def collapse(self):
'''Explicit matrix representation of the operator'''
from xii.linalg.matrix_utils import diagonal_matrix, row_matrix
from xii import ii_convert
D = row_matrix(self.dual_basis)
B = row_matrix(self.primal_basis)
I = diagonal_matrix(B.size(1), 1.)
return ii_convert(I - B.T * D)
def is_linear(L, u):
'''Is L linear in u'''
# Compute the deriative
dL = ii_convert(ii_assemble(ii_derivative(L, u)))
if dL == 0:
info('dL/du is zero')
return None
# Random guy
w = Function(u.function_space()).vector()
w.set_local(np.random.rand(w.local_size()))
# Where we evaluate the direction
dw = PETScVector(as_backend_type(dL).mat().getVecLeft())
dL.mult(w, dw)
# Now L(u) + dw
Lu_dw = ii_assemble(L) + dw
# And if the thing is linear then L(u+dw) should be the same
u.vector().axpy(1, w)
Lu_dw0 = ii_assemble(L)
return (Lu_dw - Lu_dw0).norm('linf'), Lu_dw0.norm(
'linf') #, dw.norm('linf')
def collapse(self):
'''Explicit matrix representation of the operator'''
from xii import ii_convert
return ii_convert((self.P.M) * self.P.collapse())
ub, vb = TrialFunction(Vb), TestFunction(Vb)
b = [[0, 0], [0, 0]]
b[0][0] = inner(grad(u), grad(v)) * dx + inner(u, v) * dx
b[0][1] = inner(grad(ub), grad(v)) * dx + inner(ub, v) * dx
b[1][0] = inner(grad(u), grad(vb)) * dx + inner(u, vb) * dx
b[1][1] = inner(grad(ub), grad(vb)) * dx + inner(ub, vb) * dx
BB = ii_assemble(b)
x = Function(V).vector()
x.set_local(np.random.rand(x.local_size()))
y = Function(Vb).vector()
y.set_local(np.random.rand(y.local_size()))
bb = block_vec([x, y])
z_block = BB * bb
# Make into a monolithic matrix
BB_m = ii_convert(BB)
R = ReductionOperator([2], W=[V, Vb])
z = (R.T) * BB_m * (R * bb)
print(z - z_block).norm()
y = BB_m * (R * bb)
print np.linalg.norm(
np.hstack([bi.get_local() for bi in z_block]) - y.get_local())
def analyze_cond(problem, precond, ncases, alpha, get_cond, logfile):
'''Study of module over ncases for fixed alpha'''
# Annotate columns
columns = ['ndofs', 'h', 'cond', 'flag']
header = ' '.join(columns)
# Stuff for command line printing as we go, eigenvalua bounds, cond number, nzeros
formats = ['%d'] + ['%.2E'] + ['\033[1;37;34m%g(%g)\033[0m']
msg = ' '.join(formats)
mms_data = problem.setup_mms(alpha)
case0 = 0
with open(logfile, 'a') as stream:
# Run in context manager to keep the data
history = []
for n in [4*2**i for i in range(case0, case0+ncases)]:
# Setting up the problem means obtaining a block_mat, block_vec
# and a list space or matrix, vector and function space
try:
AA, bb, W = problem.setup_problem(n, mms_data, alpha)
Z = []
except ValueError:
AA, bb, W, Z = problem.setup_problem(n, mms_data, alpha)
# Let's get the preconditioner as block operator
try:
BB = precond(W, mms_data, alpha)
except TypeError:
try:
BB = precond(W, mms_data, alpha, AA)
except TypeError:
BB = precond(W, mms_data, alpha, AA, Z)
# Need a monolithic matrix
A = ii_convert(AA)
# spectrum expects matrices
B = ii_convert(BB) # ii_convert is identity for mat
# Again monolithic kernel
Z = map(ii_convert, Z) if Z else Z
# For dirichlet cases we might not have the list space
h = W[0].mesh().hmin() if isinstance(W, list) else W.mesh().hmin()
ndofs = A.size(0)
cond, flag = get_cond(A, B, Z)
# No convergence?
if cond is None: break
# Write
# Save current
row = (ndofs, h, cond)
stream.write('%d %g %.16f %s\n' % (ndofs, h, cond, flag))
# Put the entire history because of eps monitor
history.append(row)
print '='*79
print RED % str(alpha)
print GREEN % header
for i, row in enumerate(history):
if len(history) > 1:
increment = history[i][-1] - history[i-1][-1]
else:
increment = 0
print msg % (row + (increment, ))
print '='*79
# These should be linear
L = inner(Tu, q) * dxGamma
assert is_linear(L, Function(V)) == None
test(*is_linear(L, u))
# ---------------------------------------------------------------
L = inner(Tu + Tu, q) * dxGamma
test(*is_linear(L, u))
# ---------------------------------------------------------------
# # Some simple nonlinearity where I can check things
L = inner(Tu**2, q) * dxGamma
dL0 = inner(2 * Tu * Trace(TrialFunction(V), bmesh), q) * dxGamma
A0 = ii_convert(ii_assemble(dL0)).array()
dL = ii_derivative(L, u)
A = ii_convert(ii_assemble(dL)).array()
test(np.linalg.norm(A - A0, np.inf), np.linalg.norm(A0, np.inf))
# ---------------------------------------------------------------
L = inner(2 * Tu**3 - sin(Tu), q) * dxGamma
dL0 = inner(
(6 * Tu**2 - cos(Tu)) * Trace(TrialFunction(V), bmesh), q) * dxGamma
A0 = ii_convert(ii_assemble(dL0)).array()
dL = ii_derivative(L, u)
A = ii_convert(ii_assemble(dL)).array()
def __init__(self, mesh, orientation):
if orientation is None:
# We assume convex domain and take center as ...
orientation = mesh.coordinates().mean(axis=0)
if isinstance(orientation, df.Mesh):
# We set surface normal as outer with respect to orientation mesh
assert orientation.id() in mesh.parent_entity_map
n = df.FacetNormal(orientation)
hA = df.FacetArea(orientation)
# Project normal, we have a function on mesh
DLT = df.VectorFunctionSpace(orientation,
'Discontinuous Lagrange Trace', 0)
n_ = df.Function(DLT)
df.assemble((1 / hA) * df.inner(n, df.TestFunction(DLT)) * df.ds,
tensor=n_.vector())
# Now we get it to manifold
dx_ = df.Measure('dx', domain=mesh)
V = df.VectorFunctionSpace(mesh, 'DG', 0)
df.Function.__init__(self, V)
hK = df.CellVolume(mesh)
n_vec = xii.ii_convert(
xii.ii_assemble(
(1 / hK) *
df.inner(xii.Trace(n_, mesh), df.TestFunction(V)) * dx_))
self.vector()[:] = n_vec
return None
# Manifold assumption
assert 1 <= mesh.topology().dim() < mesh.geometry().dim()
gdim = mesh.geometry().dim()
# Orientation from inside point
if isinstance(orientation, (list, np.ndarray, tuple)):
assert len(orientation) == gdim
kwargs = {'x0%d' % i: val for i, val in enumerate(orientation)}
orientation = ['x[%d] - x0%d' % (i, i) for i in range(gdim)]
orientation = df.Expression(orientation, degree=1, **kwargs)
assert orientation.ufl_shape == (gdim, )
V = df.VectorFunctionSpace(mesh, 'DG', 0, gdim)
df.Function.__init__(self, V)
n_values = self.vector().get_local()
X = mesh.coordinates()
dd = X[np.argmin(X[:, 0])] - X[np.argmax(X[:, 0])]
values = []
R = np.array([[0, -1], [1, 0]])
for cell in df.cells(mesh):
n = cell.cell_normal().array()[:gdim]
x = cell.midpoint().array()[:gdim]
# Disagree?
if np.inner(orientation(x), n) < 0:
n *= -1.
values.append(n / np.linalg.norm(n))
values = np.array(values)
for sub in range(gdim):
dofs = V.sub(sub).dofmap().dofs()
n_values[dofs] = values[:, sub]
self.vector().set_local(n_values)
self.vector().apply('insert')
