def _run(self, **kwargs):
if self.with_df:
self._pyscf = self._pyscf.density_fit()
self._pyscf.with_df.auxbasis = self.auxbasis
self._pyscf.run(**kwargs)
if self.check_stability:
for niter in range(1, self.stability_cycles + 1):
stability = self.stability()
if isinstance(stability, tuple):
internal, external = stability
else:
internal = stability
if np.allclose(internal, self._pyscf.mo_coeff):
if niter == self.stability_cycles:
if mpi.rank:
log.warn('Internal stability in HF not resolved.')
break
else:
rdm1 = self._pyscf.make_rdm1(internal, self._pyscf.mo_occ)
self._pyscf.scf(dm0=rdm1)
if not self.with_df:
self._eri_ao = util.restore(1, self.mol._pyscf.intor('int2e'),
self.nao)
else:
self._eri_ao = lib.unpack_tril(self._pyscf.with_df._cderi)
def dgemm(a, b, c=None, alpha=1.0, beta=0.0):
''' Performs dgemm in Fortran memory alignment without copying.
Input matrix should be contiguous (either F or C).
Parameters
----------
a : array
input matrix a
b : array
input matrix b
c : array, optional
output matrix c, if None then it is allocated inside the
function, default None
alpha : float, optional
scalar factor for matrix a
beta : float, optional
scalar factor for matrix c
Returns
-------
c : ndarray
output matrix
'''
#FIXME: this needs testing better - currently this should only be
# used where it has been tested against np.dot specifically for
# that use case!
if (not _is_contiguous(a)) or (not _is_contiguous(b)):
log.warn('DGEMM called on non-contiguous data')
m, k = a.shape
n = b.shape[1]
assert k == b.shape[0]
a, ta = _reorder_fortran(a)
b, tb = _reorder_fortran(b)
if c is None:
c = np.zeros((m, n), dtype=np.float64, order='C')
if m == 0 or n == 0 or k == 0:
return c
c, tc = _reorder_fortran(c)
c = blas.dgemm(alpha=alpha,
a=b,
b=a,
c=c,
beta=beta,
trans_a=not tb,
trans_b=not ta)
c, tc = _reorder_c(c)
return c
def __init__(self, uhf, **kwargs):
super().__init__(uhf, **kwargs)
self.options = _set_options(self.options, **kwargs)
if mpi.mpi is None:
log.warn(
'No MPI4Py installation detected, OptUAGF2 will therefore run in serial.'
)
self.setup()
def zgemm(a, b, c=None, alpha=1.0 + 0.0j, beta=0.0 + 0.0j):
''' Performs zgemm in Fortran memory alignment without copying.
Input matrix should be contiguous (either F or C).
Parameters
----------
a : array
input matrix a
b : array
input matrix b
c : array, optional
output matrix c, if None then it is allocated inside the
function, default None
alpha : complex, optional
scalar factor for matrix a
beta : float, optional
scalar factor for matrix c
Returns
-------
c : ndarray
output matrix
'''
if (not _is_contiguous(a)) or (not _is_contiguous(b)):
log.warn('ZGEMM called on non-contiguous data')
m, k = a.shape
n = b.shape[1]
assert k == b.shape[0]
a, ta = _reorder_fortran(a)
b, tb = _reorder_fortran(b)
if c is None:
c = np.zeros((m, n), dtype=np.complex128, order='C')
if m == 0 or n == 0 or k == 0:
return c
c, tc = _reorder_fortran(c)
c = blas.zgemm(alpha=alpha,
a=b,
b=a,
c=c,
beta=beta,
trans_a=not tb,
trans_b=not ta)
c, tc = _reorder_c(c)
return c
def run(self, **kwargs):
if self.disable_omp:
with lib.with_omp_threads(1):
self._run(**kwargs)
else:
self._run(**kwargs)
if not self._pyscf.converged:
if mpi.rank:
log.warn('%s did not converge.' % self.__class__.__name__)
return self
def cholesky_qr(a):
''' Performs the Cholesky QR decomposition. This can be unstable.
'''
try:
x = np.dot(a.T, a)
r = np.linalg.cholesky(x).T
q = np.dot(a, np.linalg.inv(r))
except np.linalg.LinAlgError: # pragma: no cover
if mpi.rank:
log.warn('Matrix not positive definite in Cholesky step of '
'util.linalg.qholesky_qr - falling back to numpy.')
return np.linalg.qr(a)
return q, r
def davidson(se, h_phys, chempot=0.0, nroots=1, which='SM', tol=1e-14, maxiter=None, ntrial=None):
''' Diagonalise the auxiliary + physical space via a
Davidson iterative eigensolver.
'''
_check_dims(se, h_phys)
ndim = se.nphys + se.naux
if maxiter is None: maxiter = 10 * ndim
if ntrial is None: ntrial = min(ndim, max(2*nroots+1, 20))
abs_op = np.absolute if which in ['SM', 'LM'] else lambda x: x
order = 1 if which in ['SM', 'SA'] else -1
matvec = lambda x : se.dot(h_phys, np.asarray(x))
diag = np.concatenate([np.diag(h_phys), se.e])
guess = [np.zeros((ndim)) for n in range(nroots)]
mask = np.argsort(abs_op(diag))[::order]
for i in range(nroots):
guess[i][mask[i]] = 1
def pick(w, v, nroots, callback):
mask = np.argsort(abs_op(w))
mask = mask[::order]
w = w[mask]
v = v[:,mask]
return w, v, 0
conv, w, v = util.davidson(matvec, guess, diag, tol=tol, nroots=nroots,
max_space=ntrial, max_cycle=maxiter, pick=pick)
if not np.all(conv):
log.warn('Davidson solver did not converge.')
return w, v
def test_warn(self):
with self.assertWarns(Warning):
log.warn('This is a test')
