def __init__(self, A, solver=default_solver, parameters={}):
from PyTrilinos import Epetra, Amesos
from dolfin import info
from time import time
self.A = A # Keep reference
T = time()
self.problem = Epetra.LinearProblem()
self.problem.SetOperator(A.down_cast().mat())
self.solver = Amesos.Factory().Create(solver, self.problem)
if (self.solver == None):
raise RuntimeError('Failed to create Amesos solver: ' + solver)
# This prevents a use-after-free crash for the MPI communicator. It
# enforces that the solver is destroyed before A.
self.solver.A = A
self.solver.SetParameters(parameters)
if self.solver is None:
raise RuntimeError("Unknown solver '%s'" % solver)
err = self.solver.SymbolicFactorization()
if err != 0:
raise RuntimeError("Amesos " + solver +
" symbolic factorization failed, err=%d" % err)
err = self.solver.NumericFactorization()
if err != 0:
raise RuntimeError("Amesos " + solver +
" numeric factorization failed, err=%d" % err)
info('constructed direct solver (using %s) in %.2f s' %
(solver, time() - T))
def __init__(self,
tolerance=1e-10,
iterations=10,
precon=None,
maxIterations=10):
"""
:Parameters:
- `tolerance`: The required error tolerance.
- `iterations`: The maximum number of iterative steps to perform.
"""
iterations = min(iterations, maxIterations)
TrilinosSolver.__init__(self,
tolerance=tolerance,
iterations=iterations,
precon=None)
if precon is not None:
import warnings
warnings.warn(
"Trilinos KLU solver does not accept preconditioners.",
UserWarning,
stacklevel=2)
self.Factory = Amesos.Factory()
def solve_direct(A, rhs, x=None):
if x is None:
x = Epetra.Vector(A.RangeMap())
problem = Epetra.LinearProblem(A, x, rhs)
solver = Amesos.Klu(problem)
ierr = solver.Solve()
assert (ierr == 0), ierr
return x
def query(which=None):
"""Return list of available solver backends, or True/False if given a solver name"""
from PyTrilinos import Amesos
factory = Amesos.Factory()
if which is None:
avail = []
for s in serial_solvers + parallel_solvers:
if factory.Query(s):
avail.append(s)
return avail
return factory.Query(which)
def direct_solve(self, jac, rhs):
'''Currently unused direct solver that was used for testing.'''
A = Epetra.CrsMatrix(Epetra.Copy, self.map, 27)
for i in range(len(jac.begA)-1):
if i == self.dim:
A[i, i] = -1
continue
for j in range(jac.begA[i], jac.begA[i+1]):
if jac.jcoA[j] != self.dim:
A[i, jac.jcoA[j]] = jac.coA[j]
A.FillComplete()
rhs[self.dim] = 0
x = Vector(rhs)
problem = Epetra.LinearProblem(A, x, rhs)
factory = Amesos.Factory()
solver = factory.Create('Klu', problem)
solver.SymbolicFactorization()
solver.NumericFactorization()
solver.Solve()
return x
def __init__(self, ckt):
# Init base class
_NLFunction.__init__(self)
# Save ckt instance
self.ckt = ckt
# Make sure circuit is ready (analysis should take care)
assert ckt.nD_ref
# Allocate matrices/vectors
# Create Epetra CrsMatrix: need communicator and map
comm = Epetra.SerialComm()
# Map: numelem, base index, communicator
map = Epetra.Map(self.ckt.nD_dimension, 0, comm)
# TODO: Find a good estimate for nnzRow
nnzRow = min(10, self.ckt.nD_dimension)
# G here is G1 = G + G0 in documentation
self.G = Epetra.CrsMatrix(Epetra.Copy, map, nnzRow)
# Jac is (G1 + di/dv) in doc
self.Jac = Epetra.CrsMatrix(Epetra.Copy, map, nnzRow)
# Allocate several vectors
self.xVec = Epetra.Vector(map)
self.iVec = Epetra.Vector(map)
self.sVec = Epetra.Vector(map)
self.errVec = Epetra.Vector(map)
self.deltaxVec = Epetra.Vector(map)
if hasattr(self.ckt, 'nD_namRClist'):
# Allocate external currents vector
self.extSVec = np.empty(self.ckt.nD_dimension)
# Generate G and Jac sparse matrices
# ----------------------------------
self._get_Gnz()
self.Jacnz = [([], []) for i in xrange(self.ckt.nD_dimension)]
# Insert linear contributions into G and Jac (to fix structure)
for row, data in enumerate(self.Gnz):
self.G.InsertGlobalValues(row, *data)
self.Jac.InsertGlobalValues(row, *data)
# Finalize G (which should not change any more unless parameter sweep)
assert not self.G.FillComplete()
assert not self.G.OptimizeStorage()
# Add nonlinear contribution to Jac. Insert ones for now as the
# matrix values will be discarded anyway.
for elem in self.ckt.nD_nlinElem:
outJac = np.ones((len(elem.csOutPorts), len(elem.controlPorts)),
dtype=float)
set_Jac(self.Jacnz, elem.nD_cpos, elem.nD_cneg, elem.nD_vpos,
elem.nD_vneg, outJac)
for row, data in enumerate(self.Jacnz):
self.Jac.InsertGlobalValues(row, *data)
# Finalize and symbolically factor Jac
assert not self.Jac.FillComplete()
assert not self.Jac.OptimizeStorage()
# Create pytrilinos solver
problem = Epetra.LinearProblem(self.Jac, self.deltaxVec, self.errVec)
factory = Amesos.Factory()
self.solver = factory.Create("Klu", problem)
assert not self.solver.SymbolicFactorization()
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)
# }
# 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(20000, 1e-10)
tic()
problem = Epetra.LinearProblem(A_epetra, x_epetra, b_epetra)
print '\n\n\n\n\n\n'
factory = Amesos.Factory()
solver = factory.Create("Amesos_Umfpack", problem)
# solver = factory.Create("MUMPS", problem)
amesosList = {"PrintTiming": True, "PrintStatus": True}
solver.SetParameters(amesosList)
solver.SymbolicFactorization()
solver.NumericFactorization()
solver.Solve()
soln = problem.GetLHS()
print "||x_computed||_2 =", soln.Norm2()
# solver.PrintTiming()
print '\n\n\n\n\n\n'
if case == 1:
ue = Expression(("20*x[0]*pow(x[1],3)", "5*pow(x[0],4)-5*pow(x[1],4)"))
pe = Expression("60*pow(x[0],2)*x[1]-20*pow(x[1],3)+5")
print "Matrix fill took", time.time() - t0, "seconds... Ran over ", len(
volumes), "elems"
mb.tag_set_data(pres_tag, volumes, np.asarray(b[v_ids]))
if USE_DIRECT_SOLVER:
outfile_template = "Results/output_direct_{0}.vtk"
else:
outfile_template = "output_iterative_fine_grid{0}.vtk"
mb.write_file(outfile_template.format(0))
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
