def index_sum(self, o, ops):
summand, mi = o.ufl_operands
ic = mi[0].count()
fi = summand.ufl_free_indices
fid = summand.ufl_index_dimensions
ipos = fi.index(ic)
d = fid[ipos]
# Compute "macro-dimensions" before and after i in the total shape of a
predim = product(summand.ufl_shape) * product(fid[:ipos])
postdim = product(fid[ipos+1:])
# Map each flattened total component of summand to
# flattened total component of indexsum o by removing
# axis corresponding to summation index ii.
ss = ops[0] # Scalar subexpressions of summand
if len(ss) != predim * postdim * d:
error("Mismatching number of subexpressions.")
sops = []
for i in range(predim):
iind = i * (postdim * d)
for k in range(postdim):
ind = iind + k
sops.append([ss[ind + j * postdim] for j in range(d)])
# For each scalar output component, sum over collected subcomponents
# TODO: Need to split this into binary additions to work with future CRSArray format,
# i.e. emitting more expressions than there are symbols for this node.
results = [sum(sop) for sop in sops]
return results
def handle_index_sum(o, ops):
summand, mi = o.ufl_operands
ic = mi[0].count()
fi = summand.ufl_free_indices
fid = summand.ufl_index_dimensions
ipos = fi.index(ic)
d = fid[ipos]
# Compute "macro-dimensions" before and after i in the total shape of a
predim = ufl.product(summand.ufl_shape) * ufl.product(fid[:ipos])
postdim = ufl.product(fid[ipos + 1:])
# Map each flattened total component of summand to
# flattened total component of indexsum o by removing
# axis corresponding to summation index ii.
ss = ops[0] # Scalar subexpressions of summand
if len(ss) != predim * postdim * d:
raise RuntimeError("Mismatching number of subexpressions.")
sops = []
for i in range(predim):
iind = i * (postdim * d)
for k in range(postdim):
ind = iind + k
sops.append([ss[ind + j * postdim] for j in range(d)])
# For each scalar output component, sum over collected subcomponents
# TODO: Need to split this into binary additions to work with future CRSArray format,
# i.e. emitting more expressions than there are symbols for this node.
results = [sum(sop) for sop in sops]
return results
def generator(ir, parameters):
"""Generate UFC code for a finite element."""
d = {}
d["factory_name"] = ir.classname
d["signature"] = "\"{}\"".format(ir.signature)
d["geometric_dimension"] = ir.geometric_dimension
d["topological_dimension"] = ir.topological_dimension
d["cell_shape"] = ir.cell_shape
d["space_dimension"] = ir.space_dimension
d["value_rank"] = len(ir.value_shape)
d["value_size"] = ufl.product(ir.value_shape)
d["reference_value_rank"] = len(ir.reference_value_shape)
d["reference_value_size"] = ufl.product(ir.reference_value_shape)
d["degree"] = ir.degree
d["family"] = "\"{}\"".format(ir.family)
d["num_sub_elements"] = ir.num_sub_elements
import ffc.codegeneration.C.cnodes as L
d["value_dimension"] = value_dimension(L, ir.value_shape)
d["reference_value_dimension"] = reference_value_dimension(L, ir.reference_value_shape)
statements = evaluate_reference_basis(L, ir, parameters)
d["evaluate_reference_basis"] = L.StatementList(statements)
statements = evaluate_reference_basis_derivatives(L, ir, parameters)
d["evaluate_reference_basis_derivatives"] = L.StatementList(statements)
statements = transform_reference_basis_derivatives(L, ir, parameters)
d["transform_reference_basis_derivatives"] = L.StatementList(statements)
statements = transform_values(L, ir, parameters)
d["transform_values"] = L.StatementList(statements)
statements = tabulate_reference_dof_coordinates(L, ir, parameters)
d["tabulate_reference_dof_coordinates"] = L.StatementList(statements)
statements = create_sub_element(L, ir)
d["sub_element_declaration"] = sub_element_declaration(L, ir)
d["create_sub_element"] = statements
# Check that no keys are redundant or have been missed
from string import Formatter
fieldnames = [
fname for _, fname, _, _ in Formatter().parse(ufc_finite_element.factory) if fname
]
assert set(fieldnames) == set(
d.keys()), "Mismatch between keys in template and in formattting dict"
# Format implementation code
implementation = ufc_finite_element.factory.format_map(d)
# Format declaration
declaration = ufc_finite_element.declaration.format(factory_name=ir.classname)
return declaration, implementation
def _read_from_xdmf_file(fun, directory, filename, suffix, components=None):
if components is not None:
filename = filename + "_component_" + "".join(components)
function_name = "function_" + "".join(components)
else:
function_name = "function"
fun_rank = fun.value_rank()
fun_dim = product(fun.value_shape())
assert fun_rank <= 2
if ((fun_rank is 1 and fun_dim not in (2, 3))
or (fun_rank is 2 and fun_dim not in (4, 9))):
fun_V = fun.function_space()
for i in range(fun_dim):
if components is not None:
filename_i = filename + "_subcomponent_" + str(i)
else:
filename_i = filename + "_component_" + str(i)
fun_i_V = get_function_subspace(fun_V, i)
fun_i = Function(fun_i_V)
if not _read_from_xdmf_file(fun_i, directory, filename_i, suffix,
None):
return False
else:
assign(fun.sub(i), fun_i)
return True
else:
full_filename_checkpoint = os.path.join(str(directory),
filename + "_checkpoint.xdmf")
file_exists = False
if is_io_process() and os.path.exists(full_filename_checkpoint):
file_exists = True
file_exists = is_io_process.mpi_comm.bcast(file_exists,
root=is_io_process.root)
if file_exists:
if suffix is not None:
assert SuffixIO.exists_file(directory, filename + "_suffix")
last_suffix = SuffixIO.load_file(directory,
filename + "_suffix")
if suffix <= last_suffix:
if full_filename_checkpoint in _all_xdmf_files:
assert _all_xdmf_latest_suffix[
full_filename_checkpoint] == suffix - 1
_all_xdmf_latest_suffix[
full_filename_checkpoint] = suffix
else:
assert suffix == 0
_all_xdmf_files[full_filename_checkpoint] = XDMFFile(
full_filename_checkpoint)
_all_xdmf_latest_suffix[full_filename_checkpoint] = 0
_all_xdmf_files[full_filename_checkpoint].read_checkpoint(
fun, function_name, suffix)
return True
else:
return False
else:
with XDMFFile(full_filename_checkpoint) as file_checkpoint:
file_checkpoint.read_checkpoint(fun, function_name, 0)
return True
else:
return False
def __call__(self, *args):
# Assume all args are x argument
x = np.array(args)
dim = self.ufl_domain().geometric_dimension()
if x.shape[-1] != dim:
raise TypeError("expected the geometry argument to be of "
"length %d" % dim)
value_size = ufl.product(self.ufl_element().value_shape())
if cpp.common.has_petsc_complex():
values = np.empty((1, value_size), dtype=np.complex128)
else:
values = np.empty((1, value_size))
# The actual evaluation
self._cpp_object.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1:
return values[0]
print("Returning values: ", values)
return values
def _read_from_file(fun, directory, filename, suffix, components=None):
if components is not None:
filename = filename + "_component_" + "".join(components)
function_name = "function_" + "".join(components)
else:
function_name = "function"
fun_rank = fun.value_rank()
fun_dim = product(fun.value_shape())
assert fun_rank <= 2
if ((fun_rank is 1 and fun_dim not in (2, 3))
or (fun_rank is 2 and fun_dim not in (4, 9))):
funs = fun.split(deepcopy=True)
for (i, fun_i) in enumerate(funs):
if components is not None:
filename_i = filename + "_subcomponent_" + str(i)
else:
filename_i = filename + "_component_" + str(i)
_read_from_file(fun_i, directory, filename_i, suffix, None)
assign(fun.sub(i), fun_i)
else:
if suffix is not None:
if suffix is 0:
# Remove from storage and re-create
try:
del _all_solution_files[(directory, filename)]
except KeyError:
pass
_all_solution_files[(directory, filename)] = SolutionFile(
directory, filename)
file_ = _all_solution_files[(directory, filename)]
file_.read(fun, function_name, suffix)
else:
file_ = SolutionFile(directory, filename)
file_.read(fun, function_name, 0)
def eval(self, x: np.ndarray, cells: np.ndarray, u=None) -> np.ndarray:
"""Evaluate Function at points x, where x has shape (num_points, 3),
and cells has shape (num_points,) and cell[i] is the index of the
cell containing point x[i]. If the cell index is negative the
point is ignored."""
# Make sure input coordinates are a NumPy array
x = np.asarray(x, dtype=np.float64)
assert x.ndim < 3
num_points = x.shape[0] if x.ndim == 2 else 1
x = np.reshape(x, (num_points, -1))
if x.shape[1] != 3:
raise ValueError(
"Coordinate(s) for Function evaluation must have length 3.")
# Make sure cells are a NumPy array
cells = np.asarray(cells, dtype=np.int32)
assert cells.ndim < 2
num_points_c = cells.shape[0] if cells.ndim == 1 else 1
cells = np.reshape(cells, num_points_c)
# Allocate memory for return value if not provided
if u is None:
value_size = ufl.product(self.ufl_element().value_shape())
if common.has_petsc_complex:
u = np.empty((num_points, value_size), dtype=np.complex128)
else:
u = np.empty((num_points, value_size))
self._cpp_object.eval(x, cells, u)
if num_points == 1:
u = np.reshape(u, (-1, ))
return u
def build_scalar_graph(expressions):
"""Build list representation of expression graph covering the given expressions.
TODO: Renaming, refactoring and cleanup of the graph building algorithms used in here
"""
# Build the initial coarse computational graph of the expression
G = build_graph(expressions)
assert len(expressions) == 1, "FIXME: Multiple expressions in graph building needs more work from this point on."
# Build more fine grained computational graph of scalar subexpressions
# TODO: Make it so that
# expressions[k] <-> NV[nvs[k][:]],
# len(nvs[k]) == value_size(expressions[k])
scalar_expressions = rebuild_with_scalar_subexpressions(G)
# Sanity check on number of scalar symbols/components
assert len(scalar_expressions) == sum(product(expr.ufl_shape) for expr in expressions)
# Build new list representation of graph where all
# vertices of V represent single scalar operations
e2i, V, V_targets = build_scalar_graph_vertices(scalar_expressions)
# Compute sparse dependency matrix
V_deps = compute_dependencies(e2i, V)
return V, V_deps, V_targets
def _write_to_pvd_file(fun, directory, filename, suffix, components=None):
if components is not None:
filename = filename + "_component_" + "".join(components)
fun_rank = fun.value_rank()
fun_dim = product(fun.value_shape())
assert fun_rank <= 2
if ((fun_rank is 1 and fun_dim not in (2, 3))
or (fun_rank is 2 and fun_dim not in (4, 9))):
funs = fun.split(deepcopy=True)
for (i, fun_i) in enumerate(funs):
if components is not None:
filename_i = filename + "_subcomponent_" + str(i)
else:
filename_i = filename + "_component_" + str(i)
_write_to_pvd_file(fun_i, directory, filename_i, suffix)
else:
full_filename = os.path.join(str(directory), filename + ".pvd")
if suffix is not None:
if full_filename in _all_pvd_files:
assert _all_pvd_latest_suffix[full_filename] == suffix - 1
_all_pvd_latest_suffix[full_filename] = suffix
else:
assert suffix == 0
_all_pvd_files[full_filename] = File(full_filename,
"compressed")
_all_pvd_latest_suffix[full_filename] = 0
_all_pvd_functions[full_filename] = fun.copy(deepcopy=True)
# Make sure to always use the same function, otherwise dolfin
# changes the numbering and visualization is difficult in ParaView
assign(_all_pvd_functions[full_filename], fun)
_all_pvd_files[full_filename] << _all_pvd_functions[full_filename]
else:
file_ = File(full_filename, "compressed")
file_ << fun
def __init__(self,
ufl_expression: ufl.core.expr.Expr,
x: np.ndarray,
form_compiler_parameters: dict = {}, jit_parameters: dict = {}):
"""Create dolfinx Expression.
Represents a mathematical expression evaluated at a pre-defined set of
points on the reference cell. This class closely follows the concept of a
UFC Expression.
This functionality can be used to evaluate a gradient of a Function at
the quadrature points in all cells. This evaluated gradient can then be
used as input to a non-FEniCS function that calculates a material
constitutive model.
Parameters
----------
ufl_expression
Pure UFL expression
x
Array of points of shape (num_points, tdim) on the reference
element.
form_compiler_parameters
Parameters used in FFCX compilation of this Expression. Run `ffcx
--help` in the commandline to see all available options.
jit_parameters
Parameters controlling JIT compilation of C code.
Note
----
This wrapper is responsible for the FFCX compilation of the UFL Expr
and attaching the correct data to the underlying C++ Expression.
"""
assert x.ndim < 3
num_points = x.shape[0] if x.ndim == 2 else 1
x = np.reshape(x, (num_points, -1))
mesh = ufl_expression.ufl_domain().ufl_cargo()
# Compile UFL expression with JIT
ufc_expression = jit.ffcx_jit(mesh.mpi_comm(), (ufl_expression, x),
form_compiler_parameters=form_compiler_parameters,
jit_parameters=jit_parameters)
self._ufl_expression = ufl_expression
self._ufc_expression = ufc_expression
# Setup data (evaluation points, coefficients, constants, mesh, value_size).
# Tabulation function.
ffi = cffi.FFI()
fn = ffi.cast("uintptr_t", ufc_expression.tabulate_expression)
value_size = ufl.product(self.ufl_expression.ufl_shape)
ufl_coefficients = ufl.algorithms.extract_coefficients(ufl_expression)
coefficients = [ufl_coefficient._cpp_object for ufl_coefficient in ufl_coefficients]
ufl_constants = ufl.algorithms.analysis.extract_constants(ufl_expression)
constants = [ufl_constant._cpp_object for ufl_constant in ufl_constants]
self._cpp_object = cpp.function.Expression(coefficients, constants, mesh, x, fn, value_size)
def build_node_sizes(V_shapes):
"Compute all the products of a sequence of shapes."
nv = len(V_shapes)
V_sizes = numpy.zeros(nv, dtype=int)
for i, sh in enumerate(V_shapes):
V_sizes[i] = product(sh)
return V_sizes
def derivative(self, y):
if len(self.roots) == 0:
deta = Function(y.function_space()).vector()
return deta
p = self.power
factors = []
dfactors = []
dnormsqs = []
normsqs = []
for root in self.roots:
form = self.normsq(y, root)
normsqs.append(assemble(form))
dnormsqs.append(assemble(derivative(form, y)))
for normsq in normsqs:
factor = normsq**(-p / 2.0) + self.shift
dfactor = (-p / 2.0) * normsq**((-p / 2.0) - 1.0)
factors.append(factor)
dfactors.append(dfactor)
eta = product(factors)
deta = Function(y.function_space()).vector()
for (solution, factor, dfactor,
dnormsq) in zip(self.roots, factors, dfactors, dnormsqs):
deta.axpy(float((eta / factor) * dfactor), dnormsq)
return deta
def build_scalar_graph(expressions):
"""Build list representation of expression graph covering the given expressions.
TODO: Renaming, refactoring and cleanup of the graph building algorithms used in here
"""
# Build the initial coarse computational graph of the expression
G = build_graph(expressions)
assert len(
expressions
) == 1, "FIXME: Multiple expressions in graph building needs more work from this point on."
# Build more fine grained computational graph of scalar subexpressions
# TODO: Make it so that
# expressions[k] <-> NV[nvs[k][:]],
# len(nvs[k]) == value_size(expressions[k])
scalar_expressions = rebuild_with_scalar_subexpressions(G)
# Sanity check on number of scalar symbols/components
assert len(scalar_expressions) == sum(
product(expr.ufl_shape) for expr in expressions)
# Build new list representation of graph where all
# vertices of V represent single scalar operations
e2i, V, V_targets = build_scalar_graph_vertices(scalar_expressions)
# Compute sparse dependency matrix
V_deps = compute_dependencies(e2i, V)
return V, V_deps, V_targets
def build_node_sizes(V_shapes):
"Compute all the products of a sequence of shapes."
nv = len(V_shapes)
V_sizes = numpy.zeros(nv, dtype=int)
for i, sh in enumerate(V_shapes):
V_sizes[i] = product(sh)
return V_sizes
def generate_quadrature_tables(self):
"""Generate static tables of quadrature points and weights."""
L = self.backend.language
parts = []
# No quadrature tables for custom (given argument)
# or point (evaluation in single vertex)
skip = ufl.measure.custom_integral_types + ufl.measure.point_integral_types
if self.ir.integral_type in skip:
return parts
alignas = self.ir.params["alignas"]
# Loop over quadrature rules
for num_points in self.ir.all_num_points:
varying_ir = self.ir.varying_irs[num_points]
points, weights = self.ir.quadrature_rules[num_points]
assert num_points == len(weights)
assert num_points == points.shape[0]
# Generate quadrature weights array
if varying_ir["need_weights"]:
wsym = self.backend.symbols.weights_table(num_points)
parts += [
L.ArrayDecl("static const ufc_scalar_t",
wsym,
num_points,
weights,
alignas=alignas)
]
# Generate quadrature points array
N = ufl.product(points.shape)
if varying_ir["need_points"] and N:
# Flatten array: (TODO: avoid flattening here, it makes padding harder)
flattened_points = points.reshape(N)
psym = self.backend.symbols.points_table(num_points)
parts += [
L.ArrayDecl("static const ufc_scalar_t",
psym,
N,
flattened_points,
alignas=alignas)
]
# Add leading comment if there are any tables
parts = L.commented_code_list(parts, "Quadrature rules")
return parts
def ufl_evaluate(self, x, component, derivatives):
"""Function used by ufl to evaluate the Expression"""
assert derivatives == () # TODO: Handle derivatives
if component:
shape = self.ufl_shape
assert len(shape) == len(component)
value_size = product(shape)
index = flatten_multiindex(component, shape_to_strides(shape))
values = numpy.zeros(value_size)
# FIXME: use a function with a return value
self(*x, values=values)
return values[index]
else:
# Scalar evaluation
return self(*x)
def ufl_evaluate(self, x, component, derivatives):
"""Function used by ufl to evaluate the Expression"""
assert derivatives == () # TODO: Handle derivatives
if component:
shape = self.ufl_shape
assert len(shape) == len(component)
value_size = product(shape)
index = flatten_multiindex(component, shape_to_strides(shape))
values = numpy.zeros(value_size)
# FIXME: use a function with a return value
self(*x, values=values)
return values[index]
else:
# Scalar evaluation
return self(*x)
def ufl_evaluate(self, x, component, derivatives):
"""Function used by ufl to evaluate the Expression"""
# FIXME: same as dolfinx.expression.Expression version. Find way
# to re-use.
assert derivatives == () # TODO: Handle derivatives
if component:
shape = self.ufl_shape
assert len(shape) == len(component)
value_size = ufl.product(shape)
index = ufl.utils.indexflattening.flatten_multiindex(
component, ufl.utils.indexflattening.shape_to_strides(shape))
values = np.zeros(value_size)
# FIXME: use a function with a return value
self(*x, values=values)
return values[index]
else:
# Scalar evaluation
return self(*x)
def __call__(self, x: np.ndarray) -> np.ndarray:
"""Evaluate Function at points x, where x has shape (num_points, gdim)"""
_x = np.asarray(x, dtype=np.float)
num_points = _x.shape[0] if len(_x.shape) > 1 else 1
_x = np.reshape(_x, (num_points, -1))
if _x.shape[1] != self.geometric_dimension():
raise ValueError("Wrong geometric dimension for coordinate(s).")
value_size = ufl.product(self.ufl_element().value_shape())
if common.has_petsc_complex:
values = np.empty((num_points, value_size), dtype=np.complex128)
else:
values = np.empty((num_points, value_size))
# Call the evaluation
self._cpp_object.eval(values, _x)
if num_points == 1:
values = np.reshape(values, (-1, ))
return values
def generate(self):
L = self.backend.language
parts = []
parts += self.generate_element_tables()
parts += self.generate_unstructured_piecewise_partition()
all_preparts = []
all_quadparts = []
all_postparts = []
preparts, quadparts, postparts = self.generate_quadrature_loop()
all_preparts += preparts
all_quadparts += quadparts
all_postparts += postparts
preparts, quadparts, postparts = self.generate_dofblock_partition(
quadrature_independent=True)
all_preparts += preparts
all_quadparts += quadparts
all_postparts += postparts
all_finalizeparts = []
# Initialize a tensor to zeros
A_values = [0.0
] * ufl.product(self.ir.expression_shape +
[self.num_points] + self.ir.tensor_shape)
all_finalizeparts = self.generate_tensor_value_initialization(A_values)
# Generate code to add reusable blocks B* to element tensor A
all_finalizeparts += self.generate_copyout_statements()
# Collect parts before, during, and after quadrature loops
parts += all_preparts
parts += all_quadparts
parts += all_postparts
parts += all_finalizeparts
return L.StatementList(parts)
def _write_to_file(fun, directory, filename, suffix, components=None):
if components is not None:
filename = filename + "_component_" + "".join(components)
function_name = "function_" + "".join(components)
else:
function_name = "function"
fun_rank = fun.value_rank()
fun_dim = product(fun.value_shape())
assert fun_rank <= 2
if ((fun_rank is 1 and fun_dim not in (2, 3))
or (fun_rank is 2 and fun_dim not in (4, 9))):
funs = fun.split(deepcopy=True)
for (i, fun_i) in enumerate(funs):
if components is not None:
filename_i = filename + "_subcomponent_" + str(i)
else:
filename_i = filename + "_component_" + str(i)
_write_to_file(fun_i, directory, filename_i, suffix, None)
else:
if suffix is not None:
if suffix is 0:
# Remove existing files if any, as new functions should not be appended, but rather overwrite existing functions
SolutionFile.remove_files(directory, filename)
# Remove from storage and re-create
try:
del _all_solution_files[(directory, filename)]
except KeyError:
pass
_all_solution_files[(directory, filename)] = SolutionFile(
directory, filename)
file_ = _all_solution_files[(directory, filename)]
file_.write(fun, function_name, suffix)
else:
# Remove existing files if any, as new functions should not be appended, but rather overwrite existing functions
SolutionFile.remove_files(directory, filename)
# Write function to file
file_ = SolutionFile(directory, filename)
file_.write(fun, function_name, 0)
def form_argument(self, v):
"""Create new symbols for expressions that represent new values."""
symmetry = v.ufl_element().symmetry()
if symmetry:
# Build symbols with symmetric components skipped
symbols = []
mapped_symbols = {}
for c in ufl.permutation.compute_indices(v.ufl_shape):
# Build mapped component mc with symmetries from element considered
mc = symmetry.get(c, c)
# Get existing symbol or create new and store with mapped component mc as key
s = mapped_symbols.get(mc)
if s is None:
s = self.new_symbol()
mapped_symbols[mc] = s
symbols.append(s)
else:
n = ufl.product(v.ufl_shape + v.ufl_index_dimensions)
symbols = self.new_symbols(n)
return symbols
def map_indexed_arg_components(indexed):
"""Build integer list mapping between flattended components
of indexed expression and its underlying tensor-valued subexpression."""
assert isinstance(indexed, Indexed)
# AKA indexed = tensor[multiindex]
tensor, multiindex = indexed.ufl_operands
# AKA e1 = e2[multiindex]
# (this renaming is historical, but kept for consistency with all the variables *1,*2 below)
e2 = tensor
e1 = indexed
# Get tensor and index shape
sh1 = e1.ufl_shape
sh2 = e2.ufl_shape
fi1 = e1.ufl_free_indices
fi2 = e2.ufl_free_indices
fid1 = e1.ufl_index_dimensions
fid2 = e2.ufl_index_dimensions
# Compute regular and total shape
tsh1 = sh1 + fid1
tsh2 = sh2 + fid2
# r1 = len(tsh1)
r2 = len(tsh2)
# str1 = shape_to_strides(tsh1)
str2 = shape_to_strides(tsh2)
assert not sh1
assert sh2 # Must have shape to be indexed in the first place
assert product(tsh1) <= product(tsh2)
# Build map from fi2/fid2 position (-offset nmui) to fi1/fid1 position
ind2_to_ind1_map = [None] * len(fi2)
for k, i in enumerate(fi2):
ind2_to_ind1_map[k] = fi1.index(i)
# Build map from fi1/fid1 position to mi position
nmui = len(multiindex)
multiindex_to_ind1_map = [None] * nmui
for k, i in enumerate(multiindex):
if isinstance(i, Index):
multiindex_to_ind1_map[k] = fi1.index(i.count())
# Build map from flattened e1 component to flattened e2 component
perm1 = compute_indices(tsh1)
ni1 = product(tsh1)
# Situation: e1 = e2[mi]
d1 = [None] * ni1
p2 = [None] * r2
assert len(sh2) == nmui
for k, i in enumerate(multiindex):
if isinstance(i, FixedIndex):
p2[k] = int(i)
for c1, p1 in enumerate(perm1):
for k, i in enumerate(multiindex):
if isinstance(i, Index):
p2[k] = p1[multiindex_to_ind1_map[k]]
for k, i in enumerate(ind2_to_ind1_map):
p2[nmui + k] = p1[i]
c2 = flatten_multiindex(p2, str2)
d1[c1] = c2
# Consistency checks
assert all(isinstance(x, int) for x in d1)
assert len(set(d1)) == len(d1)
return d1
def reference_value_size(L, reference_value_shape):
return L.Return(ufl.product(reference_value_shape))
def _modified_terminal(self, v):
"""Handle modified terminal.
Modifiers:
---------
terminal - the underlying Terminal object
global_derivatives - tuple of ints, each meaning derivative in that global direction
local_derivatives - tuple of ints, each meaning derivative in that local direction
reference_value - bool, whether this is represented in reference frame
averaged - None, 'facet' or 'cell'
restriction - None, '+' or '-'
component - tuple of ints, the global component of the Terminal
flat_component - single int, flattened local component of the Terminal, considering symmetry
"""
# (1) mt.terminal.ufl_shape defines a core indexing space UNLESS mt.reference_value,
# in which case the reference value shape of the element must be used.
# (2) mt.terminal.ufl_element().symmetry() defines core symmetries
# (3) averaging and restrictions define distinct symbols, no additional symmetries
# (4) two or more grad/reference_grad defines distinct symbols with additional symmetries
# v is not necessary scalar here, indexing in (0,...,0) picks the first scalar component
# to analyse, which should be sufficient to get the base shape and derivatives
if v.ufl_shape:
mt = analyse_modified_terminal(v[(0, ) * len(v.ufl_shape)])
else:
mt = analyse_modified_terminal(v)
# Get derivatives
num_ld = len(mt.local_derivatives)
num_gd = len(mt.global_derivatives)
assert not (num_ld and num_gd)
if num_ld:
domain = mt.terminal.ufl_domain()
tdim = domain.topological_dimension()
d_components = ufl.permutation.compute_indices((tdim, ) * num_ld)
elif num_gd:
domain = mt.terminal.ufl_domain()
gdim = domain.geometric_dimension()
d_components = ufl.permutation.compute_indices((gdim, ) * num_gd)
else:
d_components = [()]
# Get base shape without the derivative axes
base_components = ufl.permutation.compute_indices(mt.base_shape)
# Build symbols with symmetric components and derivatives skipped
symbols = []
mapped_symbols = {}
for bc in base_components:
for dc in d_components:
# Build mapped component mc with symmetries from element
# and derivatives combined
mbc = mt.base_symmetry.get(bc, bc)
mdc = tuple(sorted(dc))
mc = mbc + mdc
# Get existing symbol or create new and store with
# mapped component mc as key
s = mapped_symbols.get(mc)
if s is None:
s = self.new_symbol()
mapped_symbols[mc] = s
symbols.append(s)
# Consistency check before returning symbols
assert not v.ufl_free_indices
if ufl.product(v.ufl_shape) != len(symbols):
raise RuntimeError("Internal error in value numbering.")
return symbols
def map_indexed_arg_components(indexed):
"""Build a map from flattened components to subexpression.
Builds integer list mapping between flattened components
of indexed expression and its underlying tensor-valued subexpression."""
assert isinstance(indexed, Indexed)
# AKA indexed = tensor[multiindex]
tensor, multiindex = indexed.ufl_operands
# AKA e1 = e2[multiindex]
# (this renaming is historical, but kept for consistency with all the variables *1,*2 below)
e2 = tensor
e1 = indexed
# Get tensor and index shape
sh1 = e1.ufl_shape
sh2 = e2.ufl_shape
fi1 = e1.ufl_free_indices
fi2 = e2.ufl_free_indices
fid1 = e1.ufl_index_dimensions
fid2 = e2.ufl_index_dimensions
# Compute regular and total shape
tsh1 = sh1 + fid1
tsh2 = sh2 + fid2
# r1 = len(tsh1)
r2 = len(tsh2)
# str1 = shape_to_strides(tsh1)
str2 = shape_to_strides(tsh2)
assert not sh1
assert sh2 # Must have shape to be indexed in the first place
assert ufl.product(tsh1) <= ufl.product(tsh2)
# Build map from fi2/fid2 position (-offset nmui) to fi1/fid1 position
ind2_to_ind1_map = [None] * len(fi2)
for k, i in enumerate(fi2):
ind2_to_ind1_map[k] = fi1.index(i)
# Build map from fi1/fid1 position to mi position
nmui = len(multiindex)
multiindex_to_ind1_map = [None] * nmui
for k, i in enumerate(multiindex):
if isinstance(i, Index):
multiindex_to_ind1_map[k] = fi1.index(i.count())
# Build map from flattened e1 component to flattened e2 component
perm1 = compute_indices(tsh1)
ni1 = ufl.product(tsh1)
# Situation: e1 = e2[mi]
d1 = [None] * ni1
p2 = [None] * r2
assert len(sh2) == nmui
for k, i in enumerate(multiindex):
if isinstance(i, FixedIndex):
p2[k] = int(i)
for c1, p1 in enumerate(perm1):
for k, i in enumerate(multiindex):
if isinstance(i, Index):
p2[k] = p1[multiindex_to_ind1_map[k]]
for k, i in enumerate(ind2_to_ind1_map):
p2[nmui + k] = p1[i]
c2 = flatten_multiindex(p2, str2)
d1[c1] = c2
# Consistency checks
assert all(isinstance(x, int) for x in d1)
assert len(set(d1)) == len(d1)
return d1
def __call__(self, *args, **kwargs):
# GNW: This function is copied from the old DOLFIN Python
# code. It is far too complicated. There is no need to provide
# so many ways of doing the same thing.
#
# Deprecate as many options as possible, and maybe share with
# dolfin.expression.Expresssion.
if len(args) == 0:
raise TypeError("expected at least 1 argument")
# Test for ufl restriction
if len(args) == 1 and isinstance(args[0], str):
if args[0] in ('+', '-'):
return ufl.Coefficient.__call__(self, *args)
# Test for ufl mapping
if len(args) == 2 and isinstance(args[1], dict) and self in args[1]:
return ufl.Coefficient.__call__(self, *args)
# Some help variables
value_size = ufl.product(self.ufl_element().value_shape())
# If values (return argument) is passed, check the type and length
values = kwargs.get("values", None)
if values is not None:
if not isinstance(values, np.ndarray):
raise TypeError("expected a NumPy array for 'values'")
if len(values) != value_size or \
not np.issubdtype(values.dtype, np.float64):
raise TypeError("expected a double NumPy array of length"
" %d for return values." % value_size)
values_provided = True
else:
values_provided = False
values = np.zeros(value_size, dtype='d')
# Get the geometric dimension we live in
dim = self.ufl_domain().geometric_dimension()
# Assume all args are x argument
x = args
# If only one x argument has been provided, unpack it if it's
# an iterable
if len(x) == 1:
if isinstance(x[0], cpp.geometry.Point):
x = [x[0][i] for i in range(dim)]
elif hasattr(x[0], '__iter__'):
x = x[0]
# Convert it to an 1D numpy array
try:
x = np.fromiter(x, 'd')
except (TypeError, ValueError, AssertionError):
raise TypeError("expected scalar arguments for the coordinates")
if len(x) == 0:
raise TypeError("coordinate argument too short")
if len(x) != dim:
raise TypeError("expected the geometry argument to be of "
"length %d" % dim)
# The actual evaluation
self._cpp_object.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1 and not values_provided:
return values[0]
return values
def reference_value_size(self, L, reference_value_shape):
return L.Return(product(reference_value_shape))
def value_size(self, L, value_shape):
return L.Return(product(value_shape))
def __call__(self, *args, **kwargs):
# GNW: This function is copied from the old DOLFIN Python
# code. It is far too complicated. There is no need to provide
# so many ways of doing the same thing.
#
# Deprecate as many options as possible
if len(args) == 0:
raise TypeError("expected at least 1 argument")
# Test for ufl restriction
if len(args) == 1 and isinstance(args[0], str):
if args[0] in ('+', '-'):
return ufl.Coefficient.__call__(self, *args)
# Test for ufl mapping
if len(args) == 2 and isinstance(args[1], dict) and self in args[1]:
return ufl.Coefficient.__call__(self, *args)
# Some help variables
value_size = product(self.ufl_element().value_shape())
# If values (return argument) is passed, check the type and
# length
values = kwargs.get("values", None)
if values is not None:
if not isinstance(values, numpy.ndarray):
raise TypeError("expected a NumPy array for 'values'")
if len(values) != value_size or not numpy.issubdtype(values.dtype, numpy.float64):
raise TypeError("expected a double NumPy array of length"
" %d for return values." % value_size)
values_provided = True
else:
values_provided = False
values = numpy.zeros(value_size, dtype='d')
# Get dim if element has any domains
cell = self.ufl_element().cell()
dim = None if cell is None else cell.geometric_dimension()
# Assume all args are x argument
x = args
# If only one x argument has been provided, unpack it if it's
# an iterable
if len(x) == 1:
if isinstance(x[0], cpp.geometry.Point):
if dim is not None:
x = [x[0][i] for i in range(dim)]
else:
x = [x[0][i] for i in range(3)]
elif hasattr(x[0], '__iter__'):
x = x[0]
# Convert it to an 1D numpy array
try:
x = numpy.fromiter(x, 'd')
except (TypeError, ValueError, AssertionError):
raise TypeError("expected scalar arguments for the coordinates")
if len(x) == 0:
raise TypeError("coordinate argument too short")
if dim is None:
# Disabled warning as it breaks py.test due to excessive
# output, and that code that is warned about is still
# officially supported. See
# https://bitbucket.org/fenics-project/dolfin/issues/355/
# warning("Evaluating an Expression without knowing the right geometric dimension, assuming %d is correct." % len(x))
pass
else:
if len(x) != dim:
raise TypeError("expected the geometry argument to be of "
"length %d" % dim)
# The actual evaluation
self._cpp_object.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1 and not values_provided:
return values[0]
return values
def __call__(self, *args, **kwargs):
# GNW: This function is copied from the old DOLFIN Python
# code. It is far too complicated. There is no need to provide
# so many ways of doing the same thing.
#
# Deprecate as many options as possible, and maybe share with
# dolfin.expression.Expresssion.
if len(args) == 0:
raise TypeError("expected at least 1 argument")
# Test for ufl restriction
if len(args) == 1 and isinstance(args[0], str):
if args[0] in ('+', '-'):
return ufl.Coefficient.__call__(self, *args)
# Test for ufl mapping
if len(args) == 2 and isinstance(args[1], dict) and self in args[1]:
return ufl.Coefficient.__call__(self, *args)
# Some help variables
value_size = ufl.product(self.ufl_element().value_shape())
# If values (return argument) is passed, check the type and length
values = kwargs.get("values", None)
if values is not None:
if not isinstance(values, np.ndarray):
raise TypeError("expected a NumPy array for 'values'")
if len(values) != value_size or \
not np.issubdtype(values.dtype, np.float64):
raise TypeError("expected a double NumPy array of length"
" %d for return values." % value_size)
values_provided = True
else:
values_provided = False
values = np.zeros(value_size, dtype='d')
# Get the geometric dimension we live in
dim = self.ufl_domain().geometric_dimension()
# Assume all args are x argument
x = args
# If only one x argument has been provided, unpack it if it's
# an iterable
if len(x) == 1:
if isinstance(x[0], cpp.geometry.Point):
x = [x[0][i] for i in range(dim)]
elif hasattr(x[0], '__iter__'):
x = x[0]
# Convert it to an 1D numpy array
try:
x = np.fromiter(x, 'd')
except (TypeError, ValueError, AssertionError):
raise TypeError("expected scalar arguments for the coordinates")
if len(x) == 0:
raise TypeError("coordinate argument too short")
if len(x) != dim:
raise TypeError("expected the geometry argument to be of "
"length %d" % dim)
# The actual evaluation
self._cpp_object.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1 and not values_provided:
return values[0]
return values
def compile_expression1(expr):
code = "double values[%d];" % (product(expr.ufl_shape),)
return code
def generate_preintegrated_dofblock_partition(self):
# FIXME: Generalize this to unrolling all A[] += ... loops,
# or all loops with noncontiguous DM??
L = self.backend.language
block_contributions = self.ir.piecewise_ir["block_contributions"]
blocks = [
(blockmap, blockdata)
for blockmap, contributions in sorted(block_contributions.items())
for blockdata in contributions
if blockdata.block_mode == "preintegrated"
]
# Get symbol, dimensions, and loop index symbols for A
A_shape = self.ir.tensor_shape
A_size = ufl.product(A_shape)
A_rank = len(A_shape)
# TODO: there's something like shape2strides(A_shape) somewhere
# A_strides = ufl.utils.indexflattening.shape_to_strides(A_shape)
A_strides = [1] * A_rank
for i in reversed(range(0, A_rank - 1)):
A_strides[i] = A_strides[i + 1] * A_shape[i + 1]
A_values = [0.0] * A_size
for blockmap, blockdata in blocks:
# Accumulate A[blockmap[...]] += f*PI[...]
# Get table for inlining
tables = self.ir.unique_tables
table = tables[blockdata.name]
inline_table = self.ir.integral_type == "cell"
# Get factor expression
v = self.ir.piecewise_ir["factorization"].nodes[
blockdata.factor_index]['expression']
f = self.get_var(None, v)
# Define rhs expression for A[blockmap[arg_indices]] += A_rhs
# A_rhs = f * PI where PI = sum_q weight * u * v
PI = L.Symbol(blockdata.name)
# Define indices into preintegrated block
P_entity_indices = self.get_entities(blockdata)
if inline_table:
assert P_entity_indices == (L.LiteralInt(0), )
assert table.shape[0] == 1
# Unroll loop
blockshape = [len(DM) for DM in blockmap]
blockrange = [range(d) for d in blockshape]
for ii in itertools.product(*blockrange):
A_ii = sum(A_strides[i] * blockmap[i][ii[i]]
for i in range(len(ii)))
if blockdata.transposed:
P_arg_indices = (ii[1], ii[0])
else:
P_arg_indices = ii
if inline_table:
# Extract float value of PI[P_ii]
Pval = table[0] # always entity 0
for i in P_arg_indices:
Pval = Pval[i]
A_rhs = Pval * f
else:
# Index the static preintegrated table:
P_ii = P_entity_indices + P_arg_indices
A_rhs = f * PI[P_ii]
A_values[A_ii] = A_values[A_ii] + A_rhs
code = self.generate_tensor_value_initialization(A_values)
return L.commented_code_list(code, "UFLACS block mode: preintegrated")
def value_size(self, L, value_shape):
return L.Return(product(value_shape))
def _modified_terminal(self, v, i):
"""Modifiers:
terminal - the underlying Terminal object
global_derivatives - tuple of ints, each meaning derivative in that global direction
local_derivatives - tuple of ints, each meaning derivative in that local direction
reference_value - bool, whether this is represented in reference frame
averaged - None, 'facet' or 'cell'
restriction - None, '+' or '-'
component - tuple of ints, the global component of the Terminal
flat_component - single int, flattened local component of the Terminal, considering symmetry
"""
# (1) mt.terminal.ufl_shape defines a core indexing space UNLESS mt.reference_value,
# in which case the reference value shape of the element must be used.
# (2) mt.terminal.ufl_element().symmetry() defines core symmetries
# (3) averaging and restrictions define distinct symbols, no additional symmetries
# (4) two or more grad/reference_grad defines distinct symbols with additional symmetries
# v is not necessary scalar here, indexing in (0,...,0) picks the first scalar component
# to analyse, which should be sufficient to get the base shape and derivatives
if v.ufl_shape:
mt = analyse_modified_terminal(v[(0,) * len(v.ufl_shape)])
else:
mt = analyse_modified_terminal(v)
# Get derivatives
num_ld = len(mt.local_derivatives)
num_gd = len(mt.global_derivatives)
assert not (num_ld and num_gd)
if num_ld:
domain = mt.terminal.ufl_domain()
tdim = domain.topological_dimension()
d_components = compute_indices((tdim,) * num_ld)
elif num_gd:
domain = mt.terminal.ufl_domain()
gdim = domain.geometric_dimension()
d_components = compute_indices((gdim,) * num_gd)
else:
d_components = [()]
# Get base shape without the derivative axes
base_components = compute_indices(mt.base_shape)
# Build symbols with symmetric components and derivatives skipped
symbols = []
mapped_symbols = {}
for bc in base_components:
for dc in d_components:
# Build mapped component mc with symmetries from element and derivatives combined
mbc = mt.base_symmetry.get(bc, bc)
mdc = tuple(sorted(dc))
mc = mbc + mdc
# Get existing symbol or create new and store with mapped component mc as key
s = mapped_symbols.get(mc)
if s is None:
s = self.new_symbol()
mapped_symbols[mc] = s
symbols.append(s)
# Consistency check before returning symbols
assert not v.ufl_free_indices
if product(v.ufl_shape) != len(symbols):
error("Internal error in value numbering.")
return symbols
def reference_value_size(self, L, reference_value_shape):
return L.Return(product(reference_value_shape))
def map_component_tensor_arg_components(tensor):
"""Build a map from flattened components to subexpression.
Builds integer list mapping between flattended components
of tensor and its underlying indexed subexpression."""
assert isinstance(tensor, ComponentTensor)
# AKA tensor = as_tensor(indexed, multiindex)
indexed, multiindex = tensor.ufl_operands
e1 = indexed
e2 = tensor # e2 = as_tensor(e1, multiindex)
mi = [i for i in multiindex if isinstance(i, Index)]
# Get tensor and index shapes
sh1 = e1.ufl_shape # (sh)ape of e1
sh2 = e2.ufl_shape # (sh)ape of e2
fi1 = e1.ufl_free_indices # (f)ree (i)ndices of e1
fi2 = e2.ufl_free_indices # ...
fid1 = e1.ufl_index_dimensions # (f)ree (i)ndex (d)imensions of e1
fid2 = e2.ufl_index_dimensions # ...
# Compute total shape (tsh) of e1 and e2
tsh1 = sh1 + fid1
tsh2 = sh2 + fid2
r1 = len(tsh1) # 'total rank' or e1
r2 = len(tsh2) # ...
str1 = shape_to_strides(tsh1)
assert not sh1
assert sh2
assert len(mi) == len(multiindex)
assert ufl.product(tsh1) == ufl.product(tsh2)
assert fi1
assert all(i in fi1 for i in fi2)
nmui = len(multiindex)
assert nmui == len(sh2)
# Build map from fi2/fid2 position (-offset nmui) to fi1/fid1 position
p2_to_p1_map = [None] * r2
for k, i in enumerate(fi2):
p2_to_p1_map[k + nmui] = fi1.index(i)
# Build map from fi1/fid1 position to mi position
for k, i in enumerate(mi):
p2_to_p1_map[k] = fi1.index(mi[k].count())
# Build map from flattened e1 component to flattened e2 component
perm2 = compute_indices(tsh2)
ni2 = ufl.product(tsh2)
# Situation: e2 = as_tensor(e1, mi)
d2 = [None] * ni2
p1 = [None] * r1
for c2, p2 in enumerate(perm2):
for k2, k1 in enumerate(p2_to_p1_map):
p1[k1] = p2[k2]
c1 = flatten_multiindex(p1, str1)
d2[c2] = c1
# Consistency checks
assert all(isinstance(x, int) for x in d2)
assert len(set(d2)) == len(d2)
return d2
def __call__(self, *args, **kwargs):
"""
Evaluates the Function.
*Examples*
1) Using an iterable as x:
.. code-block:: python
fs = Expression("sin(x[0])*cos(x[1])*sin(x[3])")
x0 = (1.,0.5,0.5)
x1 = [1.,0.5,0.5]
x2 = numpy.array([1.,0.5,0.5])
v0 = fs(x0)
v1 = fs(x1)
v2 = fs(x2)
2) Using multiple scalar args for x, interpreted as a
point coordinate
.. code-block:: python
v0 = f(1.,0.5,0.5)
3) Using a Point
.. code-block:: python
p0 = Point(1.,0.5,0.5)
v0 = f(p0)
3) Passing return array
.. code-block:: python
fv = Expression(("sin(x[0])*cos(x[1])*sin(x[3])",
"2.0","0.0"))
x0 = numpy.array([1.,0.5,0.5])
v0 = numpy.zeros(3)
fv(x0, values = v0)
.. note::
A longer values array may be passed. In this way one can fast
fill up an array with different evaluations.
.. code-block:: python
values = numpy.zeros(9)
for i in xrange(0,10,3):
fv(x[i:i+3], values = values[i:i+3])
"""
if len(args)==0:
raise TypeError("expected at least 1 argument")
# Test for ufl restriction
if len(args) == 1 and args[0] in ('+','-'):
return ufl.Coefficient.__call__(self, *args)
# Test for ufl mapping
if len(args) == 2 and isinstance(args[1], dict) and self in args[1]:
return ufl.Coefficient.__call__(self, *args)
# Some help variables
value_size = product(self.ufl_element().value_shape())
# If values (return argument) is passed, check the type and length
values = kwargs.get("values", None)
if values is not None:
if not isinstance(values, numpy.ndarray):
raise TypeError("expected a NumPy array for 'values'")
if len(values) != value_size or \
not numpy.issubdtype(values.dtype, 'd'):
raise TypeError("expected a double NumPy array of length"\
" %d for return values."%value_size)
values_provided = True
else:
values_provided = False
values = numpy.zeros(value_size, dtype='d')
# Get the geometric dimension we live in
dim = self.ufl_domain().geometric_dimension()
# Assume all args are x argument
x = args
# If only one x argument has been provided, unpack it if it's
# an iterable
if len(x) == 1:
if isinstance(x[0], cpp.Point):
x = [x[0][i] for i in range(dim)]
elif hasattr(x[0], '__iter__'):
x = x[0]
# Convert it to an 1D numpy array
try:
x = numpy.fromiter(x, 'd')
except (TypeError, ValueError, AssertionError) as e:
raise TypeError("expected scalar arguments for the coordinates")
if len(x) == 0:
raise TypeError("coordinate argument too short")
if len(x) != dim:
raise TypeError("expected the geometry argument to be of "\
"length %d"%dim)
# The actual evaluation
self.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1 and not values_provided:
return values[0]
return values
def __call__(self, *args, **kwargs):
"""
Evaluates the Function.
*Examples*
1) Using an iterable as x:
.. code-block:: python
fs = Expression("sin(x[0])*cos(x[1])*sin(x[3])")
x0 = (1.,0.5,0.5)
x1 = [1.,0.5,0.5]
x2 = numpy.array([1.,0.5,0.5])
v0 = fs(x0)
v1 = fs(x1)
v2 = fs(x2)
2) Using multiple scalar args for x, interpreted as a
point coordinate
.. code-block:: python
v0 = f(1.,0.5,0.5)
3) Using a Point
.. code-block:: python
p0 = Point(1.,0.5,0.5)
v0 = f(p0)
3) Passing return array
.. code-block:: python
fv = Expression(("sin(x[0])*cos(x[1])*sin(x[3])",
"2.0","0.0"))
x0 = numpy.array([1.,0.5,0.5])
v0 = numpy.zeros(3)
fv(x0, values = v0)
.. note::
A longer values array may be passed. In this way one can fast
fill up an array with different evaluations.
.. code-block:: python
values = numpy.zeros(9)
for i in xrange(0,10,3):
fv(x[i:i+3], values = values[i:i+3])
"""
if len(args) == 0:
raise TypeError("expected at least 1 argument")
# Test for ufl restriction
if len(args) == 1 and isinstance(args[0], string_types):
if args[0] in ('+', '-'):
return ufl.Coefficient.__call__(self, *args)
# Test for ufl mapping
if len(args) == 2 and isinstance(args[1], dict) and self in args[1]:
return ufl.Coefficient.__call__(self, *args)
# Some help variables
value_size = product(self.ufl_element().value_shape())
# If values (return argument) is passed, check the type and length
values = kwargs.get("values", None)
if values is not None:
if not isinstance(values, numpy.ndarray):
raise TypeError("expected a NumPy array for 'values'")
if len(values) != value_size or \
not numpy.issubdtype(values.dtype, 'd'):
raise TypeError("expected a double NumPy array of length"\
" %d for return values."%value_size)
values_provided = True
else:
values_provided = False
values = numpy.zeros(value_size, dtype='d')
# Get the geometric dimension we live in
dim = self.ufl_domain().geometric_dimension()
# Assume all args are x argument
x = args
# If only one x argument has been provided, unpack it if it's
# an iterable
if len(x) == 1:
if isinstance(x[0], cpp.Point):
x = [x[0][i] for i in range(dim)]
elif hasattr(x[0], '__iter__'):
x = x[0]
# Convert it to an 1D numpy array
try:
x = numpy.fromiter(x, 'd')
except (TypeError, ValueError, AssertionError) as e:
raise TypeError("expected scalar arguments for the coordinates")
if len(x) == 0:
raise TypeError("coordinate argument too short")
if len(x) != dim:
raise TypeError("expected the geometry argument to be of "\
"length %d"%dim)
# The actual evaluation
self.eval(values, x)
# If scalar return statement, return scalar value.
if value_size == 1 and not values_provided:
return values[0]
return values
def _generate_element_code(ir, parameters):
"Generate code for finite element from intermediate representation."
# Skip code generation if ir is None
if ir is None:
return None
# Prefetch formatting to speedup code generation
ret = format["return"]
do_nothing = format["do nothing"]
create = format["create foo"]
# Codes generated together
(evaluate_dof_code, evaluate_dofs_code) \
= evaluate_dof_and_dofs(ir["evaluate_dof"])
element_number = ir["id"]
# Generate code
code = {}
code["classname"] = ir["classname"]
code["members"] = ""
code["constructor"] = do_nothing
code["constructor_arguments"] = ""
code["initializer_list"] = ""
code["destructor"] = do_nothing
code["signature"] = ret('"%s"' % ir["signature"])
code["cell_shape"] = ret(format["cell"](ir["cell"].cellname()))
code["topological_dimension"] = ret(ir["cell"].topological_dimension())
code["geometric_dimension"] = ret(ir["cell"].geometric_dimension())
code["space_dimension"] = ret(ir["space_dimension"])
code["value_rank"] = ret(len(ir["value_dimension"]))
code["value_dimension"] = _value_dimension(ir["value_dimension"])
code["value_size"] = ret(product(ir["value_dimension"]))
code["reference_value_rank"] = ret(len(ir["reference_value_dimension"]))
code["reference_value_dimension"] = _value_dimension(ir["reference_value_dimension"])
code["reference_value_size"] = ret(product(ir["reference_value_dimension"]))
code["evaluate_basis"] = _evaluate_basis(ir["evaluate_basis"])
code["evaluate_basis_all"] = _evaluate_basis_all(ir["evaluate_basis"])
code["evaluate_basis_derivatives"] \
= _evaluate_basis_derivatives(ir["evaluate_basis"])
code["evaluate_basis_derivatives_all"] \
= _evaluate_basis_derivatives_all(ir["evaluate_basis"])
code["evaluate_dof"] = evaluate_dof_code
code["evaluate_dofs"] = evaluate_dofs_code
code["interpolate_vertex_values"] \
= interpolate_vertex_values(ir["interpolate_vertex_values"])
code["tabulate_dof_coordinates"] \
= _tabulate_dof_coordinates(ir["tabulate_dof_coordinates"])
code["num_sub_elements"] = ret(ir["num_sub_elements"])
code["create_sub_element"] = _create_sub_element(ir)
code["create"] = ret(create(code["classname"]))
# Postprocess code
_postprocess_code(code, parameters)
return code
def expr(self, v):
"""Create new symbols for expressions that represent new values."""
n = ufl.product(v.ufl_shape + v.ufl_index_dimensions)
return self.new_symbols(n)
def map_component_tensor_arg_components(tensor):
"""Build integer list mapping between flattended components
of tensor and its underlying indexed subexpression."""
assert isinstance(tensor, ComponentTensor)
# AKA tensor = as_tensor(indexed, multiindex)
indexed, multiindex = tensor.ufl_operands
e1 = indexed
e2 = tensor # e2 = as_tensor(e1, multiindex)
mi = [i for i in multiindex if isinstance(i, Index)]
# Get tensor and index shapes
sh1 = e1.ufl_shape # (sh)ape of e1
sh2 = e2.ufl_shape # (sh)ape of e2
fi1 = e1.ufl_free_indices # (f)ree (i)ndices of e1
fi2 = e2.ufl_free_indices # ...
fid1 = e1.ufl_index_dimensions # (f)ree (i)ndex (d)imensions of e1
fid2 = e2.ufl_index_dimensions # ...
# Compute total shape (tsh) of e1 and e2
tsh1 = sh1 + fid1
tsh2 = sh2 + fid2
r1 = len(tsh1) # 'total rank' or e1
r2 = len(tsh2) # ...
str1 = shape_to_strides(tsh1)
assert not sh1
assert sh2
assert len(mi) == len(multiindex)
assert product(tsh1) == product(tsh2)
assert fi1
assert all(i in fi1 for i in fi2)
nmui = len(multiindex)
assert nmui == len(sh2)
# Build map from fi2/fid2 position (-offset nmui) to fi1/fid1 position
p2_to_p1_map = [None] * r2
for k, i in enumerate(fi2):
p2_to_p1_map[k + nmui] = fi1.index(i)
# Build map from fi1/fid1 position to mi position
for k, i in enumerate(mi):
p2_to_p1_map[k] = fi1.index(mi[k].count())
# Build map from flattened e1 component to flattened e2 component
perm2 = compute_indices(tsh2)
ni2 = product(tsh2)
# Situation: e2 = as_tensor(e1, mi)
d2 = [None] * ni2
p1 = [None] * r1
for c2, p2 in enumerate(perm2):
for k2, k1 in enumerate(p2_to_p1_map):
p1[k1] = p2[k2]
c1 = flatten_multiindex(p1, str1)
d2[c2] = c1
# Consistency checks
assert all(isinstance(x, int) for x in d2)
assert len(set(d2)) == len(d2)
return d2
def generate_block_parts(self, blockmap, blockdata):
"""Generate and return code parts for a given block."""
L = self.backend.language
# The parts to return
preparts = []
quadparts = []
block_rank = len(blockmap)
blockdims = tuple(len(dofmap) for dofmap in blockmap)
ttypes = blockdata.ttypes
if "zeros" in ttypes:
raise RuntimeError(
"Not expecting zero arguments to be left in dofblock generation."
)
arg_indices = tuple(
self.backend.symbols.argument_loop_index(i)
for i in range(block_rank))
F = self.ir.integrand[self.quadrature_rule]["factorization"]
assert not blockdata.transposed, "Not handled yet"
components = ufl.product(self.ir.expression_shape)
num_points = self.quadrature_rule.points.shape[0]
A_shape = self.ir.tensor_shape
Asym = self.backend.symbols.element_tensor()
A = L.FlattenedArray(Asym, dims=[components] + [num_points] + A_shape)
iq = self.backend.symbols.quadrature_loop_index()
# Prepend dimensions of dofmap block with free index
# for quadrature points and expression components
B_indices = tuple([iq] + list(arg_indices))
# Fetch code to access modified arguments
# An access to FE table data
arg_factors = self.get_arg_factors(blockdata, block_rank, B_indices)
A_indices = []
for i in range(len(blockmap)):
offset = blockmap[i][0]
A_indices.append(arg_indices[i] + offset)
A_indices = tuple([iq] + A_indices)
# Multiply collected factors
# For each component of the factor expression
# add result inside quadloop
body = []
for fi_ci in blockdata.factor_indices_comp_indices:
f = self.get_var(F.nodes[fi_ci[0]]["expression"])
Brhs = L.float_product([f] + arg_factors)
body.append(L.AssignAdd(A[(fi_ci[1], ) + A_indices], Brhs))
for i in reversed(range(block_rank)):
body = L.ForRange(B_indices[i + 1], 0, blockdims[i], body=body)
quadparts += [body]
return preparts, quadparts
def value_size(L, value_shape):
return L.Return(ufl.product(value_shape))
