1) then read (iunigk, err = 100, iostat = ios) npw, igk 100 call errore ('solve_linter', 'reading igk', abs (ios) ) endif ! lgamma is a q=0 computation if (lgamma) npwq = npw ! k and k+q mesh defined in initialize_gw: ! ikks(ik) = 2 * ik - 1 ! ikqs(ik) = 2 * ik ikk = ikks(ik) ikq = ikqs(ik) if (lsda) current_spin = isk (ikk) if (.not.lgamma.and.nksq.gt.1) then read (iunigk, err = 200, iostat = ios) npwq, igkq 200 call errore ('solve_linter', 'reading igkq', abs (ios) ) endif !Calculates beta functions (Kleinman-Bylander projectors), with !structure factor, for all atoms, in reciprocal space !HL the beta functions (vkb) are being generated properly. call init_us_2 (npwq, igkq, xk (1, ikq), vkb) !Reads unperturbed wavefuctions psi(k) and psi(k+q) if (nksq.gt.1) then if (lgamma) then call davcio (evc, lrwfc, iuwfc, ikk, - 1) else call davcio (evc, lrwfc, iuwfc, ikk, - 1) call davcio (evq, lrwfc, iuwfc, ikq, - 1) endif endif !IS TPA preconditioner better? do ig = 1, npwq g2kin (ig) = ( (xk (1,ikq) + g (1, igkq(ig)) ) **2 + & (xk (2,ikq) + g (2, igkq(ig)) ) **2 + & (xk (3,ikq) + g (3, igkq(ig)) ) **2 ) * tpiba2 enddo h_diag = 0.d0 do ibnd = 1, nbnd_occ (ikk) do ig = 1, npwq h_diag(ig,ibnd)= 1.d0/max(1.0d0,g2kin(ig)/eprec(ibnd,ik)) enddo IF (noncolin) THEN do ig = 1, npwq h_diag(ig+npwx,ibnd)=1.d0/max(1.0d0,g2kin(ig)/eprec(ibnd,ik)) ! h_diag(ig+npwx,ibnd)=1.d0/max(1.0d0,g2kin(ig) + (fiu/eprec(ibnd,ik)) enddo END IF enddo !HL indices freezing perturbations. mode = 1 nrec = ik !and now adds the contribution of the self consistent term if (where_rec =='solve_lint'.or.iter>1) then !HL is this necessary? ! After the first iteration dvbare_q*psi_kpoint is read from file call davcio (dvpsi, lrbar, iubar, nrec, - 1) ! calculates dvscf_q*psi_k in G_space, for all bands, k=kpoint ! dvscf_q from previous iteration (mix_potential) call start_clock ('vpsifft') do ibnd = 1, nbnd_occ (ikk) !FFT translated according to igk call cft_wave (evc (1, ibnd), aux1, +1) call apply_dpot (aux1, dvscfins(1,iw), current_spin) !FFT translated according to igkq: DeltaV(q)psi(k). call cft_wave (dvpsi (1, ibnd), aux1, -1) enddo call stop_clock ('vpsifft') !Need to check this for ultrasoft !HL THIS TERM PROBABLY NEEDS TO BE INCLUDED. !This is the variation in the D matrix due to the electric field. !call adddvscf (ipert, ik) else ! At the first iteration dvbare_q*psi_kpoint is calculated ! and written to file !call dvqpsi_us (dvbarein, ik, 1, .false.) call dvqpsi_us (dvbarein, ik, 1, .false.) call davcio (dvpsi, lrbar, iubar, nrec, +1) endif ! Orthogonalize dvpsi to valence states: ps = ! Apply -P_c^+. ! -P_c^ = - (1-P_v^): CALL orthogonalize(dvpsi, evq, ikk, ikq, dpsi) if (where_rec=='solve_lint'.or.iter > 1) then !starting value for delta_psi is read from iudwf nrec1 = ik !HL Don't need to read/write the full wave fxn at each iteration... ! call davcio ( dpsi, lrdwf, iudwf, nrec1, -1) call davcio ( dpsip, lrdwf, iudwfp, nrec1, -1) call davcio ( dpsim, lrdwf, iudwfm, nrec1, -1) ! dpsi(:,:) = (0.d0, 0.d0) ! dpsim(:,:) = (0.d0, 0.d0) ! dpsip(:,:) = (0.d0, 0.d0) !threshold for iterative solution of the linear system !write(6,*)1.d-1*sqrt(dr2), 1.d-4 thresh = min (1.d-1 * sqrt (dr2), 1.d-2) else ! ! At the first iteration dpsi and dvscfin are set to zero ! dpsi(:,:) = (0.d0, 0.d0) dpsim(:,:) = (0.d0, 0.d0) dpsip(:,:) = (0.d0, 0.d0) dvscfin(:, :) = (0.d0, 0.d0) dvscfout(:, :) = (0.d0, 0.d0) ! ! starting threshold for iterative solution of the linear system ! thresh = 1.0d-2 endif etc(:,:) = CMPLX( et(:,:), 0.0d0 , kind=DP) cw = fiu(iw) !HL should just use cgsolve_all when fiu(iw) = 0.0d0! !probably gain at least factor of two for static case. conv_root = .true. if(iw.eq.1) then call cgsolve_all (ch_psi_all, cg_psi, et(1,ikk), dvpsi, dpsip, h_diag, & npwx, npwq, thresh, ik, lter, conv_root, anorm, nbnd_occ(ikk), npol) dpsim(:,:) = dpsip(:,:) dpsi(:,:) = dcmplx(0.5d0,0.0d0)*(dpsim(:,:) + dpsip(:,:) ) else call cbcg_solve_fix(cch_psi_all_fix, cg_psi, etc(1,ikk), dvpsi, dpsip, h_diag, & npwx, npwq, thresh, ik, lter, conv_root, anorm, nbnd_occ(ikk), npol, cw, .true.) call cbcg_solve_fix(cch_psi_all_fix, cg_psi, etc(1,ikk), dvpsi, dpsim, h_diag, & npwx, npwq, thresh, ik, lter, conv_root, anorm, nbnd_occ(ikk), npol, -cw, .true.) dpsi(:,:) = dcmplx(0.5d0,0.0d0)*(dpsim(:,:) + dpsip(:,:) ) endif ltaver = ltaver + lter lintercall = lintercall + 1 if (.not.conv_root) WRITE(1000+mpime, '(5x,"kpoint",i4," ibnd",i4, & & " solve_linter: root not converged ",e10.3)') & & ik , ibnd, anorm nrec1 = ik call davcio (dpsim, lrdwf, iudwfm, nrec1, + 1) call davcio (dpsip, lrdwf, iudwfp, nrec1, + 1) ! calculates dvscf, sum over k => dvscf_q_ipert ! incdrhoscf: This routine computes the change of the charge density due to the ! perturbation. It is called at the end of the computation of the ! change of the wavefunction for a given k point. weight = wk (ikk) !IF (noncolin) THEN ! call incdrhoscf_nc(drhoscf(1,1),weight,ik, & ! dbecsum_nc(1,1,1,1), dpsi) !ELSE call incdrhoscf ( drhoscf(1,iw) , weight, ik, dbecsum(1,1,current_spin)) !ENDIF !should do non-collin spin as well!! !call incdrhoscf (drhoscf(1,current_spin,ipert), weight, ik, & ! dbecsum(1,1,current_spin,ipert), dpsi) enddo if (doublegrid) then do is = 1, nspin_mag call cinterpolate (drhoscfh(1,iw), drhoscf(1,iw), 1) enddo else call zcopy (nspin_mag*nrxx, drhoscf(1,iw), 1, drhoscfh(1,iw), 1) endif call addusddens (drhoscfh(1,iw), dbecsum, 0) call zcopy (nrxx*nspin_mag, drhoscfh(1,iw), 1, dvscfout(1,iw),1) ! SGW: here we enforce zero average variation of the charge density ! if the bare perturbation does not have a constant term ! (otherwise the numerical error, coupled with a small denominator ! in the coulomb term, gives rise to a spurious dvscf response) ! One wing of the dielectric matrix is particularly badly behaved meandvb = sqrt ((sum(dreal(dvbarein)))**2.d0 + (sum(aimag(dvbarein)))**2.d0 )/float(nrxxs) if (meandvb.lt.1.d-8) then call cft3 (dvscfout(1,iw), nr1, nr2, nr3, nrx1, nrx2, nrx3, -1) dvscfout ( nl(1), current_spin ) = (0.d0, 0.0d0) call cft3 (dvscfout(1,iw), nr1, nr2, nr3, nrx1, nrx2, nrx3, 1) endif call dv_of_drho (1, dvscfout(1,iw), .true.) if (iw.eq.1) then !Density reponse in real space should be real at zero freq no matter what! dvscfout(:,iw) = dcmplx(real(dvscfout(:,iw)), 0.0d0) !just using standard broyden for the zero freq. case. call mix_potential_real(2*nrxx*nspin_mag, dvscfout(1,iw), dvscfin(1,iw), alpha_mix(kter), & dr2, tr2_gw, iter, nmix_gw, flmixdpot, convt) else !for pure imaginary freqs. the density reponse should be real... do ir = 1, nrxx dvscfout(ir,iw) = dcmplx(real(dvscfout(ir,iw)), 0.0d0) enddo call mix_potential_c(nrxx, dvscfout(1,iw), dvscfin(1,iw), & alpha_mix(kter), dr2, tr2_gw, iter, & nmix_gw, convt) do ir = 1, nrxx dvscfin(ir,iw) = dcmplx(real(dvscfin(ir,iw)), 0.0d0) enddo endif if (doublegrid) then do ipert = 1, npe do is = 1, nspin_mag call cinterpolate (dvscfin(1,iw), dvscfins(1,iw), -1) enddo enddo endif ! with the new change of the potential we compute the integrals ! of the change of potential and Q ! HL-Q denotes the augmentation charge. Look at Vanderbilt PRB 41 7892 ! Q_{ij} = - ! USPP valence charge density is described ! n_v(r) = \sum_{n,k} \phi^{*}_{nk} (r) \phi_{nk}(r) + \sum_{i,j} p_{i,j}Q_{j,i} ! p_{i,j} = \sum_{n,k} #ifdef __PARA aux_avg (1) = DBLE (ltaver) aux_avg (2) = DBLE (lintercall) averlt = aux_avg (1) / aux_avg (2) #else averlt = DBLE (ltaver) / lintercall #endif tcpu = get_clock ('GW') dr2 = dr2 / DBLE(npe) ! WRITE( 1000+mpime, '(/,5x," iter # ",i3," total cpu time :",f8.1, & ! " secs av.it.: ",f5.1)') iter, tcpu, averlt ! WRITE( 1000+mpime, '(5x," thresh=",e10.3, " alpha_mix = ",f6.3, & ! " |ddv_scf|^2 = ",e10.3 )') thresh, alpha_mix (kter) , dr2 CALL flush_unit( stdout ) rec_code=10 if (convt) goto 155 enddo !loop on kter (iterations) 155 iter0=0 ! WRITE( stdout, '(/,5x," iter # ",i3," total cpu time :",f8.1, & ! "secs av.it.:",f5.1)') iter, tcpu, averlt ! WRITE(1000+mpime, '(/,5x," iter # ",i3," total cpu time :",f8.1, & ! "secs av.it.: ",f5.1)') iter, tcpu, averlt drhoscf(:,iw) = dvscfin(:,iw) if (convt) then if (fildvscf.ne.' ') then write(6, '("fildvscf")') end if endif deallocate (h_diag) deallocate (aux1) deallocate (dbecsum) IF (noncolin) deallocate (dbecsum_nc) deallocate (dvscfout) deallocate (drhoscfh) if (doublegrid) deallocate (dvscfins) deallocate (dvscfin) deallocate (dpsi) call stop_clock ('solve_linter') END SUBROUTINE solve_linter SUBROUTINE setmixout(in1, in2, mix, dvscfout, dbecsum, ndim, flag ) USE kinds, ONLY : DP USE mp_global, ONLY : intra_pool_comm USE mp, ONLY : mp_sum IMPLICIT NONE INTEGER :: in1, in2, flag, ndim, startb, lastb COMPLEX(DP) :: mix(in1+in2), dvscfout(in1), dbecsum(in2) CALL divide (in2, startb, lastb) ndim=lastb-startb+1 IF (flag==-1) THEN mix(1:in1)=dvscfout(1:in1) mix(in1+1:in1+ndim)=dbecsum(startb:lastb) ELSE dvscfout(1:in1)=mix(1:in1) dbecsum=(0.0_DP,0.0_DP) dbecsum(startb:lastb)=mix(in1+1:in1+ndim) #ifdef __PARA CALL mp_sum(dbecsum, intra_pool_comm) #endif ENDIF END SUBROUTINE setmixout