def extend_hierarchy(levels,
strength,
aggregate,
smooth,
improve_candidates,
diagonal_dominance=False,
keep=True):
"""Service routine to implement the strength of connection, aggregation,
tentative prolongation construction, and prolongation smoothing. Called by
smoothed_aggregation_solver.
"""
def unpack_arg(v):
if isinstance(v, tuple):
return v[0], v[1]
else:
return v, {}
A = levels[-1].A
B = levels[-1].B
if A.symmetry == "nonsymmetric":
AH = A.H.asformat(A.format)
BH = levels[-1].BH
##
# Compute the strength-of-connection matrix C, where larger
# C[i,j] denote stronger couplings between i and j.
fn, kwargs = unpack_arg(strength[len(levels) - 1])
if fn == 'symmetric':
C = symmetric_strength_of_connection(A, **kwargs)
elif fn == 'classical':
C = classical_strength_of_connection(A, **kwargs)
elif fn == 'distance':
C = distance_strength_of_connection(A, **kwargs)
elif (fn == 'ode') or (fn == 'evolution'):
if 'B' in kwargs:
C = evolution_strength_of_connection(A, **kwargs)
else:
C = evolution_strength_of_connection(A, B, **kwargs)
elif fn == 'energy_based':
C = energy_based_strength_of_connection(A, **kwargs)
elif fn == 'predefined':
C = kwargs['C'].tocsr()
elif fn == 'algebraic_distance':
C = algebraic_distance(A, **kwargs)
elif fn is None:
C = A.tocsr()
else:
raise ValueError('unrecognized strength of connection method: %s' %
str(fn))
##
# Avoid coarsening diagonally dominant rows
flag, kwargs = unpack_arg(diagonal_dominance)
if flag:
C = eliminate_diag_dom_nodes(A, C, **kwargs)
##
# Compute the aggregation matrix AggOp (i.e., the nodal coarsening of A).
# AggOp is a boolean matrix, where the sparsity pattern for the k-th column
# denotes the fine-grid nodes agglomerated into k-th coarse-grid node.
fn, kwargs = unpack_arg(aggregate[len(levels) - 1])
if fn == 'standard':
AggOp = standard_aggregation(C, **kwargs)[0]
elif fn == 'naive':
AggOp = naive_aggregation(C, **kwargs)[0]
elif fn == 'lloyd':
AggOp = lloyd_aggregation(C, **kwargs)[0]
elif fn == 'predefined':
AggOp = kwargs['AggOp'].tocsr()
else:
raise ValueError('unrecognized aggregation method %s' % str(fn))
##
# Improve near nullspace candidates by relaxing on A B = 0
fn, kwargs = unpack_arg(improve_candidates[len(levels) - 1])
if fn is not None:
b = np.zeros((A.shape[0], 1), dtype=A.dtype)
B = relaxation_as_linear_operator((fn, kwargs), A, b) * B
levels[-1].B = B
if A.symmetry == "nonsymmetric":
BH = relaxation_as_linear_operator((fn, kwargs), AH, b) * BH
levels[-1].BH = BH
##
# Compute the tentative prolongator, T, which is a tentative interpolation
# matrix from the coarse-grid to the fine-grid. T exactly interpolates
# B_fine = T B_coarse.
T, B = fit_candidates(AggOp, B)
if A.symmetry == "nonsymmetric":
TH, BH = fit_candidates(AggOp, BH)
##
# Smooth the tentative prolongator, so that it's accuracy is greatly
# improved for algebraically smooth error.
fn, kwargs = unpack_arg(smooth[len(levels) - 1])
if fn == 'jacobi':
P = jacobi_prolongation_smoother(A, T, C, B, **kwargs)
elif fn == 'richardson':
P = richardson_prolongation_smoother(A, T, **kwargs)
elif fn == 'energy':
P = energy_prolongation_smoother(A, T, C, B, None, (False, {}),
**kwargs)
elif fn is None:
P = T
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
##
# Compute the restriction matrix, R, which interpolates from the fine-grid
# to the coarse-grid. If A is nonsymmetric, then R must be constructed
# based on A.H. Otherwise R = P.H or P.T.
symmetry = A.symmetry
if symmetry == 'hermitian':
R = P.H
elif symmetry == 'symmetric':
R = P.T
elif symmetry == 'nonsymmetric':
fn, kwargs = unpack_arg(smooth[len(levels) - 1])
if fn == 'jacobi':
R = jacobi_prolongation_smoother(AH, TH, C, BH, **kwargs).H
elif fn == 'richardson':
R = richardson_prolongation_smoother(AH, TH, **kwargs).H
elif fn == 'energy':
R = energy_prolongation_smoother(AH, TH, C, BH, None, (False, {}),
**kwargs)
R = R.H
elif fn is None:
R = T.H
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
if keep:
levels[-1].C = C # strength of connection matrix
levels[-1].AggOp = AggOp # aggregation operator
levels[-1].T = T # tentative prolongator
levels[-1].P = P # smoothed prolongator
levels[-1].R = R # restriction operator
levels.append(multilevel_solver.level())
A = R * A * P # Galerkin operator
A.symmetry = symmetry
levels[-1].A = A
levels[-1].B = B # right near nullspace candidates
if A.symmetry == "nonsymmetric":
levels[-1].BH = BH # left near nullspace candidates
def extend_hierarchy(levels, strength, aggregate, smooth, improve_candidates,
diagonal_dominance=False, keep=True):
"""Service routine to implement the strength of connection, aggregation,
tentative prolongation construction, and prolongation smoothing. Called by
smoothed_aggregation_solver.
"""
def unpack_arg(v):
if isinstance(v, tuple):
return v[0], v[1]
else:
return v, {}
A = levels[-1].A
B = levels[-1].B
if A.symmetry == "nonsymmetric":
AH = A.H.asformat(A.format)
BH = levels[-1].BH
# Compute the strength-of-connection matrix C, where larger
# C[i,j] denote stronger couplings between i and j.
fn, kwargs = unpack_arg(strength[len(levels)-1])
if fn == 'symmetric':
C = symmetric_strength_of_connection(A, **kwargs)
elif fn == 'classical':
C = classical_strength_of_connection(A, **kwargs)
elif fn == 'distance':
C = distance_strength_of_connection(A, **kwargs)
elif (fn == 'ode') or (fn == 'evolution'):
if 'B' in kwargs:
C = evolution_strength_of_connection(A, **kwargs)
else:
C = evolution_strength_of_connection(A, B, **kwargs)
elif fn == 'energy_based':
C = energy_based_strength_of_connection(A, **kwargs)
elif fn == 'predefined':
C = kwargs['C'].tocsr()
elif fn == 'algebraic_distance':
C = algebraic_distance(A, **kwargs)
elif fn is None:
C = A.tocsr()
else:
raise ValueError('unrecognized strength of connection method: %s' %
str(fn))
# Avoid coarsening diagonally dominant rows
flag, kwargs = unpack_arg(diagonal_dominance)
if flag:
C = eliminate_diag_dom_nodes(A, C, **kwargs)
# Compute the aggregation matrix AggOp (i.e., the nodal coarsening of A).
# AggOp is a boolean matrix, where the sparsity pattern for the k-th column
# denotes the fine-grid nodes agglomerated into k-th coarse-grid node.
fn, kwargs = unpack_arg(aggregate[len(levels)-1])
if fn == 'standard':
AggOp = standard_aggregation(C, **kwargs)[0]
elif fn == 'naive':
AggOp = naive_aggregation(C, **kwargs)[0]
elif fn == 'lloyd':
AggOp = lloyd_aggregation(C, **kwargs)[0]
elif fn == 'predefined':
AggOp = kwargs['AggOp'].tocsr()
else:
raise ValueError('unrecognized aggregation method %s' % str(fn))
# Improve near nullspace candidates by relaxing on A B = 0
fn, kwargs = unpack_arg(improve_candidates[len(levels)-1])
if fn is not None:
b = np.zeros((A.shape[0], 1), dtype=A.dtype)
B = relaxation_as_linear_operator((fn, kwargs), A, b) * B
levels[-1].B = B
if A.symmetry == "nonsymmetric":
BH = relaxation_as_linear_operator((fn, kwargs), AH, b) * BH
levels[-1].BH = BH
# Compute the tentative prolongator, T, which is a tentative interpolation
# matrix from the coarse-grid to the fine-grid. T exactly interpolates
# B_fine = T B_coarse.
T, B = fit_candidates(AggOp, B)
if A.symmetry == "nonsymmetric":
TH, BH = fit_candidates(AggOp, BH)
# Smooth the tentative prolongator, so that it's accuracy is greatly
# improved for algebraically smooth error.
fn, kwargs = unpack_arg(smooth[len(levels)-1])
if fn == 'jacobi':
P = jacobi_prolongation_smoother(A, T, C, B, **kwargs)
elif fn == 'richardson':
P = richardson_prolongation_smoother(A, T, **kwargs)
elif fn == 'energy':
P = energy_prolongation_smoother(A, T, C, B, None, (False, {}),
**kwargs)
elif fn is None:
P = T
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# Compute the restriction matrix, R, which interpolates from the fine-grid
# to the coarse-grid. If A is nonsymmetric, then R must be constructed
# based on A.H. Otherwise R = P.H or P.T.
symmetry = A.symmetry
if symmetry == 'hermitian':
R = P.H
elif symmetry == 'symmetric':
R = P.T
elif symmetry == 'nonsymmetric':
fn, kwargs = unpack_arg(smooth[len(levels)-1])
if fn == 'jacobi':
R = jacobi_prolongation_smoother(AH, TH, C, BH, **kwargs).H
elif fn == 'richardson':
R = richardson_prolongation_smoother(AH, TH, **kwargs).H
elif fn == 'energy':
R = energy_prolongation_smoother(AH, TH, C, BH, None, (False, {}),
**kwargs)
R = R.H
elif fn is None:
R = T.H
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
if keep:
levels[-1].C = C # strength of connection matrix
levels[-1].AggOp = AggOp # aggregation operator
levels[-1].T = T # tentative prolongator
levels[-1].P = P # smoothed prolongator
levels[-1].R = R # restriction operator
levels.append(multilevel_solver.level())
A = R * A * P # Galerkin operator
A.symmetry = symmetry
levels[-1].A = A
levels[-1].B = B # right near nullspace candidates
if A.symmetry == "nonsymmetric":
levels[-1].BH = BH # left near nullspace candidates
def general_setup_stage(ml, symmetry, candidate_iters, prepostsmoother,
smooth, eliminate_local, coarse_solver, work):
"""
Computes additional candidates and improvements
following Algorithm 4 in Brezina et al.
Parameters
----------
candidate_iters
number of test relaxation iterations
epsilon
minimum acceptable relaxation convergence factor
References
----------
.. [1] Brezina, Falgout, MacLachlan, Manteuffel, McCormick, and Ruge
"Adaptive Smoothed Aggregation (alphaSA) Multigrid"
SIAM Review Volume 47, Issue 2 (2005)
http://www.cs.umn.edu/~maclach/research/aSA2.pdf
"""
def make_bridge(T):
M, N = T.shape
K = T.blocksize[0]
bnnz = T.indptr[-1]
# the K+1 represents the new dof introduced by the new candidate. the
# bridge 'T' ignores this new dof and just maps zeros there
data = numpy.zeros((bnnz, K+1, K), dtype=T.dtype)
data[:, :-1, :] = T.data
return bsr_matrix((data, T.indices, T.indptr),
shape=((K + 1) * (M / K), N))
def expand_candidates(B_old, nodesize):
# insert a new dof that is always zero, to create NullDim+1 dofs per
# node in B
NullDim = B_old.shape[1]
nnodes = B_old.shape[0] / nodesize
Bnew = numpy.zeros((nnodes, nodesize+1, NullDim), dtype=B_old.dtype)
Bnew[:, :-1, :] = B_old.reshape(nnodes, nodesize, NullDim)
return Bnew.reshape(-1, NullDim)
levels = ml.levels
x = scipy.rand(levels[0].A.shape[0], 1)
if levels[0].A.dtype == complex:
x = x + 1.0j*scipy.rand(levels[0].A.shape[0], 1)
b = numpy.zeros_like(x)
x = ml.solve(b, x0=x, tol=float(numpy.finfo(numpy.float).tiny),
maxiter=candidate_iters)
work[:] += ml.operator_complexity()*ml.levels[0].A.nnz*candidate_iters*2
T0 = levels[0].T.copy()
#TEST FOR CONVERGENCE HERE
for i in range(len(ml.levels) - 2):
# alpha-SA paper does local elimination here, but after talking
# to Marian, its not clear that this helps things
# fn, kwargs = unpack_arg(eliminate_local)
# if fn == True:
# eliminate_local_candidates(x,levels[i].AggOp,levels[i].A,
# levels[i].T, **kwargs)
# add candidate to B
B = numpy.hstack((levels[i].B, x.reshape(-1, 1)))
# construct Ptent
T, R = fit_candidates(levels[i].AggOp, B)
levels[i].T = T
x = R[:, -1].reshape(-1, 1)
# smooth P
fn, kwargs = unpack_arg(smooth[i])
if fn == 'jacobi':
levels[i].P = jacobi_prolongation_smoother(levels[i].A, T,
levels[i].C, R,
**kwargs)
elif fn == 'richardson':
levels[i].P = richardson_prolongation_smoother(levels[i].A, T,
**kwargs)
elif fn == 'energy':
levels[i].P = energy_prolongation_smoother(levels[i].A, T,
levels[i].C, R, None,
(False, {}), **kwargs)
x = R[:, -1].reshape(-1, 1)
elif fn is None:
levels[i].P = T
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# construct R
if symmetry == 'symmetric': # R should reflect A's structure
levels[i].R = levels[i].P.T.asformat(levels[i].P.format)
elif symmetry == 'hermitian':
levels[i].R = levels[i].P.H.asformat(levels[i].P.format)
# construct coarse A
levels[i+1].A = levels[i].R * levels[i].A * levels[i].P
# construct bridging P
T_bridge = make_bridge(levels[i+1].T)
R_bridge = levels[i+2].B
# smooth bridging P
fn, kwargs = unpack_arg(smooth[i+1])
if fn == 'jacobi':
levels[i+1].P = jacobi_prolongation_smoother(levels[i+1].A,
T_bridge,
levels[i+1].C,
R_bridge, **kwargs)
elif fn == 'richardson':
levels[i+1].P = richardson_prolongation_smoother(levels[i+1].A,
T_bridge,
**kwargs)
elif fn == 'energy':
levels[i+1].P = energy_prolongation_smoother(levels[i+1].A,
T_bridge,
levels[i+1].C,
R_bridge, None,
(False, {}), **kwargs)
elif fn is None:
levels[i+1].P = T_bridge
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# construct the "bridging" R
if symmetry == 'symmetric': # R should reflect A's structure
levels[i+1].R = levels[i+1].P.T.asformat(levels[i+1].P.format)
elif symmetry == 'hermitian':
levels[i+1].R = levels[i+1].P.H.asformat(levels[i+1].P.format)
# run solver on candidate
solver = multilevel_solver(levels[i+1:], coarse_solver=coarse_solver)
change_smoothers(solver, presmoother=prepostsmoother,
postsmoother=prepostsmoother)
x = solver.solve(numpy.zeros_like(x), x0=x,
tol=float(numpy.finfo(numpy.float).tiny),
maxiter=candidate_iters)
work[:] += 2 * solver.operator_complexity() * solver.levels[0].A.nnz *\
candidate_iters*2
# update values on next level
levels[i+1].B = R[:, :-1].copy()
levels[i+1].T = T_bridge
# note that we only use the x from the second coarsest level
fn, kwargs = unpack_arg(prepostsmoother)
for lvl in reversed(levels[:-2]):
x = lvl.P * x
work[:] += lvl.A.nnz*candidate_iters*2
if fn == 'gauss_seidel':
# only relax at nonzeros, so as not to mess up any locally dropped
# candidates
indices = numpy.ravel(x).nonzero()[0]
gauss_seidel_indexed(lvl.A, x, numpy.zeros_like(x), indices,
iterations=candidate_iters, sweep='symmetric')
elif fn == 'gauss_seidel_ne':
gauss_seidel_ne(lvl.A, x, numpy.zeros_like(x),
iterations=candidate_iters, sweep='symmetric')
elif fn == 'gauss_seidel_nr':
gauss_seidel_nr(lvl.A, x, numpy.zeros_like(x),
iterations=candidate_iters, sweep='symmetric')
elif fn == 'jacobi':
jacobi(lvl.A, x, numpy.zeros_like(x), iterations=1,
omega=1.0 / rho_D_inv_A(lvl.A))
elif fn == 'richardson':
polynomial(lvl.A, x, numpy.zeros_like(x), iterations=1,
coeffients=[1.0/approximate_spectral_radius(lvl.A)])
elif fn == 'gmres':
x[:] = (gmres(lvl.A, numpy.zeros_like(x), x0=x,
maxiter=candidate_iters)[0]).reshape(x.shape)
else:
raise TypeError('Unrecognized smoother')
# x will be dense again, so we have to drop locally again
elim, elim_kwargs = unpack_arg(eliminate_local)
if elim is True:
x = x/norm(x, 'inf')
eliminate_local_candidates(x, levels[0].AggOp, levels[0].A, T0,
**elim_kwargs)
return x.reshape(-1, 1)
def initial_setup_stage(A, symmetry, pdef, candidate_iters, epsilon,
max_levels, max_coarse, aggregate, prepostsmoother,
smooth, strength, work, initial_candidate=None):
"""
Computes a complete aggregation and the first near-nullspace candidate
following Algorithm 3 in Brezina et al.
Parameters
----------
candidate_iters
number of test relaxation iterations
epsilon
minimum acceptable relaxation convergence factor
References
----------
.. [1] Brezina, Falgout, MacLachlan, Manteuffel, McCormick, and Ruge
"Adaptive Smoothed Aggregation ($\alpha$SA) Multigrid"
SIAM Review Volume 47, Issue 2 (2005)
http://www.cs.umn.edu/~maclach/research/aSA2.pdf
"""
##
# Define relaxation routine
def relax(A, x):
fn, kwargs = unpack_arg(prepostsmoother)
if fn == 'gauss_seidel':
gauss_seidel(A, x, numpy.zeros_like(x),
iterations=candidate_iters, sweep='symmetric')
elif fn == 'gauss_seidel_nr':
gauss_seidel_nr(A, x, numpy.zeros_like(x),
iterations=candidate_iters, sweep='symmetric')
elif fn == 'gauss_seidel_ne':
gauss_seidel_ne(A, x, numpy.zeros_like(x),
iterations=candidate_iters, sweep='symmetric')
elif fn == 'jacobi':
jacobi(A, x, numpy.zeros_like(x), iterations=1,
omega=1.0 / rho_D_inv_A(A))
elif fn == 'richardson':
polynomial(A, x, numpy.zeros_like(x), iterations=1,
coeffients=[1.0/approximate_spectral_radius(A)])
elif fn == 'gmres':
x[:] = (gmres(A, numpy.zeros_like(x), x0=x,
maxiter=candidate_iters)[0]).reshape(x.shape)
else:
raise TypeError('Unrecognized smoother')
# flag for skipping steps f-i in step 4
skip_f_to_i = True
#step 1
A_l = A
if initial_candidate is None:
x = scipy.rand(A_l.shape[0], 1)
if A_l.dtype == complex:
x = x + 1.0j*scipy.rand(A_l.shape[0], 1)
else:
x = numpy.array(initial_candidate, dtype=A_l.dtype)
#step 2
relax(A_l, x)
work[:] += A_l.nnz * candidate_iters*2
# step 3
# not advised to stop the iteration here: often the first relaxation pass
# _is_ good, but the remaining passes are poor
# if x_A_x/x_A_x_old < epsilon:
# # relaxation alone is sufficient
# print 'relaxation alone works: %g'%(x_A_x/x_A_x_old)
# return x, []
# step 4
As = [A]
xs = [x]
Ps = []
AggOps = []
StrengthOps = []
while A.shape[0] > max_coarse and max_levels > 1:
# The real check to break from the while loop is below
# Begin constructing next level
fn, kwargs = unpack_arg(strength[len(As)-1]) # step 4b
if fn == 'symmetric':
C_l = symmetric_strength_of_connection(A_l, **kwargs)
# Diagonal must be nonzero
C_l = C_l + eye(C_l.shape[0], C_l.shape[1], format='csr')
elif fn == 'classical':
C_l = classical_strength_of_connection(A_l, **kwargs)
# Diagonal must be nonzero
C_l = C_l + eye(C_l.shape[0], C_l.shape[1], format='csr')
if isspmatrix_bsr(A_l):
C_l = amalgamate(C, A_l.blocksize[0])
elif (fn == 'ode') or (fn == 'evolution'):
C_l = evolution_strength_of_connection(A_l,
numpy.ones(
(A_l.shape[0], 1),
dtype=A.dtype),
**kwargs)
elif fn == 'predefined':
C_l = kwargs['C'].tocsr()
elif fn is None:
C_l = A_l.tocsr()
else:
raise ValueError('unrecognized strength of connection method: %s' %
str(fn))
# In SA, strength represents "distance", so we take magnitude of
# complex values
if C_l.dtype == complex:
C_l.data = numpy.abs(C_l.data)
# Create a unified strength framework so that large values represent
# strong connections and small values represent weak connections
if (fn == 'ode') or (fn == 'evolutin') or (fn == 'energy_based'):
C_l.data = 1.0 / C_l.data
# aggregation
fn, kwargs = unpack_arg(aggregate[len(As) - 1])
if fn == 'standard':
AggOp = standard_aggregation(C_l, **kwargs)[0]
elif fn == 'lloyd':
AggOp = lloyd_aggregation(C_l, **kwargs)[0]
elif fn == 'predefined':
AggOp = kwargs['AggOp'].tocsr()
else:
raise ValueError('unrecognized aggregation method %s' % str(fn))
T_l, x = fit_candidates(AggOp, x) # step 4c
fn, kwargs = unpack_arg(smooth[len(As)-1]) # step 4d
if fn == 'jacobi':
P_l = jacobi_prolongation_smoother(A_l, T_l, C_l, x, **kwargs)
elif fn == 'richardson':
P_l = richardson_prolongation_smoother(A_l, T_l, **kwargs)
elif fn == 'energy':
P_l = energy_prolongation_smoother(A_l, T_l, C_l, x, None,
(False, {}), **kwargs)
elif fn is None:
P_l = T_l
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# R should reflect A's structure # step 4e
if symmetry == 'symmetric':
A_l = P_l.T.asformat(P_l.format) * A_l * P_l
elif symmetry == 'hermitian':
A_l = P_l.H.asformat(P_l.format) * A_l * P_l
StrengthOps.append(C_l)
AggOps.append(AggOp)
Ps.append(P_l)
As.append(A_l)
# skip to step 5 as in step 4e
if (A_l.shape[0] <= max_coarse) or (len(AggOps) + 1 >= max_levels):
break
if not skip_f_to_i:
x_hat = x.copy() # step 4g
relax(A_l, x) # step 4h
work[:] += A_l.nnz*candidate_iters*2
if pdef is True:
x_A_x = numpy.dot(numpy.conjugate(x).T, A_l*x)
xhat_A_xhat = numpy.dot(numpy.conjugate(x_hat).T, A_l*x_hat)
err_ratio = (x_A_x/xhat_A_xhat)**(1.0/candidate_iters)
else:
# use A.H A inner-product
Ax = A_l * x
#Axhat = A_l * x_hat
x_A_x = numpy.dot(numpy.conjugate(Ax).T, Ax)
xhat_A_xhat = numpy.dot(numpy.conjugate(x_hat).T, A_l*x_hat)
err_ratio = (x_A_x/xhat_A_xhat)**(1.0/candidate_iters)
if err_ratio < epsilon: # step 4i
#print "sufficient convergence, skipping"
skip_f_to_i = True
if x_A_x == 0:
x = x_hat # need to restore x
else:
# just carry out relaxation, don't check for convergence
relax(A_l, x) # step 4h
work[:] += 2 * A_l.nnz * candidate_iters
# store xs for diagnostic use and for use in step 5
xs.append(x)
# step 5
# Extend coarse-level candidate to the finest level
# --> note that we start with the x from the second coarsest level
x = xs[-1]
# make sure that xs[-1] has been relaxed by step 4h, i.e. relax(As[-2], x)
for lev in range(len(Ps)-2, -1, -1): # lev = coarsest ... finest-1
P = Ps[lev] # I: lev --> lev+1
A = As[lev] # A on lev+1
x = P * x
relax(A, x)
work[:] += A.nnz*candidate_iters*2
# Set predefined strength of connection and aggregation
if len(AggOps) > 1:
aggregate = [('predefined', {'AggOp': AggOps[i]})
for i in range(len(AggOps))]
strength = [('predefined', {'C': StrengthOps[i]})
for i in range(len(StrengthOps))]
return x, aggregate, strength # first candidate
def general_setup_stage(ml, symmetry, candidate_iters, prepostsmoother, smooth,
eliminate_local, coarse_solver, work):
"""
Computes additional candidates and improvements
following Algorithm 4 in Brezina et al.
Parameters
----------
candidate_iters
number of test relaxation iterations
epsilon
minimum acceptable relaxation convergence factor
References
----------
.. [1] Brezina, Falgout, MacLachlan, Manteuffel, McCormick, and Ruge
"Adaptive Smoothed Aggregation (alphaSA) Multigrid"
SIAM Review Volume 47, Issue 2 (2005)
http://www.cs.umn.edu/~maclach/research/aSA2.pdf
"""
def make_bridge(T):
M, N = T.shape
K = T.blocksize[0]
bnnz = T.indptr[-1]
# the K+1 represents the new dof introduced by the new candidate. the
# bridge 'T' ignores this new dof and just maps zeros there
data = np.zeros((bnnz, K + 1, K), dtype=T.dtype)
data[:, :-1, :] = T.data
return bsr_matrix((data, T.indices, T.indptr),
shape=((K + 1) * (M / K), N))
def expand_candidates(B_old, nodesize):
# insert a new dof that is always zero, to create NullDim+1 dofs per
# node in B
NullDim = B_old.shape[1]
nnodes = B_old.shape[0] / nodesize
Bnew = np.zeros((nnodes, nodesize + 1, NullDim), dtype=B_old.dtype)
Bnew[:, :-1, :] = B_old.reshape(nnodes, nodesize, NullDim)
return Bnew.reshape(-1, NullDim)
levels = ml.levels
x = sp.rand(levels[0].A.shape[0], 1)
if levels[0].A.dtype == complex:
x = x + 1.0j * sp.rand(levels[0].A.shape[0], 1)
b = np.zeros_like(x)
x = ml.solve(b,
x0=x,
tol=float(np.finfo(np.float).tiny),
maxiter=candidate_iters)
work[:] += ml.operator_complexity(
) * ml.levels[0].A.nnz * candidate_iters * 2
T0 = levels[0].T.copy()
# TEST FOR CONVERGENCE HERE
for i in range(len(ml.levels) - 2):
# alpha-SA paper does local elimination here, but after talking
# to Marian, its not clear that this helps things
# fn, kwargs = unpack_arg(eliminate_local)
# if fn == True:
# eliminate_local_candidates(x,levels[i].AggOp,levels[i].A,
# levels[i].T, **kwargs)
# add candidate to B
B = np.hstack((levels[i].B, x.reshape(-1, 1)))
# construct Ptent
T, R = fit_candidates(levels[i].AggOp, B)
levels[i].T = T
x = R[:, -1].reshape(-1, 1)
# smooth P
fn, kwargs = unpack_arg(smooth[i])
if fn == 'jacobi':
levels[i].P = jacobi_prolongation_smoother(levels[i].A, T,
levels[i].C, R,
**kwargs)
elif fn == 'richardson':
levels[i].P = richardson_prolongation_smoother(
levels[i].A, T, **kwargs)
elif fn == 'energy':
levels[i].P = energy_prolongation_smoother(levels[i].A, T,
levels[i].C, R, None,
(False, {}), **kwargs)
x = R[:, -1].reshape(-1, 1)
elif fn is None:
levels[i].P = T
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# construct R
if symmetry == 'symmetric': # R should reflect A's structure
levels[i].R = levels[i].P.T.asformat(levels[i].P.format)
elif symmetry == 'hermitian':
levels[i].R = levels[i].P.H.asformat(levels[i].P.format)
# construct coarse A
levels[i + 1].A = levels[i].R * levels[i].A * levels[i].P
# construct bridging P
T_bridge = make_bridge(levels[i + 1].T)
R_bridge = levels[i + 2].B
# smooth bridging P
fn, kwargs = unpack_arg(smooth[i + 1])
if fn == 'jacobi':
levels[i + 1].P = jacobi_prolongation_smoother(
levels[i + 1].A, T_bridge, levels[i + 1].C, R_bridge, **kwargs)
elif fn == 'richardson':
levels[i + 1].P = richardson_prolongation_smoother(
levels[i + 1].A, T_bridge, **kwargs)
elif fn == 'energy':
levels[i + 1].P = energy_prolongation_smoother(
levels[i + 1].A, T_bridge, levels[i + 1].C, R_bridge, None,
(False, {}), **kwargs)
elif fn is None:
levels[i + 1].P = T_bridge
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# construct the "bridging" R
if symmetry == 'symmetric': # R should reflect A's structure
levels[i + 1].R = levels[i + 1].P.T.asformat(levels[i +
1].P.format)
elif symmetry == 'hermitian':
levels[i + 1].R = levels[i + 1].P.H.asformat(levels[i +
1].P.format)
# run solver on candidate
solver = multilevel_solver(levels[i + 1:], coarse_solver=coarse_solver)
change_smoothers(solver,
presmoother=prepostsmoother,
postsmoother=prepostsmoother)
x = solver.solve(np.zeros_like(x),
x0=x,
tol=float(np.finfo(np.float).tiny),
maxiter=candidate_iters)
work[:] += 2 * solver.operator_complexity() * solver.levels[0].A.nnz *\
candidate_iters*2
# update values on next level
levels[i + 1].B = R[:, :-1].copy()
levels[i + 1].T = T_bridge
# note that we only use the x from the second coarsest level
fn, kwargs = unpack_arg(prepostsmoother)
for lvl in reversed(levels[:-2]):
x = lvl.P * x
work[:] += lvl.A.nnz * candidate_iters * 2
if fn == 'gauss_seidel':
# only relax at nonzeros, so as not to mess up any locally dropped
# candidates
indices = np.ravel(x).nonzero()[0]
gauss_seidel_indexed(lvl.A,
x,
np.zeros_like(x),
indices,
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'gauss_seidel_ne':
gauss_seidel_ne(lvl.A,
x,
np.zeros_like(x),
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'gauss_seidel_nr':
gauss_seidel_nr(lvl.A,
x,
np.zeros_like(x),
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'jacobi':
jacobi(lvl.A,
x,
np.zeros_like(x),
iterations=1,
omega=1.0 / rho_D_inv_A(lvl.A))
elif fn == 'richardson':
polynomial(lvl.A,
x,
np.zeros_like(x),
iterations=1,
coeffients=[1.0 / approximate_spectral_radius(lvl.A)])
elif fn == 'gmres':
x[:] = (gmres(lvl.A,
np.zeros_like(x),
x0=x,
maxiter=candidate_iters)[0]).reshape(x.shape)
else:
raise TypeError('Unrecognized smoother')
# x will be dense again, so we have to drop locally again
elim, elim_kwargs = unpack_arg(eliminate_local)
if elim is True:
x = x / norm(x, 'inf')
eliminate_local_candidates(x, levels[0].AggOp, levels[0].A, T0,
**elim_kwargs)
return x.reshape(-1, 1)
def initial_setup_stage(A,
symmetry,
pdef,
candidate_iters,
epsilon,
max_levels,
max_coarse,
aggregate,
prepostsmoother,
smooth,
strength,
work,
initial_candidate=None):
"""
Computes a complete aggregation and the first near-nullspace candidate
following Algorithm 3 in Brezina et al.
Parameters
----------
candidate_iters
number of test relaxation iterations
epsilon
minimum acceptable relaxation convergence factor
References
----------
.. [1] Brezina, Falgout, MacLachlan, Manteuffel, McCormick, and Ruge
"Adaptive Smoothed Aggregation ($\alpha$SA) Multigrid"
SIAM Review Volume 47, Issue 2 (2005)
http://www.cs.umn.edu/~maclach/research/aSA2.pdf
"""
# Define relaxation routine
def relax(A, x):
fn, kwargs = unpack_arg(prepostsmoother)
if fn == 'gauss_seidel':
gauss_seidel(A,
x,
np.zeros_like(x),
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'gauss_seidel_nr':
gauss_seidel_nr(A,
x,
np.zeros_like(x),
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'gauss_seidel_ne':
gauss_seidel_ne(A,
x,
np.zeros_like(x),
iterations=candidate_iters,
sweep='symmetric')
elif fn == 'jacobi':
jacobi(A,
x,
np.zeros_like(x),
iterations=1,
omega=1.0 / rho_D_inv_A(A))
elif fn == 'richardson':
polynomial(A,
x,
np.zeros_like(x),
iterations=1,
coeffients=[1.0 / approximate_spectral_radius(A)])
elif fn == 'gmres':
x[:] = (gmres(A, np.zeros_like(x), x0=x,
maxiter=candidate_iters)[0]).reshape(x.shape)
else:
raise TypeError('Unrecognized smoother')
# flag for skipping steps f-i in step 4
skip_f_to_i = True
# step 1
A_l = A
if initial_candidate is None:
x = sp.rand(A_l.shape[0], 1)
if A_l.dtype == complex:
x = x + 1.0j * sp.rand(A_l.shape[0], 1)
else:
x = np.array(initial_candidate, dtype=A_l.dtype)
# step 2
relax(A_l, x)
work[:] += A_l.nnz * candidate_iters * 2
# step 3
# not advised to stop the iteration here: often the first relaxation pass
# _is_ good, but the remaining passes are poor
# if x_A_x/x_A_x_old < epsilon:
# # relaxation alone is sufficient
# print 'relaxation alone works: %g'%(x_A_x/x_A_x_old)
# return x, []
# step 4
As = [A]
xs = [x]
Ps = []
AggOps = []
StrengthOps = []
while A.shape[0] > max_coarse and max_levels > 1:
# The real check to break from the while loop is below
# Begin constructing next level
fn, kwargs = unpack_arg(strength[len(As) - 1]) # step 4b
if fn == 'symmetric':
C_l = symmetric_strength_of_connection(A_l, **kwargs)
# Diagonal must be nonzero
C_l = C_l + eye(C_l.shape[0], C_l.shape[1], format='csr')
elif fn == 'classical':
C_l = classical_strength_of_connection(A_l, **kwargs)
# Diagonal must be nonzero
C_l = C_l + eye(C_l.shape[0], C_l.shape[1], format='csr')
if isspmatrix_bsr(A_l):
C_l = amalgamate(C_l, A_l.blocksize[0])
elif (fn == 'ode') or (fn == 'evolution'):
C_l = evolution_strength_of_connection(
A_l, np.ones((A_l.shape[0], 1), dtype=A.dtype), **kwargs)
elif fn == 'predefined':
C_l = kwargs['C'].tocsr()
elif fn is None:
C_l = A_l.tocsr()
else:
raise ValueError('unrecognized strength of connection method: %s' %
str(fn))
# In SA, strength represents "distance", so we take magnitude of
# complex values
if C_l.dtype == complex:
C_l.data = np.abs(C_l.data)
# Create a unified strength framework so that large values represent
# strong connections and small values represent weak connections
if (fn == 'ode') or (fn == 'evolution') or (fn == 'energy_based'):
C_l.data = 1.0 / C_l.data
# aggregation
fn, kwargs = unpack_arg(aggregate[len(As) - 1])
if fn == 'standard':
AggOp = standard_aggregation(C_l, **kwargs)[0]
elif fn == 'lloyd':
AggOp = lloyd_aggregation(C_l, **kwargs)[0]
elif fn == 'predefined':
AggOp = kwargs['AggOp'].tocsr()
else:
raise ValueError('unrecognized aggregation method %s' % str(fn))
T_l, x = fit_candidates(AggOp, x) # step 4c
fn, kwargs = unpack_arg(smooth[len(As) - 1]) # step 4d
if fn == 'jacobi':
P_l = jacobi_prolongation_smoother(A_l, T_l, C_l, x, **kwargs)
elif fn == 'richardson':
P_l = richardson_prolongation_smoother(A_l, T_l, **kwargs)
elif fn == 'energy':
P_l = energy_prolongation_smoother(A_l, T_l, C_l, x, None,
(False, {}), **kwargs)
elif fn is None:
P_l = T_l
else:
raise ValueError('unrecognized prolongation smoother method %s' %
str(fn))
# R should reflect A's structure # step 4e
if symmetry == 'symmetric':
A_l = P_l.T.asformat(P_l.format) * A_l * P_l
elif symmetry == 'hermitian':
A_l = P_l.H.asformat(P_l.format) * A_l * P_l
StrengthOps.append(C_l)
AggOps.append(AggOp)
Ps.append(P_l)
As.append(A_l)
# skip to step 5 as in step 4e
if (A_l.shape[0] <= max_coarse) or (len(AggOps) + 1 >= max_levels):
break
if not skip_f_to_i:
x_hat = x.copy() # step 4g
relax(A_l, x) # step 4h
work[:] += A_l.nnz * candidate_iters * 2
if pdef is True:
x_A_x = np.dot(np.conjugate(x).T, A_l * x)
xhat_A_xhat = np.dot(np.conjugate(x_hat).T, A_l * x_hat)
err_ratio = (x_A_x / xhat_A_xhat)**(1.0 / candidate_iters)
else:
# use A.H A inner-product
Ax = A_l * x
# Axhat = A_l * x_hat
x_A_x = np.dot(np.conjugate(Ax).T, Ax)
xhat_A_xhat = np.dot(np.conjugate(x_hat).T, A_l * x_hat)
err_ratio = (x_A_x / xhat_A_xhat)**(1.0 / candidate_iters)
if err_ratio < epsilon: # step 4i
# print "sufficient convergence, skipping"
skip_f_to_i = True
if x_A_x == 0:
x = x_hat # need to restore x
else:
# just carry out relaxation, don't check for convergence
relax(A_l, x) # step 4h
work[:] += 2 * A_l.nnz * candidate_iters
# store xs for diagnostic use and for use in step 5
xs.append(x)
# step 5
# Extend coarse-level candidate to the finest level
# --> note that we start with the x from the second coarsest level
x = xs[-1]
# make sure that xs[-1] has been relaxed by step 4h, i.e. relax(As[-2], x)
for lev in range(len(Ps) - 2, -1, -1): # lev = coarsest ... finest-1
P = Ps[lev] # I: lev --> lev+1
A = As[lev] # A on lev+1
x = P * x
relax(A, x)
work[:] += A.nnz * candidate_iters * 2
# Set predefined strength of connection and aggregation
if len(AggOps) > 1:
aggregate = [('predefined', {
'AggOp': AggOps[i]
}) for i in range(len(AggOps))]
strength = [('predefined', {
'C': StrengthOps[i]
}) for i in range(len(StrengthOps))]
return x, aggregate, strength # first candidate