def test(model, spaceName, dimD, dimR, storage):
# print("########################################")
# print("#### ",spaceName,storage,dimD,dimR,flush=True)
spaceD = create.space(spaceName,
grid,
dimRange=dimD,
order=1,
storage=storage)
spaceR = create.space(spaceName,
grid,
dimRange=dimR,
order=1,
storage=storage)
scheme = create.operator("galerkin", model, spaceD, spaceR)
uD = create.function("discrete", spaceD, name=storage)
uD.clear()
start = time.time()
for i in range(testLoop):
A = linearOperator(scheme) # , parameters={"petsc.blockedmode":False})
end = time.time()
# print( "setup+assembly:",(end-start)/testLoop, flush=True )
start = time.time()
for i in range(testLoop):
jacobian(scheme, uD, A)
end = time.time()
# print( "assembly only: ",(end-start)/testLoop, flush=True )
# sys.stdout.flush()
try:
import petsc4py
from petsc4py import PETSc
mat = A.as_petsc
# print(mat.getInfo(), flush=True)
except:
pass
def __init__(self, scheme):
self.model = scheme.model
self.res = scheme.space.interpolate([0],name="residual")
self.scheme = scheme
self.jacobian = linearOperator(self.scheme)
self.snes = PETSc.SNES().create()
self.snes.setFunction(self.f, self.res.as_petsc.duplicate())
self.snes.setUseMF(False)
self.snes.setJacobian(self.Df, self.jacobian.as_petsc, self.jacobian.as_petsc)
self.snes.getKSP().setType("cg")
self.snes.setFromOptions()
def __init__(self, scheme):
self.model = scheme.model
self.jacobian = linearOperator(scheme)
self.ksp = PETSc.KSP()
self.ksp.create(PETSc.COMM_WORLD)
# use conjugate gradients method
self.ksp.setType("cg")
# and incomplete Cholesky
self.ksp.getPC().setType("icc")
self.ksp.setOperators(self.jacobian.as_petsc)
self.ksp.setFromOptions()
def test(space):
if test_numpy:
numpySpace = create.space(space,
grid,
dimRange=1,
order=1,
storage='numpy')
numpyScheme = create.scheme("galerkin", model, numpySpace)
numpy_h = create.function("discrete", numpySpace, name="numpy")
numpy_dofs = numpy_h.as_numpy
# numpyScheme.solve(target = numpy_h)
start = time.time()
for i in range(testLoop):
linOp = linearOperator(numpyScheme)
numpyScheme.jacobian(numpy_h, linOp)
numpy_mat = linOp.as_numpy
end = time.time()
# print( "numpy:", (end-start)/testLoop, flush=True )
# sys.stdout.flush()
numpy_coo = numpy_mat.tocoo()
# for i,j,v in zip(numpy_coo.row,numpy_coo.col,numpy_coo.data):
# print(i,j,v)
# print("****************************",flush=True)
if test_istl:
istlSpace = create.space(space,
grid,
dimRange=1,
order=1,
storage='istl')
istlScheme = create.scheme("galerkin", model, istlSpace)
istl_h = create.function("discrete", istlSpace, name="istl")
istl_dofs = istl_h.as_istl
# istlScheme.solve(target = istl_h)
start = time.time()
for i in range(testLoop):
linOp = linearOperator(istlScheme)
istlScheme.jacobian(istl_h, linOp)
istl_mat = linOp.as_istl
end = time.time()
# print( "istl:", (end-start)/testLoop, flush=True )
# sys.stdout.flush()
# there is no way yet to go from istl to scipy - would be nice to have
# istl_coo = istl_mat.tocoo()
# for i,j,v in zip(eigen_coo.row,eigen_coo.col,eigen_coo.data):
# print(i,j,v)
# print("****************************",flush=True)
if test_petsc:
import petsc4py
from petsc4py import PETSc
petsc4py.init(sys.argv)
petscSpace = create.space(space,
grid,
dimRange=1,
order=1,
storage='petsc')
petscScheme = create.scheme("galerkin", model, petscSpace)
petsc_h = create.function("discrete", petscSpace, name="petsc")
petsc_dofs = petsc_h.as_petsc
# petscScheme.solve(target = petsc_h)
linOp = linearOperator(petscScheme)
petscScheme.jacobian(petsc_h, linOp)
petsc_mat = linOp.as_petsc
rptr, cind, vals = petsc_mat.getValuesCSR()
petsc_coo = scipy.sparse.csr_matrix((vals, cind, rptr)).tocoo()
start = time.time()
for i in range(testLoop):
linOp = linearOperator(petscScheme)
petscScheme.jacobian(petsc_h, linOp)
petsc_mat = linOp.as_petsc
end = time.time()
# print( "petsc:", (end-start)/testLoop, flush=True )
# sys.stdout.flush()
ksp = PETSc.KSP()
ksp.create(PETSc.COMM_WORLD)
ksp.setType("cg")
# ksp.getPC().setType("icc")
petsc_h.clear()
res = petsc_h.copy()
petscScheme(petsc_h, res)
petscScheme.jacobian(petsc_h, linOp)
petsc_mat = linOp.as_petsc
ksp.setOperators(petsc_mat, petsc_mat)
ksp.setFromOptions()
ksp.solve(res.as_petsc, petsc_h.as_petsc)
# print("****************************",flush=True)
# print(petsc_mat.size, petsc_mat.getSize(), petsc_mat.getSizes())
# print(petsc_mat.getType())
# print(type(petsc_mat))
# print(petscSpace.size)
# print(petsc_mat.assembled)
# rptr, cind, vals = petsc_mat.getValuesCSR()
# petsc_coo = scipy.sparse.csr_matrix((vals,cind,rptr),shape=(100,100)).tocoo()
# for i,j,v in zip(petsc_coo.row,petsc_coo.col,petsc_coo.data):
# print(i,j,v)
if test_istl:
try: # istl_coo does not exist
assert (istl_coo.row == numpy_coo.row).all()
assert (istl_coo.col == numpy_coo.col).all()
assert np.allclose(istl_coo.data, numpy_coo.data)
except:
# print("issue between istl and numpy matrices")
pass
if test_petsc:
try:
assert (petsc_coo.row == numpy_coo.row).all()
assert (petsc_coo.col == numpy_coo.col).all()
assert np.allclose(petsc_coo.data, numpy_coo.data)
except:
# print("issue between petsc and numpy matrices")
pass
velocity = spcU.interpolate([
0,
] * spcU.dimRange, name="velocity")
pressure = spcP.interpolate([0], name="pressure")
rhsVelo = velocity.copy()
rhsPress = pressure.copy()
sol_u = velocity.as_numpy
sol_p = pressure.as_numpy
rhs_u = rhsVelo.as_numpy
rhs_p = rhsPress.as_numpy
r = numpy.copy(rhs_p)
d = numpy.copy(rhs_p)
precon = numpy.copy(rhs_p)
xi = numpy.copy(rhs_u)
A = linearOperator(mainOp).as_numpy
G = linearOperator(gradOp).as_numpy
D = linearOperator(divOp).as_numpy
M = linearOperator(massOp).as_numpy
P = linearOperator(preconOp).as_numpy
def Ainv(rhs, target):
target[:] = linalg.spsolve(A, rhs)
def Minv(rhs, target):
target[:] = linalg.spsolve(M, rhs)
def Pinv(rhs, target):
def __init__(self, x_coeff):
self.shape = (x_coeff.shape[0], x_coeff.shape[0])
self.dtype = x_coeff.dtype
x = space.function("tmp", dofVector=x_coeff)
self.jacobian = linearOperator(scheme, ubar=x)
def __init__(self, scheme):
self.model = scheme.model
self.jacobian = linearOperator(scheme)
def monolithicSolve(schemes,
targets,
callback=None,
iter=10,
f_tol=1e-8,
eps=1e-6,
verbose=0):
"""Helper function to solve bulk and interface scheme coupled monolithically.
A newton method based on scipy.
The coupling jacobian blocks are evalutaed by finite difference on demand.
The Schur complement is used for the inversion of the block-wise jacobian.
Args:
schemes: pair of schemes
targets: pair of discrete functions that should be solved for AND that are used in the coupling forms
callback: update function that is called every time before solving a scheme
iter: maximum number of iterations
f_tol: objective residual two norm
eps: step size for finite difference
verbose: 1: print residuum for each iteration, 2: and for each Schur iteration
Returns:
if converged
"""
assert len(schemes) == 2
assert len(targets) == 2
(scheme, ischeme) = schemes
(uh, th) = targets
n = len(uh.as_numpy)
m = len(th.as_numpy)
if m == 0:
if verbose:
print("second scheme is empty, forward to scheme.solve()",
flush=True)
scheme.solve(target=uh)
return
from dune.fem.operator import linear as linearOperator
A = linearOperator(scheme)
D = linearOperator(ischeme)
def updateJacobians():
scheme.jacobian(uh, A)
ischeme.jacobian(th, D)
def call():
if callback is not None:
callback()
# Evaluate the coupling blocks by finite difference on-demand
Bz = uh.copy()
def B(z):
norm = np.sqrt(z.dot(z))
if norm == 0:
return np.zeros(n)
th.as_numpy[:] += z * eps / norm
call()
scheme(uh, Bz)
th.as_numpy[:] -= z * eps / norm
Bz.as_numpy[:] -= f.as_numpy
Bz.as_numpy[:] /= eps
return Bz.as_numpy * norm
Cz = th.copy()
def C(z):
norm = np.sqrt(z.dot(z))
if norm == 0:
return np.zeros(m)
uh.as_numpy[:] += z * eps / norm
call()
ischeme(th, Cz)
uh.as_numpy[:] -= z * eps / norm
Cz.as_numpy[:] -= g.as_numpy
Cz.as_numpy[:] /= eps
return Cz.as_numpy * norm
# Evaluate
f = uh.copy()
g = th.copy()
call()
scheme(uh, f)
ischeme(th, g)
i = 0
def checkResiduum():
a = f.as_numpy
b = g.as_numpy
res = np.sqrt(np.dot(a, a) + np.dot(b, b))
if verbose > 0:
print(" i:", i, " | f | =", res)
if res < f_tol:
return True
if checkResiduum():
return True
from scipy.sparse.linalg import LinearOperator, spsolve, lgmres
for i in range(1, iter):
updateJacobians()
def eval(x):
return
S = LinearOperator((n + m, n + m), lambda x: np.concatenate(
(A.as_numpy.dot(x[:n]) + B(x[n:]), C(x[:n]) + D.as_numpy.dot(x[n:])
)))
M = LinearOperator((n + m, n + m), lambda x: np.concatenate(
(spsolve(A.as_numpy, x[:n]), spsolve(D.as_numpy, x[n:]))))
r = np.concatenate((f.as_numpy, g.as_numpy))
x0 = np.concatenate((uh.as_numpy, th.as_numpy))
x, _ = lgmres(S, r, x0=x0, M=M, tol=f_tol)
uh.as_numpy[:] -= x[:n]
th.as_numpy[:] -= x[n:]
call()
scheme(uh, f)
ischeme(th, g)
if checkResiduum():
return True
return False
gradOp = galerkinOperator(gradModel)
divOp = galerkinOperator(divModel)
massOp = galerkinScheme(massModel==0)
preconOp = galerkinScheme(preconModel==0)
velocity = spcU.interpolate([0,]*spcU.dimRange, name="velocity")
pressure = spcP.interpolate([0], name="pressure")
rhsVelo = velocity.copy()
rhsPress = pressure.copy()
r = rhsPress.copy()
d = rhsPress.copy()
precon = rhsPress.copy()
xi = rhsVelo.copy()
A = linearOperator(mainOp)
G = linearOperator(gradOp)
D = linearOperator(divOp)
M = linearOperator(massOp)
P = linearOperator(preconOp)
# Note issue with using 'cg' due to
# (a) missing mainOp.setConstraints(velocity)
# (b) no bc for pressure preconder....
solver = {"method":"cg","tolerance":1e-10, "verbose":False}
Ainv = mainOp.inverseLinearOperator(A,parameters=solver)
Minv = massOp.inverseLinearOperator(M,parameters=solver)
Pinv = preconOp.inverseLinearOperator(P,solver)
mainOp(velocity,rhsVelo)
rhsVelo *= -1
a = (inner((u) / dt, v) + inner(u, v)) * dx
exact = as_vector([
exp(-2 * t) * (initial - 1) + 1,
] * dimR)
b = replace(a, {u: exact})
solverParam = {"newton.verbose": 0, "newton.linear.verbose": 0}
scheme = create.scheme("galerkin",
a == b,
space,
solver='cg',
parameters=solverParam)
scheme.setQuadratureOrders(quadOrder, quadOrder)
scheme.model.dt = 0.05
grid.writeVTK('initial', pointdata={'initial': initial})
start = time.time()
t = 0
A = linearOperator(scheme)
while t < 1.0:
scheme.model.t = t
u_h_n.assign(u_h)
scheme.solve(target=u_h)
scheme.jacobian(u_h_n, A)
t += scheme.model.dt
print("time loop:", time.time() - start)
grid.writeVTK('forchheimer', pointdata=[u_h])
0,
] * dimR), name='destD')
destE = space.interpolate(as_vector([
0,
] * dimR), name='destE')
u = TrialFunction(space)
v = TestFunction(space)
x = SpatialCoordinate(space.cell())
ubar = space.interpolate(as_vector([
dot(x, x),
] * dimR), name="ubar")
a = (inner(0.5 * dot(u, u), v[0]) + inner(u[0] * grad(u), grad(v))) * dx
op = create.operator("galerkin", a, space)
A = linearOperator(op)
op.jacobian(ubar, A)
A(arg, destA)
da = apply_derivatives(derivative(action(a, ubar), ubar, u))
dop = create.operator("galerkin", da, space)
dop(arg, destB)
err = integrate(grid, (destA - destB)**2, 5)
# print("error=",err)
assert (err < 1e-15)
A = linearOperator(dop)
dop.jacobian(arg, A)
A(arg, destC)
err = integrate(grid, (destA - destC)**2, 5)
# print("error=",err)
