def __init__(self, gd, bd, block_comm, dtype, nao, timer=nulltimer):
KohnShamLayouts.__init__(self, gd, bd, block_comm, dtype, timer)
self.mMdescriptor = MatrixDescriptor(nao, nao)
self.nMdescriptor = MatrixDescriptor(bd.mynbands, nao)
self.Mstart = 0
self.Mstop = nao
self.Mmax = nao
self.mynao = nao
self.nao = nao
self.orbital_comm = bd.comm
def __init__(self, blacsgrid, M, N, mb, nb, rsrc=0, csrc=0):
assert M > 0
assert N > 0
assert 1 <= mb
assert 1 <= nb
if mb > M:
mb = M
if nb > N:
nb = N
assert 0 <= rsrc < blacsgrid.nprow
assert 0 <= csrc < blacsgrid.npcol
self.blacsgrid = blacsgrid
self.M = M # global size 1
self.N = N # global size 2
self.mb = mb # block cyclic distr dim 1
self.nb = nb # and 2. How many rows or columns are on this processor
# more info:
# http://www.netlib.org/scalapack/slug/node75.html
self.rsrc = rsrc
self.csrc = csrc
if blacsgrid.is_active():
locN, locM = _gpaw.get_blacs_local_shape(self.blacsgrid.context,
self.N, self.M, self.nb,
self.mb, self.csrc,
self.rsrc)
# max 1 is nonsensical, but appears
# to be required by PBLAS
self.lld = max(1, locN)
else:
# ScaLAPACK has no requirements as to what these values on an
# inactive blacsgrid should be. This seemed reasonable to me
# at the time.
locN, locM = 0, 0
# This applies only to empty matrices. Some functions will
# complain about lld<1, so lld=1 will make those happy.
self.lld = 1
# locM, locN is not allowed to be negative. This will cause the
# redistributor to fail. This could happen on active blacsgrid
# which does not contain any piece of the distribute matrix.
# This is why there is a final check on the value of locM, locN.
MatrixDescriptor.__init__(self, max(0, locM), max(0, locN))
# This is the definition of inactive descriptor; can occur
# on an active or inactive blacs grid.
self.active = locM > 0 and locN > 0
self.bshape = (self.mb, self.nb) # Shape of one block
self.gshape = (M, N) # Global shape of array
def __init__(self, blacsgrid, M, N, mb, nb, rsrc, csrc):
assert M > 0
assert N > 0
assert 1 <= mb
assert 1 <= nb
if mb > M:
mb = M
if nb > N:
nb = N
assert 0 <= rsrc < blacsgrid.nprow
assert 0 <= csrc < blacsgrid.npcol
self.blacsgrid = blacsgrid
self.M = M # global size 1
self.N = N # global size 2
self.mb = mb # block cyclic distr dim 1
self.nb = nb # and 2. How many rows or columns are on this processor
# more info:
# http://www.netlib.org/scalapack/slug/node75.html
self.rsrc = rsrc
self.csrc = csrc
if blacsgrid.is_active():
locN, locM = _gpaw.get_blacs_local_shape(self.blacsgrid.context,
self.N, self.M,
self.nb, self.mb,
self.csrc, self.rsrc)
self.lld = max(1, locN) # max 1 is nonsensical, but appears
# to be required by PBLAS
else:
# ScaLAPACK has no requirements as to what these values on an
# inactive blacsgrid should be. This seemed reasonable to me
# at the time.
locN, locM = 0, 0
self.lld = 0
# locM, locN is not allowed to be negative. This will cause the
# redistributor to fail. This could happen on active blacsgrid
# which does not contain any piece of the distribute matrix.
# This is why there is a final check on the value of locM, locN.
MatrixDescriptor.__init__(self, max(0, locM), max(0, locN))
# This is the definition of inactive descriptor; can occur
# on an active or inactive blacs grid.
self.active = locM > 0 and locN > 0
self.bshape = (self.mb, self.nb) # Shape of one block
self.gshape = (M, N) # Global shape of array
def __init__(self, gd, bd, dtype, nao, timer=nulltimer):
KohnShamLayouts.__init__(self, gd, bd, dtype, timer)
self.mMdescriptor = MatrixDescriptor(nao, nao)
self.nMdescriptor = MatrixDescriptor(bd.mynbands, nao)
self.Mstart = 0
self.Mstop = nao
self.Mmax = nao
self.mynao = nao
self.nao = nao
# a problem with forces when
# domain-decomposition = (1, 1, 1)
# bd.comm.size > 1
# kpt.comm.size > 1
# Uncomment line 145 in gpaw/test/parallel/lcao_parallel.py
self.orbital_comm = serial_comm
def iterate_one_k_point(self, ham, wfs, kpt):
"""Setup H and S matries and diagonalize for the kpoint"""
self.timer.start('DirectPW')
H_GG, S_GG = wfs.hs(ham, kpt.q, kpt.s)
npw = len(H_GG)
eps_n = np.empty(npw)
md = MatrixDescriptor(npw, npw)
psit_nG = md.empty(dtype=complex)
md.general_diagonalize_dc(H_GG, S_GG, psit_nG, eps_n,
iu=wfs.bd.nbands)
kpt.eps_n[:] = eps_n[:wfs.bd.nbands]
wfs.pt.integrate(kpt.psit_nG, kpt.P_ani, kpt.q)
self.timer.stop('DirectPW')
error = 0.0
return error
def diagonalize_full_hamiltonian(self, ham, atoms, occupations, txt,
nbands=None, scalapack=None, expert=False):
assert self.dtype == complex
if nbands is None:
nbands = self.pd.ngmin // self.bd.comm.size * self.bd.comm.size
else:
assert nbands <= self.pd.ngmin
if expert:
iu = nbands
else:
iu = None
self.bd = bd = BandDescriptor(nbands, self.bd.comm)
p = functools.partial(print, file=txt)
p('Diagonalizing full Hamiltonian ({0} lowest bands)'.format(nbands))
p('Matrix size (min, max): {0}, {1}'.format(self.pd.ngmin,
self.pd.ngmax))
mem = 3 * self.pd.ngmax**2 * 16 / bd.comm.size / 1024**2
p('Approximate memory usage per core: {0:.3f} MB'.format(mem))
if bd.comm.size > 1:
if isinstance(scalapack, (list, tuple)):
nprow, npcol, b = scalapack
else:
nprow = int(round(bd.comm.size**0.5))
while bd.comm.size % nprow != 0:
nprow -= 1
npcol = bd.comm.size // nprow
b = 64
p('ScaLapack grid: {0}x{1},'.format(nprow, npcol),
'block-size:', b)
bg = BlacsGrid(bd.comm, bd.comm.size, 1)
bg2 = BlacsGrid(bd.comm, nprow, npcol)
scalapack = True
else:
nprow = npcol = 1
scalapack = False
self.pt.set_positions(atoms.get_scaled_positions())
self.kpt_u[0].P_ani = None
self.allocate_arrays_for_projections(self.pt.my_atom_indices)
myslice = bd.get_slice()
pb = ProgressBar(txt)
nkpt = len(self.kpt_u)
for u, kpt in enumerate(self.kpt_u):
pb.update(u / nkpt)
npw = len(self.pd.Q_qG[kpt.q])
if scalapack:
mynpw = -(-npw // bd.comm.size)
md = BlacsDescriptor(bg, npw, npw, mynpw, npw)
md2 = BlacsDescriptor(bg2, npw, npw, b, b)
else:
md = md2 = MatrixDescriptor(npw, npw)
with self.timer('Build H and S'):
H_GG, S_GG = self.hs(ham, kpt.q, kpt.s, md)
if scalapack:
r = Redistributor(bd.comm, md, md2)
H_GG = r.redistribute(H_GG)
S_GG = r.redistribute(S_GG)
psit_nG = md2.empty(dtype=complex)
eps_n = np.empty(npw)
with self.timer('Diagonalize'):
if not scalapack:
md2.general_diagonalize_dc(H_GG, S_GG, psit_nG, eps_n,
iu=iu)
else:
md2.general_diagonalize_dc(H_GG, S_GG, psit_nG, eps_n)
del H_GG, S_GG
kpt.eps_n = eps_n[myslice].copy()
if scalapack:
md3 = BlacsDescriptor(bg, npw, npw, bd.mynbands, npw)
r = Redistributor(bd.comm, md2, md3)
psit_nG = r.redistribute(psit_nG)
kpt.psit_nG = psit_nG[:bd.mynbands].copy()
del psit_nG
with self.timer('Projections'):
self.pt.integrate(kpt.psit_nG, kpt.P_ani, kpt.q)
kpt.f_n = None
pb.finish()
occupations.calculate(self)
def diagonalize_full_hamiltonian(self, ham, atoms, occupations, log,
nbands=None, ecut=None, scalapack=None,
expert=False):
if self.dtype != complex:
raise ValueError('Your wavefunctions are not complex as '
'required by the PW diagonalization routine.\n'
'Please supply GPAW(..., dtype=complex, ...) '
'as an argument to the calculator to enforce '
'complex wavefunctions.')
if nbands is None and ecut is None:
nbands = self.pd.ngmin // self.bd.comm.size * self.bd.comm.size
elif nbands is None:
ecut /= units.Hartree
vol = abs(np.linalg.det(self.gd.cell_cv))
nbands = int(vol * ecut**1.5 * 2**0.5 / 3 / pi**2)
else:
assert nbands <= self.pd.ngmin
if expert:
iu = nbands
else:
iu = None
self.bd = bd = BandDescriptor(nbands, self.bd.comm)
log('Diagonalizing full Hamiltonian ({0} lowest bands)'.format(nbands))
log('Matrix size (min, max): {0}, {1}'.format(self.pd.ngmin,
self.pd.ngmax))
mem = 3 * self.pd.ngmax**2 * 16 / bd.comm.size / 1024**2
log('Approximate memory usage per core: {0:.3f} MB'.format(mem))
if bd.comm.size > 1:
if isinstance(scalapack, (list, tuple)):
nprow, npcol, b = scalapack
else:
nprow = int(round(bd.comm.size**0.5))
while bd.comm.size % nprow != 0:
nprow -= 1
npcol = bd.comm.size // nprow
b = 64
log('ScaLapack grid: {0}x{1},'.format(nprow, npcol),
'block-size:', b)
bg = BlacsGrid(bd.comm, bd.comm.size, 1)
bg2 = BlacsGrid(bd.comm, nprow, npcol)
scalapack = True
else:
nprow = npcol = 1
scalapack = False
self.set_positions(atoms.get_scaled_positions())
self.kpt_u[0].P_ani = None
self.allocate_arrays_for_projections(self.pt.my_atom_indices)
myslice = bd.get_slice()
pb = ProgressBar(log.fd)
nkpt = len(self.kpt_u)
for u, kpt in enumerate(self.kpt_u):
pb.update(u / nkpt)
npw = len(self.pd.Q_qG[kpt.q])
if scalapack:
mynpw = -(-npw // bd.comm.size)
md = BlacsDescriptor(bg, npw, npw, mynpw, npw)
md2 = BlacsDescriptor(bg2, npw, npw, b, b)
else:
md = md2 = MatrixDescriptor(npw, npw)
with self.timer('Build H and S'):
H_GG, S_GG = self.hs(ham, kpt.q, kpt.s, md)
if scalapack:
r = Redistributor(bd.comm, md, md2)
H_GG = r.redistribute(H_GG)
S_GG = r.redistribute(S_GG)
psit_nG = md2.empty(dtype=complex)
eps_n = np.empty(npw)
with self.timer('Diagonalize'):
if not scalapack:
md2.general_diagonalize_dc(H_GG, S_GG, psit_nG, eps_n,
iu=iu)
else:
md2.general_diagonalize_dc(H_GG, S_GG, psit_nG, eps_n)
del H_GG, S_GG
kpt.eps_n = eps_n[myslice].copy()
if scalapack:
md3 = BlacsDescriptor(bg, npw, npw, bd.mynbands, npw)
r = Redistributor(bd.comm, md2, md3)
psit_nG = r.redistribute(psit_nG)
kpt.psit_nG = psit_nG[:bd.mynbands].copy()
del psit_nG
with self.timer('Projections'):
self.pt.integrate(kpt.psit_nG, kpt.P_ani, kpt.q)
kpt.f_n = None
pb.finish()
occupations.calculate(self)
return nbands