def matvec(self, b):
from PyTrilinos import AztecOO
from dolfin import GenericVector
if not isinstance(b, GenericVector):
return NotImplemented
x = self.A.create_vec(dim=1)
if len(x) != len(b):
raise RuntimeError(
'incompatible dimensions for AztecOO matvec, %d != %d'%(len(x),len(b)))
solver = AztecOO.AztecOO(self.A.down_cast().mat(), x.down_cast().vec(), b.down_cast().vec())
#solver.SetAztecDefaults()
solver.SetAztecOption(AztecOO.AZ_solver, self.solver)
if self.precond:
if hasattr(self.precond, 'down_cast'):
solver.SetPrecOperator(self.precond.down_cast())
else:
# doesn't seem to work very well
solver.SetAztecOption(AztecOO.AZ_precond, self.precond)
# the following are from the example with precond='dom_decomp'
solver.SetAztecOption(AztecOO.AZ_subdomain_solve, AztecOO.AZ_ilu)
solver.SetAztecOption(AztecOO.AZ_overlap, 1)
solver.SetAztecOption(AztecOO.AZ_graph_fill, 1)
solver.SetAztecOption(AztecOO.AZ_output, 0)
solver.Iterate(self.maxiter, self.tolerance)
return x
def iterative_solver(List, Matrix, InputLHS, RHS, Prec):
LHS = Epetra.MultiVector(InputLHS)
Time = Epetra.Time(Matrix.Comm())
hasConditionNumber = False;
Solver = AztecOO.AztecOO(Matrix, LHS, RHS)
Solver.SetPrecOperator(Prec)
if (List['az_solver'] == "AZ_gmres"):
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres);
elif List['az_solver'] == "AZ_cg":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cg);
elif List['az_solver'] == "AZ_cg_condnum":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cg_condnum);
hasConditionNumber = True
elif List['az_solver'] == "AZ_gmres_condnum":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres_condnum);
hasConditionNumber = True
elif List['az_solver'] == "AZ_bicgstab":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_bicgstab);
elif List['az_solver'] == "AZ_tfqmr":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_tfqmr);
elif List['az_solver'] == "AZ_cgs":
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cgs);
else:
print "Solver type not correct, ", List['az_solver']
Solver.SetAztecOption(AztecOO.AZ_output, 16);
if List['iters'] < 0 | List['iters'] > MAX_ITERATIONS:
print "Maximum number of iterations either negative of > ", MAX_ITERATIONS;
raise("PARAMETER_ERROR");
if List['tol'] < 1e-12:
print "Tolerance is too small"
raise("PARAMETER_ERROR");
if List['az_kspace'] > 0 & List['az_kspace'] <= MAX_KSPACE:
Solver.SetAztecOption(AztecOO.AZ_kspace, List['az_kspace']);
else:
print "Krylov space dimension either negative of >", MAX_KSPACE
print "You have", List['az_kspace']
raise("PARAMETER_ERROR");
if List['az_output'] == "16":
Solver.SetAztecOption(AztecOO.AZ_output, 16)
elif List['az_output'] == "32":
Solver.SetAztecOption(AztecOO.AZ_output, 32)
elif List['az_output'] == "AZ_last":
Solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_last)
elif List['az_output'] == "AZ_none":
Solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_none)
err = Solver.Iterate(List['iters'], List['tol'])
if hasConditionNumber:
ConditionNumber = Solver.GetStatus(AztecOO.AZ_condnum)
else:
ConditionNumber = 0.0;
return (err, Solver.NumIters(), Time.ElapsedTime(), ConditionNumber)
def solve_linear_problem(comm, A, b):
"""
retorna a solucao do sistema linear Ax = b
input:
A: matriz do sistema
b: termo fonte
output:
x:solucao
"""
n = len(b)
assert A.NumMyCols() == A.NumMyRows()
assert A.NumMyCols() == n
if A.Filled():
pass
else:
A.FillComplete()
std_map = Epetra.Map(n, 0, comm)
x = Epetra.Vector(std_map)
linearProblem = Epetra.LinearProblem(A, x, b)
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_warnings)
solver.Iterate(10000, 1e-14)
return x
def solve_linear_problem(self, A, b, x=None, its=1000, tolerance=1e-10):
'''
resolve o problema Ax = b
input:
A: matriz quadrada esparsa do scipy
b = termo fonte
x: chute inicial
its: numero maximo de iteracoes
tolerance: tolerancia para o residuo
output:
res: informa se o residuo foi menor que a tolerancia
x2: vetor resposta
'''
comm = self._comm
n = len(b)
std_map = Epetra.Map(n, 0, comm)
x2 = Epetra.Vector(std_map)
if x:
x2[:] = x[:]
b2 = Epetra.Vector(std_map)
b2[:] = b[:]
A2 = Epetra.CrsMatrix(Epetra.Copy, std_map, 7)
indices = sp.find(A)
A2.InsertGlobalValues(indices[0], indices[1], indices[2])
irr = A2.FillComplete()
linearProblem = Epetra.LinearProblem(A2, x2, b2)
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_warnings)
# solver.SetParameters(self._params)
solver.Iterate(its, tolerance)
x2 = np.array(x2)
res = solver.ScaledResidual() < tolerance
return x2
def solve(self):
# parallel linear solver
# final project has 2D differential equation
linearProblem = Epetra.LinearProblem(self.A, self.x, self.b)
solver = AztecOO.AztecOO(linearProblem)
solver.Iterate(10000, 1.e-5)
return
def trilinos_solve(A, rhs, prec, x=None, tol=1e-5):
solver = AztecOO.AztecOO(A, x, rhs)
solver.SetPrecOperator(prec)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_conv, AztecOO.AZ_rhs)
solver.SetAztecOption(AztecOO.AZ_output, 0)
solver.Iterate(3000, tol)
return x
def solve_system(operator, rhs, process_maps, tol=1E-5, prec=None):
"""Solve the global system with GMRES"""
from PyTrilinos import AztecOO
x = Epetra.MultiVector(process_maps.single_proc_map, 1)
solver = AztecOO.AztecOO(operator, x, rhs)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
if prec is None:
solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_none)
else:
solver.SetPrecOperator(P)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_last)
solver.Iterate(500, tol)
return x
def setup_linear_problem(self, A_name, b_name):
"""Sets a linear problem. The `A_name` and `b_name` parameters are
related to the matrix and vectors names to be used in a linear problem
Ax = b.
"""
self.A = self.mesh.matrix_manager.get_matrix(A_name)
self.b = self.mesh.matrix_manager.get_vector(b_name)
self.x = Epetra.Vector(
self.mesh.matrix_manager.std_map[self.solution_dim])
self.linearProblem = Epetra.LinearProblem(self.A, self.x, self.b)
self.solver = AztecOO.AztecOO(self.linearProblem)
self.solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_last)
def solve_linear_problem(self, A, b, n):
if A.Filled():
pass
else:
A.FillComplete()
std_map = Epetra.Map(n, 0, self.comm)
x = Epetra.Vector(std_map)
linearProblem = Epetra.LinearProblem(A, x, b)
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_warnings)
solver.Iterate(10000, 1e-14)
return x
def _solve_(self, L, x, b):
Solver = AztecOO.AztecOO(L, x, b)
Solver.SetAztecOption(AztecOO.AZ_solver, self.solver)
## Solver.SetAztecOption(AztecOO.AZ_kspace, 30)
Solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_none)
if self.preconditioner is not None:
self.preconditioner._applyToSolver(solver=Solver, matrix=L)
else:
Solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_none)
output = Solver.Iterate(self.iterations, self.tolerance)
if self.preconditioner is not None:
if hasattr(self.preconditioner, 'Prec'):
del self.preconditioner.Prec
if 'FIPY_VERBOSE_SOLVER' in os.environ:
status = Solver.GetAztecStatus()
from fipy.tools.debug import PRINT
PRINT('iterations: %d / %d' %
(status[AztecOO.AZ_its], self.iterations))
failure = {
AztecOO.AZ_normal: 'AztecOO.AZ_normal',
AztecOO.AZ_param: 'AztecOO.AZ_param',
AztecOO.AZ_breakdown: 'AztecOO.AZ_breakdown',
AztecOO.AZ_loss: 'AztecOO.AZ_loss',
AztecOO.AZ_ill_cond: 'AztecOO.AZ_ill_cond',
AztecOO.AZ_maxits: 'AztecOO.AZ_maxits'
}
PRINT('failure', failure[status[AztecOO.AZ_why]])
PRINT('AztecOO.AZ_r:', status[AztecOO.AZ_r])
PRINT('AztecOO.AZ_scaled_r:', status[AztecOO.AZ_scaled_r])
PRINT('AztecOO.AZ_rec_r:', status[AztecOO.AZ_rec_r])
PRINT('AztecOO.AZ_solve_time:', status[AztecOO.AZ_solve_time])
PRINT('AztecOO.AZ_Aztec_version:',
status[AztecOO.AZ_Aztec_version])
return output
def solve_aztec(A, rhs, x=None, tol=1e-5, plist=None):
if x is None:
x = Epetra.Vector(A.RangeMap())
if plist is None:
plist = {"Solver": "bicgstab",
"Precond": "Dom_Decomp",
"subdomain_solve": "ilu",
"Output": 0
}
solver = AztecOO.AztecOO(A, x, rhs)
solver.SetParameters(plist, True)
solver.SetAztecOption(AztecOO.AZ_output, 0)
solver.SetAztecOption(AztecOO.AZ_conv, AztecOO.AZ_rhs)
ierr = solver.Iterate(300000, tol)
assert (ierr == 0) or (ierr == -3), ierr
return x
def solve(self, op, rhs, tol, maxit):
if op.dtype.char != op.dtype.char.upper():
# Real case
if abs(op.alpha.real) < abs(op.alpha.imag):
op.alpha = op.alpha.imag
else:
op.alpha = op.alpha.real
op.beta = op.beta.real
epetra_op = Operator(op)
x = Vector(rhs)
solver = AztecOO.AztecOO(epetra_op, x, rhs)
solver.SetParameters({
"Solver": "GMRES",
"Precond": "None",
"Output": -3
}) # Warnings. See az_aztec_defs.h
info = solver.Iterate(maxit, tol)
if info < 0:
raise Exception('AztecOO returned ' + str(info))
elif info > 0 and maxit > 1:
warnings.warn('GMRES did not converge in ' +
str(solver.NumIters()) + ' iterations')
return x
def run_solver(self, interpolation_method):
"""Run solver."""
self.interpolation_method = interpolation_method
t0 = time.time()
n_vertex = len(set(self.mesh_data.all_nodes) - self.dirichlet_nodes)
print("interpolation runing...")
self.get_nodes_weights(interpolation_method)
print(
"done interpolation...",
"took {0} seconds to interpolate over {1} verts".format(
time.time() - t0, n_vertex
),
)
print("filling the transmissibility matrix...")
begin = time.time()
try:
for volume in self.volumes:
volume_id = self.mb.tag_get_data(self.global_id_tag, volume)[
0
][0]
RHS = self.mb.tag_get_data(self.source_tag, volume)[0][0]
self.Q[volume_id] += RHS
# self.Q[volume_id, 0] += RHS
except Exception:
pass
for face in self.neumann_faces:
face_flow = self.mb.tag_get_data(self.neumann_tag, face)[0][0]
volume = self.mtu.get_bridge_adjacencies(face, 2, 3)
volume = np.asarray(volume, dtype="uint64")
id_volume = self.mb.tag_get_data(self.global_id_tag, volume)[0][0]
face_nodes = self.mtu.get_bridge_adjacencies(face, 0, 0)
node_crds = self.mb.get_coords(face_nodes).reshape([3, 3])
face_area = geo._area_vector(node_crds, norma=True)
RHS = face_flow * face_area
self.Q[id_volume] += -RHS
# self.Q[id_volume, 0] += - RHS
id_volumes = []
all_LHS = []
for face in self.dirichlet_faces:
# '2' argument was initially '0' but it's incorrect
I, J, K = self.mtu.get_bridge_adjacencies(face, 2, 0)
left_volume = np.asarray(
self.mtu.get_bridge_adjacencies(face, 2, 3), dtype="uint64"
)
id_volume = self.mb.tag_get_data(self.global_id_tag, left_volume)[
0
][0]
id_volumes.append(id_volume)
JI = self.mb.get_coords([I]) - self.mb.get_coords([J])
JK = self.mb.get_coords([K]) - self.mb.get_coords([J])
LJ = (
self.mb.get_coords([J])
- self.mesh_data.mb.tag_get_data(
self.volume_centre_tag, left_volume
)[0]
)
N_IJK = np.cross(JI, JK) / 2.0
_test = np.dot(LJ, N_IJK)
if _test < 0.0:
I, K = K, I
JI = self.mb.get_coords([I]) - self.mb.get_coords([J])
JK = self.mb.get_coords([K]) - self.mb.get_coords([J])
N_IJK = np.cross(JI, JK) / 2.0
tan_JI = np.cross(N_IJK, JI)
tan_JK = np.cross(N_IJK, JK)
self.mb.tag_set_data(self.normal_tag, face, N_IJK)
face_area = np.sqrt(np.dot(N_IJK, N_IJK))
h_L = geo.get_height(N_IJK, LJ)
g_I = self.get_boundary_node_pressure(I)
g_J = self.get_boundary_node_pressure(J)
g_K = self.get_boundary_node_pressure(K)
K_L = self.mb.tag_get_data(self.perm_tag, left_volume).reshape(
[3, 3]
)
K_n_L = self.vmv_multiply(N_IJK, K_L, N_IJK)
K_L_JI = self.vmv_multiply(N_IJK, K_L, tan_JI)
K_L_JK = self.vmv_multiply(N_IJK, K_L, tan_JK)
D_JK = self.get_cross_diffusion_term(
tan_JK, LJ, face_area, h_L, K_n_L, K_L_JK, boundary=True
)
D_JI = self.get_cross_diffusion_term(
tan_JI, LJ, face_area, h_L, K_n_L, K_L_JI, boundary=True
)
K_eq = (1 / h_L) * (face_area * K_n_L)
RHS = D_JK * (g_I - g_J) - K_eq * g_J + D_JI * (g_J - g_K)
LHS = K_eq
all_LHS.append(LHS)
self.Q[id_volume] += -RHS
# self.Q[id_volume, 0] += - RHS
# self.mb.tag_set_data(self.flux_info_tag, face,
# [D_JK, D_JI, K_eq, I, J, K, face_area])
all_cols = []
all_rows = []
all_values = []
self.ids = []
self.v_ids = []
self.ivalues = []
for face in self.intern_faces:
left_volume, right_volume = self.mtu.get_bridge_adjacencies(
face, 2, 3
)
L = self.mesh_data.mb.tag_get_data(
self.volume_centre_tag, left_volume
)[0]
R = self.mesh_data.mb.tag_get_data(
self.volume_centre_tag, right_volume
)[0]
dist_LR = R - L
I, J, K = self.mtu.get_bridge_adjacencies(face, 0, 0)
JI = self.mb.get_coords([I]) - self.mb.get_coords([J])
JK = self.mb.get_coords([K]) - self.mb.get_coords([J])
N_IJK = np.cross(JI, JK) / 2.0
test = np.dot(N_IJK, dist_LR)
if test < 0:
left_volume, right_volume = right_volume, left_volume
L = self.mesh_data.mb.tag_get_data(
self.volume_centre_tag, left_volume
)[0]
R = self.mesh_data.mb.tag_get_data(
self.volume_centre_tag, right_volume
)[0]
dist_LR = R - L
face_area = np.sqrt(np.dot(N_IJK, N_IJK))
tan_JI = np.cross(N_IJK, JI)
tan_JK = np.cross(N_IJK, JK)
K_R = self.mb.tag_get_data(self.perm_tag, right_volume).reshape(
[3, 3]
)
RJ = R - self.mb.get_coords([J])
h_R = geo.get_height(N_IJK, RJ)
K_R_n = self.vmv_multiply(N_IJK, K_R, N_IJK)
K_R_JI = self.vmv_multiply(N_IJK, K_R, tan_JI)
K_R_JK = self.vmv_multiply(N_IJK, K_R, tan_JK)
K_L = self.mb.tag_get_data(self.perm_tag, left_volume).reshape(
[3, 3]
)
LJ = L - self.mb.get_coords([J])
h_L = geo.get_height(N_IJK, LJ)
K_L_n = self.vmv_multiply(N_IJK, K_L, N_IJK)
K_L_JI = self.vmv_multiply(N_IJK, K_L, tan_JI)
K_L_JK = self.vmv_multiply(N_IJK, K_L, tan_JK)
D_JI = self.get_cross_diffusion_term(
tan_JI,
dist_LR,
face_area,
h_L,
K_L_n,
K_L_JI,
h_R,
K_R_JI,
K_R_n,
)
D_JK = self.get_cross_diffusion_term(
tan_JK,
dist_LR,
face_area,
h_L,
K_L_n,
K_L_JK,
h_R,
K_R_JK,
K_R_n,
)
K_eq = (K_R_n * K_L_n) / (K_R_n * h_L + K_L_n * h_R) * face_area
id_right = self.mb.tag_get_data(self.global_id_tag, right_volume)
id_left = self.mb.tag_get_data(self.global_id_tag, left_volume)
col_ids = [id_right, id_right, id_left, id_left]
row_ids = [id_right, id_left, id_left, id_right]
values = [K_eq, -K_eq, K_eq, -K_eq]
all_cols.append(col_ids)
all_rows.append(row_ids)
all_values.append(values)
# wait for interpolation to be done
self._node_treatment(I, id_left, id_right, K_eq, D_JK=D_JK)
self._node_treatment(
J, id_left, id_right, K_eq, D_JI=D_JI, D_JK=-D_JK
)
self._node_treatment(K, id_left, id_right, K_eq, D_JI=-D_JI)
# self.mb.tag_set_data(self.flux_info_tag, face,
# [D_JK, D_JI, K_eq, I, J, K, face_area])
self.T.InsertGlobalValues(self.ids, self.v_ids, self.ivalues)
# self.T[
# np.asarray(self.ids)[:, :, 0, 0], np.asarray(self.v_ids)
# ] = np.asarray(self.ivalues)
self.T.InsertGlobalValues(id_volumes, id_volumes, all_LHS)
# self.T[
# np.asarray(id_volumes), np.asarray(id_volumes)
# ] = np.asarray(all_LHS)
self.T.InsertGlobalValues(all_cols, all_rows, all_values)
# self.T[
# np.asarray(all_cols)[:, 0, 0, 0],
# np.asarray(all_rows)[:, 0, 0, 0]
# ] = np.asarray(all_values)[:, 0]
self.T.FillComplete()
mat_fill_time = time.time() - begin
print("matrix fill took {0} seconds...".format(mat_fill_time))
mesh_size = len(self.volumes)
print("running solver...")
USE_DIRECT_SOLVER = False
linearProblem = Epetra.LinearProblem(self.T, self.x, self.Q)
if USE_DIRECT_SOLVER:
solver = Amesos.Lapack(linearProblem)
print("1) Performing symbolic factorizations...")
solver.SymbolicFactorization()
print("2) Performing numeric factorizations...")
solver.NumericFactorization()
print("3) Solving the linear system...")
solver.Solve()
t = time.time() - t0
print(
"Solver took {0} seconds to run over {1} volumes".format(
t, mesh_size
)
)
else:
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_none)
# solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_Jacobi)
# solver.SetAztecOption(AztecOO.AZ_kspace, 1251)
# solver.SetAztecOption(AztecOO.AZ_orthog, AztecOO.AZ_modified)
# solver.SetAztecOption(AztecOO.AZ_conv, AztecOO.AZ_Anorm)
solver.Iterate(8000, 1e-10)
t = time.time() - t0
its = solver.GetAztecStatus()[0]
solver_time = solver.GetAztecStatus()[6]
print(
"Solver took {0} seconds to run over {1} volumes".format(
t, mesh_size
)
)
print(
"Solver converged at %.dth iteration in %3f seconds."
% (int(its), solver_time)
)
# self.T = self.T.tocsc()
# self.Q = self.Q.tocsc()
# self.x = spsolve(self.T, self.Q)
# print(np.sum(self.T[50]), self.Q[50])
self.mb.tag_set_data(self.pressure_tag, self.volumes, self.x)
def runit(size, prec, diff):
# builds the linear system matrix and sets up starting solution and
# right-hand side
Comm = Epetra.PyComm()
print "SIZE = ", size
print "PREC = ", prec
print "DIFF = ", diff
GaleriList = {
"nx": size,
"ny": size,
"mx": 1,
"my": 1,
"conv": 1.0,
"diff": diff
}
Map = Galeri.CreateMap("Cartesian2D", Comm, GaleriList)
Matrix = Galeri.CreateCrsMatrix("BentPipe2D", Map, GaleriList)
Util = Epetra.Util()
Util.SetSeed(0)
LHS = Epetra.Vector(Map)
RHS = Epetra.Vector(Map)
n = LHS.MyLength()
for i in xrange(0, n):
LHS[i] = sin(i * 3.1415 / n) * sin(i * 3.1415 / n)
Matrix.Multiply(False, LHS, RHS)
LHS.PutScalar(0.0)
# sets up the parameters for ML using a python dictionary
if prec == "SA":
MLList = {
"max levels": 10,
"output": 10,
"increasing or decreasing": "increasing",
"aggregation: type": "Uncoupled-MIS",
"smoother: type (level 0)": "symmetric Gauss-Seidel",
"smoother: damping factor": 0.67,
"smoother: sweeps": 1,
"smoother: pre or post": "both",
"PDE equations": 1,
"aggregation: damping factor": 1.333,
"eigen-analysis: type": "power-method"
}
elif prec == "NSA":
MLList = {
"max levels": 10,
"output": 10,
"increasing or decreasing": "increasing",
"aggregation: type": "Uncoupled-MIS",
"smoother: type (level 0)": "symmetric Gauss-Seidel",
"smoother: damping factor": 0.67,
"smoother: sweeps": 1,
"smoother: pre or post": "both",
"PDE equations": 1,
"aggregation: damping factor": 0.0000,
"eigen-analysis: type": "power-method"
}
elif prec == "NSR":
MLList = {
"max levels": 10,
"output": 10,
"increasing or decreasing": "increasing",
"aggregation: type": "Uncoupled-MIS",
"smoother: type (level 0)": "symmetric Gauss-Seidel",
"smoother: damping factor": 0.67,
"smoother: sweeps": 1,
"smoother: pre or post": "both",
"PDE equations": 1,
"aggregation: use tentative restriction": True,
"aggregation: damping factor": 1.0000,
"eigen-analysis: type": "power-method"
}
elif prec == "EMIN":
MLList = {
"max levels": 10,
"output": 10,
"increasing or decreasing": "increasing",
"aggregation: type": "Uncoupled-MIS",
"smoother: type (level 0)": "symmetric Gauss-Seidel",
"smoother: damping factor": 0.67,
"smoother: sweeps": 1,
"smoother: pre or post": "both",
"PDE equations": 1,
"energy minimization: enable": True,
"energy minimization: type": 2
}
# creates the preconditioner and computes it
Prec = ML.MultiLevelPreconditioner(Matrix, False)
Prec.SetParameterList(MLList)
Prec.ComputePreconditioner()
# sets up the solver, specifies Prec as preconditioner, and
# solves using CG.
Solver = AztecOO.AztecOO(Matrix, LHS, RHS)
Solver.SetPrecOperator(Prec)
Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
Solver.SetAztecOption(AztecOO.AZ_kspace, 30)
Solver.SetAztecOption(AztecOO.AZ_output, 32)
Solver.Iterate(400, 1e-8)
del Prec
# b_epetra = NullSpace(bb,'vec')
# x_epetra = NullSpace(x.vector(),'vec')
print '\n\n\n DoF = ', Nb[0], '\n\n\n'
DoF[xx - 1] = Nb[0] - 1
mlList = {
"max levels": 200,
"output": 10,
"smoother: type": "symmetric Gauss-Seidel",
"aggregation: type": "Uncoupled"
}
prec = ML.MultiLevelPreconditioner(P_epetra, False)
prec.SetParameterList(mlList)
prec.ComputePreconditioner()
solver = AztecOO.AztecOO(A_epetra, x_epetra, b_epetra)
solver.SetPrecOperator(prec)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_output, 100)
err = solver.Iterate(1550, 1e-5)
print 'done'
# Ap = sp.save(A)
# SaveEpertaMatrix(A,"A")
# SaveEpertaMatrix(P.down_cast().mat(),"P")
# tic()
# solve(a==L1,uu,bcs)
# # solver = KrylovSolver("gmres", "amg")
def iterative_solve(self,
A,
rhs,
x0,
tol=1.e-5,
max_iter=200,
n_smooth=10,
smoother="gmres",
conv_history=False,
with_multiscale=True,
adapt_smoothing=True,
verbose=False):
history = dict()
history["n_smooth"] = []
history["residual"] = []
assert smoother in ["gmres", "ilu", "jacobi"]
rhs_sum = rhs.Comm().SumAll(np.sum(rhs[:]))
if not abs(rhs_sum) < abs(rhs.NormInf() * 1e-8):
logger.warn("sum of rhs does not equal to zero")
residual = Epetra.Vector(A.RangeMap())
ref_pressure = Epetra.Vector(A.RangeMap())
ref_residual = Epetra.Vector(A.RangeMap())
temp = Epetra.Vector(A.RangeMap())
A_diagonal = Epetra.Vector(A.RangeMap())
A.ExtractDiagonalCopy(A_diagonal)
ref_pressure[:] = rhs[:] / A_diagonal[:]
A.Multiply(False, ref_pressure, temp)
ref_residual[:] = rhs[:] - temp[:]
ref_residual_norm = ref_residual.NormInf()
AP = mat_multiply(A, self.P)
RAP_msfv = mat_multiply(self.R, AP, transpose_1=True)
RAP_msfe = mat_multiply(self.P, AP, transpose_1=True)
residual = self.__compute_residual(A, rhs, x0, residual)
error = Epetra.Vector(self.P.RangeMap())
if with_multiscale:
error = self.__solve_one_step(residual, RAP_msfv, self.R)
x0[:] += error[:]
if smoother == "ilu":
ilu = IFPACK.ILU(self.A)
ilu.Initialize()
ilu.Compute()
residual_prev_norm = 1.e50
solver = AztecOO.AztecOO()
if smoother == "ilu":
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_fixed_pt)
solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_dom_decomp)
solver.SetAztecOption(AztecOO.AZ_subdomain_solve, AztecOO.AZ_ilu)
if smoother == "gmres":
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_dom_decomp)
solver.SetAztecOption(AztecOO.AZ_subdomain_solve, AztecOO.AZ_ilu)
if smoother == "jacobi":
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_fixed_pt)
solver.SetAztecOption(AztecOO.AZ_precond, AztecOO.AZ_Jacobi)
solver.SetAztecOption(AztecOO.AZ_conv, AztecOO.AZ_rhs)
solver.SetAztecOption(AztecOO.AZ_output, 0)
for iteration in xrange(max_iter):
residual = self.__compute_residual(A, rhs, x0, residual)
logger.debug("Residual at iteration %d: %g", iteration,
residual.NormInf()[0] / ref_residual_norm)
error[:] = 0.0
solver.Iterate(A, error, residual, n_smooth, 1e-20)
x0[:] += error[:]
residual = self.__compute_residual(A, rhs, x0, residual)
if with_multiscale:
error = self.__solve_one_step(residual, RAP_msfe, self.P)
x0[:] += error[:]
if residual.NormInf()[0] / ref_residual_norm < tol:
logger.debug("Number of iterations for convergence: %d",
iteration)
break
if residual.NormInf(
)[0] >= 1.01 * residual_prev_norm and adapt_smoothing:
n_smooth = int(1.4 * n_smooth)
logger.warn(
"Solver stagnated, increasing number of smoothing steps to %d",
n_smooth)
residual_prev_norm = residual.NormInf()
history["n_smooth"].append(n_smooth)
history["residual"].append(residual_prev_norm[0] /
ref_residual_norm)
if verbose:
print iteration, history["residual"][-1]
residual = self.__compute_residual(A, rhs, x0, residual)
error[:] = 0.0
solver.Iterate(A, error, residual, n_smooth, 1e-20)
x0[:] += error[:]
residual = self.__compute_residual(A, rhs, x0, residual)
error = self.__solve_one_step(residual, RAP_msfv, self.R)
x0[:] += error[:]
history["n_smooth"].append(n_smooth)
history["residual"].append(residual.NormInf()[0] / ref_residual_norm)
# Check for convergence.:
lhs_fine = Epetra.Vector(A.DomainMap())
rhs_coarse = Epetra.Vector(self.R.DomainMap())
lhs_coarse = Epetra.Vector(self.R.DomainMap())
error = Epetra.Vector(self.R.DomainMap())
self.R.Multiply(True, rhs, rhs_coarse)
A.Multiply(False, x0, lhs_fine)
self.R.Multiply(True, lhs_fine, lhs_coarse)
max_lhs = lhs_coarse.NormInf()
max_rhs = rhs_coarse.NormInf()
error[:] = lhs_coarse[:] - rhs_coarse[:]
max_err = error.NormInf()
tol = max(max_lhs, max_rhs) * 1.e-4
assert np.allclose(
rhs_coarse[:], lhs_coarse[:],
atol=tol), "max_lhs:%e max_rhs:%e, max_error:%e " % (
max_lhs, max_rhs, max_err)
if conv_history:
return x0, history
else:
return x0
linearProblem = Epetra.LinearProblem(A, x, b)
if USE_DIRECT_SOLVER:
solver = Amesos.Lapack(linearProblem)
print "1) Performing symbolic factorizations..."
solver.SymbolicFactorization()
print "2) Performing numeric factorizations..."
solver.NumericFactorization()
print "3) Solving the linear system..."
ierr = solver.Solve()
else:
solver = AztecOO.AztecOO(linearProblem)
ierr = solver.Iterate(1000, 1e-9)
print " solver.Solve() return code = ", ierr
if ierr <= 1e-9:
print
print "|--------------------|"
print "|convergence achieved|"
print "|--------------------|"
print
mb.tag_set_data(pres_tag, volumes, np.asarray(x[v_ids]))
def upscale_perm_flow_based(self, domain, dim, boundary_meshset):
self.average_method = 'flow-based'
area = (
self.block_size[1] * self.block_size[2],
self.block_size[0] * self.block_size[2],
self.block_size[0] * self.block_size[1],
)
pres_tag = self.mb.tag_get_handle("Pressure", 1, types.MB_TYPE_DOUBLE,
types.MB_TAG_SPARSE, True)
std_map = Epetra.Map(len(domain), 0, self.comm)
linear_vals = np.arange(0, len(domain))
id_map = dict(zip(domain, linear_vals))
boundary_elms = set()
b = Epetra.Vector(std_map)
x = Epetra.Vector(std_map)
A = Epetra.CrsMatrix(Epetra.Copy, std_map, 3)
for elem in boundary_meshset:
if elem in boundary_elms:
continue
boundary_elms.add(elem)
idx = id_map[elem]
A.InsertGlobalValues(idx, [1], [idx])
b[idx] = self.mb.tag_get_data(self.boundary_dir[dim],
elem,
flat=True)
self.mb.tag_set_data(pres_tag, domain, np.repeat(0.0, len(domain)))
for elem in (set(domain) ^ boundary_elms):
adj_volumes = self.mesh_topo_util.get_bridge_adjacencies(
np.asarray([elem]), 2, 3, 0)
adj_volumes = [elems for elems in adj_volumes if elems in domain]
adj_volumes_set = set(adj_volumes)
elem_center = self.mesh_topo_util.get_average_position(
np.asarray([elem]))
K1 = self.mb.tag_get_data(self.perm_tag, [elem], flat=True)
adj_perms = []
for adjacencies in range(len(adj_volumes)):
adj_perms.append(
self.mb.tag_get_data(self.perm_tag, adj_volumes,
flat=True)[adjacencies *
9:(adjacencies + 1) * 9])
values = []
ids = []
for K2, adj in zip(adj_perms, adj_volumes_set):
adj_center = self.mesh_topo_util.get_average_position(
np.asarray([adj]))
N = elem_center - adj_center
N = N / np.sqrt(N[0]**2 + N[1]**2 + N[2]**2)
K1proj = np.dot(np.dot(N, K1.reshape([3, 3])), N)
K2proj = np.dot(np.dot(N, K2.reshape([3, 3])), N)
dl = np.linalg.norm((elem_center - adj_center) / 2)
K_eq = (2 * K1proj * K2proj) / (K1proj * dl + K2proj * dl)
values.append(-K_eq)
if adj in id_map:
ids.append(id_map[adj])
values.append(-sum(values))
idx = id_map[elem]
ids.append(idx)
A.InsertGlobalValues(idx, values, ids)
A.FillComplete()
"""
prec = ML.MultiLevelPreconditioner(A, False)
prec.SetParameterList(mlList)
solver = AztecOO.AztecOO(A, x, b)
solver.SetPrecOperator(prec)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cg)
solver.SetAztecOption(AztecOO.AZ_output, 16)
solver.Iterate(1550, 1e-5)
"""
linearProblem = Epetra.LinearProblem(A, x, b)
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_warnings)
solver.Iterate(200, 1e-9)
self.mb.tag_set_data(pres_tag, domain, np.asarray(x))
# Get the flux - should break down in another part
flow_rate = 0.0
total_area = 0.0
for elem in boundary_meshset:
elem_center = self.mesh_topo_util.get_average_position(
np.asarray([elem]))
adj_volumes = self.mesh_topo_util.get_bridge_adjacencies(
np.asarray([elem]), 2, 3)
adj_volumes_set = set(adj_volumes).intersection(set(domain))
adj_to_boundary_volumes = set()
for el in adj_volumes_set:
if el in boundary_meshset:
adj_to_boundary_volumes.add(el)
adj_volumes_set = adj_volumes_set - adj_to_boundary_volumes
for adj in adj_volumes_set:
adj_center = self.mesh_topo_util.get_average_position(
np.asarray([adj]))
N = elem_center - adj_center
N = N / np.sqrt(N[0]**2 + N[1]**2 + N[2]**2)
adj_pres = self.mb.tag_get_data(pres_tag, adj)
adj_perm = np.dot(
N,
np.dot(
self.mb.tag_get_data(self.perm_tag,
adj).reshape([3, 3]), N))
elem_perm = np.dot(
N,
np.dot(
self.mb.tag_get_data(self.perm_tag,
elem).reshape([3, 3]), N))
dl = np.linalg.norm((elem_center - adj_center) / 2)
K_equiv = (2 * adj_perm * elem_perm) / (adj_perm * dl +
elem_perm * dl)
flow_rate = flow_rate + area[dim] * K_equiv * adj_pres / dl
total_area = total_area + area[dim]
perm = flow_rate * dl / total_area
return perm
"aggregation: type" : "Uncoupled", "smoother: type (level 0)" : "Aztec", "smoother: damping factor": 0.67, "smoother: sweeps": 1, "smoother: pre or post": "post", "smoother: type (level 1)" : "Aztec", "PDE equations": 5, "aggregation: use tentative restriction": False, "aggregation: damping factor": 0.000, "eigen-analysis: type": "power-method", "aggregation: block scaling": False, "energy minimization: enable": False, "energy minimization: type": 2 } # creates the preconditioner and computes it Prec = ML.MultiLevelPreconditioner(Matrix, False) Prec.SetParameterList(MLList) Prec.ComputePreconditioner() # sets up the solver, specifies Prec as preconditioner, and # solves using CG. Solver = AztecOO.AztecOO(Matrix, LHS, RHS) Solver.SetPrecOperator(Prec) Solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres); Solver.SetAztecOption(AztecOO.AZ_kspace, 30); Solver.SetAztecOption(AztecOO.AZ_output, 1); Solver.Iterate(200, 1e-8) del Prec
def solve_linear_problem(self):
linearProblem = Epetra.LinearProblem(self.T, self.x, self.Q)
solver = AztecOO.AztecOO(linearProblem)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_output, AztecOO.AZ_none)
solver.Iterate(2000, 1e-16)
"smoother: type (level 4)": "Hiptmair",
"smoother: type (level 5)": "Hiptmair",
"smoother: type (level 6)": "Hiptmair",
"smoother: Hiptmair efficient symmetric": True,
"subsmoother: type": "Chebyshev",
"subsmoother: Chebyshev alpha": 27.0,
"subsmoother: node sweeps": 4,
"subsmoother: edge sweeps": 4,
"coarse: type": "Amesos-KLU",
"coarse: max size": 128,
"coarse: pre or post": "post",
"coarse: sweeps": 1
}
comm = Epetra.PyComm()
C = scipy_csr_matrix2CrsMatrix(System["C"].tocsr(), comm)
CurlCurl = scipy_csr_matrix2CrsMatrix(System["CurlCurl"].tocsr(), comm)
node = scipy_csr_matrix2CrsMatrix(System["node"].tocsr(), comm)
ML_Hiptmair = ML.MultiLevelPreconditioner(CurlCurl, C, node, MLList)
ML_Hiptmair.ComputePreconditioner()
x = System["rhs"][0]
b_epetra = TrilinosIO._numpyToTrilinosVector(x)
x_epetra = TrilinosIO._numpyToTrilinosVector(System["rhs"] * 0)
solver = AztecOO.AztecOO(CurlCurl, x_epetra, b_epetra)
solver.SetPrecOperator(ML_Hiptmair)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cg)
solver.SetAztecOption(AztecOO.AZ_output, 50)
err = solver.Iterate(155000, 1e-10)
bcs.apply(PP)
# bcs.apply(BB)make
P = as_backend_type(PP).mat()
B = as_backend_type(BB).mat()
L = as_backend_type(LL).mat()
b = as_backend_type(bb).vec()
x = Epetra.Vector(0 * bb.array())
# exit
# A = as_backend_type(AA).mat()
print toc()
ML_Hiptmair = ml.MultiLevelPreconditioner(P, B, L, MLList)
ML_Hiptmair.ComputePreconditioner()
# ML_Hiptmair.ComputePreconditioner()
solver = AztecOO.AztecOO(A, x, b)
solver.SetPrecOperator(ML_Hiptmair)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_gmres)
solver.SetAztecOption(AztecOO.AZ_output, 16)
err = solver.Iterate(1550, 1e-5)
# if (Solving == 'Direct'):
# ksp = PETSc.KSP().create()
# ksp.setOperators(A)
# ksp.setFromOptions()
# ksp.setType(ksp.Type.PREONLY)
# pc = ksp.getPC()
# pc.setFromOptions()
# pc.setType("lu")
if pol:
DOp = QUDestripeOperator(pix, tmap_local, tmap_glob, umap_local,
umap_glob, hits_glob, baseline_lengths,
data, comm, num_baselines)
else:
DOp = TDestripeOperator(pix, tmap_local, tmap_glob, hits_glob,
baseline_lengths, comm)
# Create the preconditioning operator
PrecOp = PrecOperator(baseline_lengths, comm)
# Aggregate the Destriping Operator, first guess and RHS in Linear Problem object
LinProb = Epetra.LinearProblem(DOp, baselines, RHS)
# Solve the problem iteratively using GMRES
IterSolver = AztecOO.AztecOO(LinProb)
IterSolver.SetPrecOperator(PrecOp)
with tm.TimeMonitor("Iterations"):
IterSolver.Iterate(gmres_iterations, gmres_residual)
# Write the baselines to disk, remove baselines from data, bin and write destriped maps
with tm.TimeMonitor("After destriping"):
if pol:
mpi_writer.write("baselinesQ.bin",
baselines.array[0][:num_baselines])
mpi_writer.write("baselinesU.bin",
baselines.array[0][num_baselines:])
data['Q'] -= np.repeat(baselines.array[0][:num_baselines],
baseline_lengths)
scipy.io.savemat( "System.mat", {"CurlCurl":Acurl.sparray(),"node":Anode.sparray(),"C":C,"rhs":b.array()},oned_as='row')
scipy.io.savemat( "node.mat", {"node":Anode.sparray()},oned_as='row')
scipy.io.savemat( "rhs.mat", {"rhs":b.array()},oned_as='row')
C = scipy_csr_matrix2CrsMatrix(C, comm)
Acurl = scipy_csr_matrix2CrsMatrix(Acurl.sparray(), comm)
Anode = scipy_csr_matrix2CrsMatrix(Anode.sparray(), comm)
# Acurl = as_backend_type(Acurl).mat()
# Anode = as_backend_type(Anode).mat()
ML_Hiptmair = ML.MultiLevelPreconditioner(Acurl,C,Anode,MLList,True)
ML_Hiptmair.ComputePreconditioner()
x = Function(V)
b_epetra = x_epetra = TrilinosIO._numpyToTrilinosVector(b.array())
x_epetra = TrilinosIO._numpyToTrilinosVector(x.vector().array())
tic()
#u = M.SolveSystem(A,b,V,"cg","amg",1e-6,1e-6,1)
print toc()
import PyTrilinos.AztecOO as AztecOO
solver = AztecOO.AztecOO(Acurl, x_epetra, b_epetra)
solver.SetPrecOperator(ML_Hiptmair)
solver.SetAztecOption(AztecOO.AZ_solver, AztecOO.AZ_cg);
solver.SetAztecOption(AztecOO.AZ_output, 50);
err = solver.Iterate(155000, 1e-10)
x = Function(V)
x.vector()[:] = TrilinosIO._trilinosToNumpyVector(x_epetra)
def solve(self):
linear_problem = Epetra.LinearProblem(self.A, self.x, self.b)
solver = AztecOO.AztecOO(linear_problem)
solver.Iterate(10000, 1.e-5)
return
