def get_jk(mol_or_mf=None, dm=None, hermi=1, with_j=True, with_k=True, omega=None): '''MPI version of scf.hf.get_jk function''' #vj = get_j(mol_or_mf, dm, hermi) #vk = get_k(mol_or_mf, dm, hermi) if isinstance(mol_or_mf, gto.mole.Mole): mf = hf.SCF(mol_or_mf).view(SCF) else: mf = mol_or_mf # dm may be too big for mpi4py library to serialize. Broadcast dm here. if any(comm.allgather(dm is mpi.Message.SkippedArg)): dm = mpi.bcast_tagged_array(dm) mf.unpack_(comm.bcast(mf.pack())) if mf.opt is None: mf.opt = mf.init_direct_scf() if omega is None: vj, vk = _eval_jk(mf, dm, hermi, _jk_jobs_s8) else: with mf.mol.with_range_coulomb(omega): vj, vk = _eval_jk(mf, dm, hermi, _jk_jobs_s8) if rank == 0: for i in range(vj.shape[0]): lib.hermi_triu(vj[i], 1, inplace=True) return vj.reshape(dm.shape), vk.reshape(dm.shape)
def get_k(mol_or_mf, dm, hermi=1): if isinstance(mol_or_mf, gto.mole.Mole): mf = hf.SCF(mol_or_mf).view(SCF) else: mf = mol_or_mf # dm may be too big for mpi4py library to serialize. Broadcast dm here. if any(comm.allgather(isinstance(dm, str) and dm == 'SKIPPED_ARG')): dm = mpi.bcast_tagged_array(dm) mf.unpack_(comm.bcast(mf.pack())) if mf.opt is None: mf.opt = mf.init_direct_scf() with lib.temporary_env(mf.opt._this.contents, fprescreen=_vhf._fpointer('CVHFnrs8_vk_prescreen')): vk = _eval_jk(mf, dm, hermi, _vk_jobs_s8) return vk.reshape(dm.shape)
def get_k(mol_or_mf=None, dm=None, hermi=1, omega=None): if isinstance(mol_or_mf, gto.mole.Mole): mf = hf.SCF(mol_or_mf).view(SCF) else: mf = mol_or_mf # dm may be too big for mpi4py library to serialize. Broadcast dm here. if any(comm.allgather(dm is mpi.Message.SkippedArg)): dm = mpi.bcast_tagged_array(dm) mf.unpack_(comm.bcast(mf.pack())) if mf.opt is None: mf.opt = mf.init_direct_scf() with lib.temporary_env(mf.opt._this.contents, fprescreen=_vhf._fpointer('CVHFnrs8_vk_prescreen')): if omega is None: vk = _eval_jk(mf, dm, hermi, _vk_jobs_s8) else: with mf.mol.with_range_coulomb(omega): vk = _eval_jk(mf, dm, hermi, _vk_jobs_s8) return vk.reshape(dm.shape)
def get_jk(mol_or_mf, dm, hermi=1): '''MPI version of scf.hf.get_jk function''' #vj = get_j(mol_or_mf, dm, hermi) #vk = get_k(mol_or_mf, dm, hermi) if isinstance(mol_or_mf, gto.mole.Mole): mf = hf.SCF(mol_or_mf).view(SCF) else: mf = mol_or_mf # dm may be too big for mpi4py library to serialize. Broadcast dm here. if any(comm.allgather(isinstance(dm, str) and dm == 'SKIPPED_ARG')): dm = mpi.bcast_tagged_array(dm) mf.unpack_(comm.bcast(mf.pack())) if mf.opt is None: mf.opt = mf.init_direct_scf() vj, vk = _eval_jk(mf, dm, hermi, _jk_jobs_s8) if rank == 0: for i in range(vj.shape[0]): lib.hermi_triu(vj[i], 1, inplace=True) return vj.reshape(dm.shape), vk.reshape(dm.shape)
def get_veff(mf, mol=None, dm=None, dm_last=0, vhf_last=0, hermi=1): t0 = (time.clock(), time.time()) mf.unpack_(comm.bcast(mf.pack())) mol = mf.mol ni = mf._numint if mf.nlc != '': raise NotImplementedError omega, alpha, hyb = ni.rsh_and_hybrid_coeff(mf.xc, spin=mol.spin) if abs(omega) > 1e-10: # For range separated Coulomb operator raise NotImplementedError # Broadcast the large input arrays here. if any(comm.allgather(isinstance(dm, str) and dm == 'SKIPPED_ARG')): if rank == 0 and dm is None: dm = mf.make_rdm1() dm = mpi.bcast_tagged_array(dm) if any(comm.allgather(isinstance(dm_last, str) and dm_last == 'SKIPPED_ARG')): dm_last = mpi.bcast_tagged_array(dm_last) if any(comm.allgather(isinstance(vhf_last, str) and vhf_last == 'SKIPPED_ARG')): vhf_last = mpi.bcast_tagged_array(vhf_last) ground_state = (isinstance(dm, numpy.ndarray) and dm.ndim == 2) if mf.grids.coords is None: _setup_grids_(mf, dm) t0 = logger.timer(mf, 'setting up grids', *t0) if hermi == 2: # because rho = 0 n, exc, vxc = 0, 0, 0 else: n, exc, vxc = ni.nr_rks(mol, mf.grids, mf.xc, dm) n = comm.allreduce(n) exc = comm.allreduce(exc) vxc = mpi.reduce(vxc) logger.debug(mf, 'nelec by numeric integration = %s', n) t0 = logger.timer(mf, 'vxc', *t0) if abs(hyb) < 1e-10 and abs(alpha) < 1e-10: vk = None if getattr(vhf_last, 'vj', None) is not None: ddm = numpy.asarray(dm) - dm_last vj = mpi.reduce(mf.get_j(mol, ddm, hermi)) vj += vhf_last.vj else: vj = mf.get_j(mol, dm, hermi) vj = mpi.reduce(vj) vxc += vj else: if getattr(vhf_last, 'vk', None) is not None: ddm = numpy.asarray(dm) - dm_last vj, vk = mf.get_jk(mol, ddm, hermi) ddm = None vj = mpi.reduce(vj) vk = mpi.reduce(vk) * hyb vj += vhf_last.vj vk += vhf_last.vk else: vj, vk = mf.get_jk(mol, dm, hermi) vj = mpi.reduce(vj) vk = mpi.reduce(vk) * hyb vxc += vj - vk * .5 if ground_state: exc -= numpy.einsum('ij,ji', dm, vk) * .5 * .5 if ground_state: ecoul = numpy.einsum('ij,ji', dm, vj) * .5 else: ecoul = None vxc = lib.tag_array(vxc, ecoul=ecoul, exc=exc, vj=vj, vk=vk) return vxc
def get_veff(mf, mol=None, dm=None, dm_last=0, vhf_last=0, hermi=1): t0 = (logger.process_clock(), logger.perf_counter()) mf.unpack_(comm.bcast(mf.pack())) mol = mf.mol ni = mf._numint if mf.nlc != '': raise NotImplementedError omega, alpha, hyb = ni.rsh_and_hybrid_coeff(mf.xc, spin=mol.spin) # Broadcast the large input arrays here. if any(comm.allgather(dm is mpi.Message.SkippedArg)): if rank == 0 and dm is None: dm = mf.make_rdm1() dm = mpi.bcast_tagged_array(dm) if any(comm.allgather(dm_last is mpi.Message.SkippedArg)): dm_last = mpi.bcast_tagged_array(dm_last) if any(comm.allgather(vhf_last is mpi.Message.SkippedArg)): vhf_last = mpi.bcast_tagged_array(vhf_last) ground_state = (dm.ndim == 3 and dm.shape[0] == 2) if mf.grids.coords is None: mpi_rks._setup_grids_(mf, dm[0]+dm[1]) t0 = logger.timer(mf, 'setting up grids', *t0) if hermi == 2: # because rho = 0 n, exc, vxc = 0, 0, 0 else: n, exc, vxc = ni.nr_uks(mol, mf.grids, mf.xc, dm) n = comm.allreduce(n) exc = comm.allreduce(exc) vxc = mpi.reduce(vxc) logger.debug(mf, 'nelec by numeric integration = %s', n) t0 = logger.timer(mf, 'vxc', *t0) if abs(hyb) < 1e-10 and abs(alpha) < 1e-10: vk = None if getattr(vhf_last, 'vj', None) is not None: ddm = numpy.asarray(dm) - dm_last ddm = ddm[0] + ddm[1] vj = mf.get_j(mol, ddm, hermi) vj += vhf_last.vj else: vj = mf.get_j(mol, dm[0]+dm[1], hermi) vxc += vj else: if getattr(vhf_last, 'vk', None) is not None: ddm = numpy.asarray(dm) - dm_last vj, vk = mf.get_jk(mol, ddm, hermi) vj = vj[0] + vj[1] vk *= hyb if abs(omega) > 1e-10: vklr = mf.get_k(mol, ddm, hermi, omega=omega) vk += vklr * (alpha - hyb) ddm = None vj += vhf_last.vj vk += vhf_last.vk else: vj, vk = mf.get_jk(mol, dm, hermi) vj = vj[0] + vj[1] vk *= hyb if abs(omega) > 1e-10: vklr = mf.get_k(mol, dm, hermi, omega=omega) vk += vklr * (alpha - hyb) vxc += vj vxc -= vk if ground_state: exc -=(numpy.einsum('ij,ji', dm[0], vk[0]) + numpy.einsum('ij,ji', dm[1], vk[1])) * .5 if ground_state: ecoul = numpy.einsum('ij,ji', dm[0]+dm[1], vj) * .5 else: ecoul = None vxc = lib.tag_array(vxc, ecoul=ecoul, exc=exc, vj=vj, vk=vk) return vxc
def get_jk(mol_or_mf, dm, hermi, dmcur): #, *args, **kwargs '''MPI version of scf.hf.get_jk function''' #vj = get_j(mol_or_mf, dm, hermi) #vk = get_k(mol_or_mf, dm, hermi) if isinstance(mol_or_mf, gto.mole.Mole): mf = hf.SCF(mol_or_mf).view(SCF) else: mf = mol_or_mf # dm may be too big for mpi4py library to serialize. Broadcast dm here. if any(comm.allgather(isinstance(dm, str) and dm == 'SKIPPED_ARG')): #dm = mpi.bcast_tagged_array_occdf(dm) dm = mpi.bcast_tagged_array(dm) mf.unpack_(comm.bcast(mf.pack())) # initial and final grids level grdlvl_i = 0 grdlvl_f = 1 # norm_ddm threshold for grids change thrd_nddm = 0.01 # set block size to adapt memory sblk = 100 # interspace betweeen v shell intsp = 1 # threshold for u and v gthrdu = 1e-7 gthrdvs = 1e-4 gthrdvd = 1e-4 global cond, wao_vx, ngridsx, coordsx, weightsx dms = numpy.asarray(dm) dm_shape = dms.shape nao = dm_shape[-1] dms = dms.reshape(-1,nao,nao) nset = dms.shape[0] vj = [0] * nset vk = [0] * nset # DF-J and sgX set mf.with_df = mf mol = mf.mol global int2c, ovlp, ao_loc, rao_loc, ov_scr # need set mf.initsgx = None in scf.SCF __init_ if mf.initsgx is None: mf.initsgx = 0 cond = 0 # set auxbasis in input file, need self.auxbasis = None in __init__ of hf.py # mf.auxbasis = 'weigend' auxbasis = mf.auxbasis auxbasis = comm.bcast(auxbasis) mf.auxbasis = comm.bcast(mf.auxbasis) auxmol = df.addons.make_auxmol(mol, auxbasis) # (P|Q) int2c = auxmol.intor('int2c2e', aosym='s1', comp=1) if rank == 0: print('auxmol.basis',auxmol.basis,'number of aux basis',int2c.shape[0]) # for sgX # ao_loc and rao_loc intbn = mol._add_suffix('int3c2e') intbn = gto.moleintor.ascint3(intbn) ao_loc = gto.moleintor.make_loc(mol._bas, intbn) rao_loc = numpy.zeros((nao),dtype=int) ovlp = mol.intor_symmetric('int1e_ovlp') for i in range(mol.nbas): for jj in range(ao_loc[i],ao_loc[i+1]): rao_loc[jj] = i # ovlp screening ov_scr = numpy.zeros((mol.nbas,mol.nbas),dtype=int) for i in range(mol.nbas): for j in range(mol.nbas): if mol.bas_atom(i) == mol.bas_atom(j): ov_scr[i,j] = 1 else: movlp = numpy.amax(numpy.absolute(ovlp[ao_loc[i]:ao_loc[i+1],ao_loc[j]:ao_loc[j+1]])) if movlp > gthrdvs: ov_scr[i,j] = 1 if rank == 0: print('thrd_nddm',thrd_nddm, 'sblk',sblk, 'intsp',intsp, 'gthrdu',gthrdu) if rank == 0: print('gthrdvs',gthrdvs, 'gthrdvd',gthrdvd) # coase and fine grids change grdchg = 0 norm_ddm = 0 for k in range(nset): norm_ddm += numpy.linalg.norm(dms[k]) if norm_ddm < thrd_nddm and cond == 2 : cond = 1 if cond == 0: wao_vx, ngridsx, coordsx, weightsx = get_gridss(mol,grdlvl_i, sblk) if rank == 0: print('grids level at first is', grdlvl_i) cond = 2 elif cond == 1: wao_vx, ngridsx, coordsx, weightsx = get_gridss(mol,grdlvl_f, sblk) if rank == 0: print('grids level change to', grdlvl_f) dms = numpy.asarray(dmcur) dms = dms.reshape(-1,nao,nao) grdchg = 1 cond = 3 if rank==0: Jtime=time.time() # DF-J dmtril = [] for k in range(nset): dmtril.append(lib.pack_tril(dms[k]+dms[k].T)) i = numpy.arange(nao) dmtril[k][i*(i+1)//2+i] *= .5 rho = [] b0 = 0 for eri1 in loop(mf.with_df): naux, nao_pair = eri1.shape #if rank==0: print('slice-naux',naux,'rank',rank) b1 = b0 + naux assert(nao_pair == nao*(nao+1)//2) for k in range(nset): if b0 == 0: rho.append(numpy.empty(paux[rank])) rho[k][b0:b1] = numpy.dot(eri1, dmtril[k]) b0 = b1 orho = [] rec = [] for k in range(nset): orho.append(mpi.gather(rho[k])) if rank == 0: ivj0 = scipy.linalg.solve(int2c, orho[k]) else: ivj0 = None rec.append(numpy.empty(paux[rank])) comm.Scatterv([ivj0,paux],rec[k],root=0) b0 = 0 for eri1 in loop(mf.with_df): naux, nao_pair = eri1.shape b1 = b0 + naux assert(nao_pair == nao*(nao+1)//2) for k in range(nset): vj[k] += numpy.dot(rec[k][b0:b1].T, eri1) b0 = b1 for k in range(nset): vj[k] = comm.reduce(vj[k]) if rank==0: print "Took this long for J: ", time.time()-Jtime if rank==0: Jtime=time.time() # sgX wao_v = wao_vx coords = coordsx weights = weightsx for k in range(nset): '''# Plus density screening ov_scr = numpy.zeros((mol.nbas,mol.nbas),dtype=int) for i in range(mol.nbas): for j in range(mol.nbas): movlp = numpy.amax(numpy.absolute(dms[k][ao_loc[i]:ao_loc[i+1],ao_loc[j]:ao_loc[j+1]])) if movlp > gthrdvd: ov_scr[i,j] = 1''' # xfg screening ngrids = coords.shape[0] blksize = min(ngrids, sblk) gscr = [] for i0, i1 in lib.prange(0, ngrids, blksize): # screening u by value of grids umaxg = numpy.amax(numpy.absolute(wao_v[i0:i1]), axis=0) usi = numpy.argwhere(umaxg > gthrdu).reshape(-1) if len(usi) != 0: # screening v by ovlp then triangle matrix bn uovl = ovlp[usi, :] vmaxu = numpy.amax(numpy.absolute(uovl), axis=0) osi = numpy.argwhere(vmaxu > gthrdvs).reshape(-1) udms = dms[k][usi, :] # screening v by dm and ovlp then triangle matrix bn dmaxg = numpy.amax(numpy.absolute(udms), axis=0) dsi = numpy.argwhere(dmaxg > gthrdvd).reshape(-1) vsi = numpy.intersect1d(dsi, osi) if len(vsi) != 0: vsh = numpy.unique(rao_loc[vsi]) vshi = [] for i in range(vsh.shape[0]): ista = ao_loc[vsh[i]] iend = ao_loc[vsh[i]+1] vshi.append(numpy.arange(ista, iend)) vshi = numpy.asarray(numpy.hstack(vshi)) dmsi = dms[k][usi] fg = weights[i0:i1,None] * numpy.dot(wao_v[i0:i1,usi],dmsi[:,vshi]) gmaxfg = numpy.amax(numpy.absolute(fg), axis=1) gmaxwao_v = numpy.amax(numpy.absolute(wao_v[i0:i1,usi]), axis=1) gmaxtt = gmaxfg * gmaxwao_v gscr0 = numpy.argwhere(gmaxtt > gthrdu).reshape(-1) if gscr0.shape[0] > 0: gscr.append(gscr0 + i0) hgscr = numpy.hstack(gscr).reshape(-1) coords = mpi.gather(coords[hgscr]) wao_v = mpi.gather(wao_v[hgscr]) weights = mpi.gather(weights[hgscr]) if rank == 0: print('screened grids', coords.shape[0]) coords = numpy.array_split(coords, mpi.pool.size) wao_v = numpy.array_split(wao_v, mpi.pool.size) weights = numpy.array_split(weights, mpi.pool.size) coords = mpi.scatter(coords) wao_v = mpi.scatter(wao_v) weights = mpi.scatter(weights) # Kuv = Sum(Xug Avt Dkt Xkg) ngrids = coords.shape[0] for i0, i1 in lib.prange(0, ngrids, blksize): # screening u by value of grids umaxg = numpy.amax(numpy.absolute(wao_v[i0:i1]), axis=0) usi = numpy.argwhere(umaxg > gthrdu).reshape(-1) if len(usi) != 0: # screening v by ovlp then triangle matrix bn uovl = ovlp[usi, :] vmaxu = numpy.amax(numpy.absolute(uovl), axis=0) osi = numpy.argwhere(vmaxu > gthrdvs).reshape(-1) udms = dms[k][usi, :] # screening v by dm and ovlp then triangle matrix bn dmaxg = numpy.amax(numpy.absolute(udms), axis=0) dsi = numpy.argwhere(dmaxg > gthrdvd).reshape(-1) vsi = numpy.intersect1d(dsi, osi) if len(vsi) != 0: vsh = numpy.unique(rao_loc[vsi]) nvsh = vsh.shape[0] vov0 = ov_scr[vsh] vov = vov0[:,vsh] vshi = [] xvsh = vsh ivx = [0] vx = 0 for i in range(vsh.shape[0]): ista = ao_loc[vsh[i]] iend = ao_loc[vsh[i]+1] vshi.append(numpy.arange(ista, iend)) vx += iend - ista ivx.append(vx) vshi = numpy.asarray(numpy.hstack(vshi)) nvshi = vshi.shape[0] #print('ee',nvshi) ivx = numpy.asarray(ivx) vshbeg = vsh[0] vshfin = vsh[-1]+1 dmsi = dms[k][usi] fg = weights[i0:i1,None] * numpy.dot(wao_v[i0:i1,usi],dmsi[:,vshi]) fakemol = gto.fakemol_for_charges(coords[i0:i1]) #pmol = gto.mole.conc_mol(mol, fakemol) intor = mol._add_suffix('int3c2e') atm, bas, env = gto.mole.conc_env(mol._atm, mol._bas, mol._env, fakemol._atm, fakemol._bas, fakemol._env) shls_slice = (vshbeg, vshfin, vshbeg, vshfin, mol.nbas, mol.nbas+fakemol.nbas) comp=1 #aosym='s1' aosym='s2ij' if aosym == 's2ij': gv = getints3c_scr(intor, atm, bas, env, shls_slice, comp, xvsh, nvshi, ivx, vov, fg, aosym) else: gv = getints3c_scr(intor, atm, bas, env, shls_slice, comp, xvsh, nvshi, ivx, vov, fg, aosym) vk0 = numpy.zeros((nao,nao)) vksp = lib.einsum('gu,gv->uv', wao_v[i0:i1,usi], gv) vk1 = vk0[usi] vk1[:,vshi] = vksp vk0[usi] = vk1 vk[k] += vk0 wao_vw = weights[:,None] * wao_v sn = lib.einsum('gu,gv->uv', wao_v, wao_vw) vk[k] = comm.reduce(vk[k]) sn = comm.reduce(sn) # SSn^-1 for grids to analitic if rank == 0: snsgk = scipy.linalg.solve(sn, vk[k]) vk[k] = numpy.matmul(ovlp, snsgk) if hermi == 1: vk[k] = (vk[k] + vk[k].T)*.5 if rank == 0: print "Took this long for K: ", time.time()-Jtime vj = lib.unpack_tril(numpy.asarray(vj), 1).reshape(dm_shape) vk = numpy.asarray(vk).reshape(dm_shape) return vj, vk, grdchg