def unroll(body_gen, total_number, max_unroll=None, start=0):
from cgen import For, Line, Block
from pytools import flatten
if max_unroll is None:
max_unroll = total_number
result = []
if total_number > max_unroll:
loop_items = (total_number // max_unroll) * max_unroll
result.extend([
For("unsigned j = 0",
"j < %d" % loop_items,
"j += %d" % max_unroll,
Block(list(flatten(
body_gen("(j+%d)" % i)
for i in range(max_unroll))))
),
Line()
])
start += loop_items
result.extend(flatten(
body_gen(i) for i in range(start, total_number)))
return result
def unroll(body_gen, total_number, max_unroll=None, start=0):
from cgen import For, Line, Block
from pytools import flatten
if max_unroll is None:
max_unroll = total_number
result = []
if total_number > max_unroll:
loop_items = (total_number // max_unroll) * max_unroll
result.extend([
For("unsigned j = 0",
"j < %d" % loop_items,
"j += %d" % max_unroll,
Block(list(flatten(
body_gen("(j+%d)" % i)
for i in range(max_unroll))))
),
Line()
])
start += loop_items
result.extend(flatten(
body_gen(i) for i in range(start, total_number)))
return result
def __call__(self, *args):
result = numpy.empty(self.shape, self.result_dtype)
from pytools import flatten
self.kernel(result, *tuple(flatten(args)))
return result
def map_num_reference_derivative(self, expr):
discr = self.places.get_discretization(expr.dofdesc.geometry,
expr.dofdesc.discr_stage)
from pytools import flatten
ref_axes = flatten([axis] * mult for axis, mult in expr.ref_axes)
return discr.num_reference_derivative(ref_axes, self.rec(expr.operand))
def __call__(self, *args):
result = numpy.empty(self.shape, self.result_dtype)
from pytools import flatten
self.kernel(result, *tuple(flatten(args)))
return result
def generate_icosahedron(r, order):
# http://en.wikipedia.org/w/index.php?title=Icosahedron&oldid=387737307
phi = (1+5**(1/2))/2
from pytools import flatten
vertices = np.array(sorted(flatten([
(0, pm1*1, pm2*phi),
(pm1*1, pm2*phi, 0),
(pm1*phi, 0, pm2*1)]
for pm1 in [-1, 1]
for pm2 in [-1, 1]))).T.copy()
top_ring = [11, 7, 1, 2, 8]
bottom_ring = [10, 9, 3, 0, 4]
bottom_point = 6
top_point = 5
tris = []
l = len(top_ring)
for i in range(l):
tris.append([top_ring[i], top_ring[(i+1) % l], top_point])
tris.append([bottom_ring[i], bottom_point, bottom_ring[(i+1) % l], ])
tris.append([bottom_ring[i], bottom_ring[(i+1) % l], top_ring[i]])
tris.append([top_ring[i], bottom_ring[(i+1) % l], top_ring[(i+1) % l]])
vertices *= r/la.norm(vertices[:, 0])
vertex_indices = np.array(tris, dtype=np.int32)
grp = make_group_from_vertices(vertices, vertex_indices, order)
from meshmode.mesh import Mesh
return Mesh(vertices, [grp],
element_connectivity=None)
def index_list_backend(self, ilists):
from pytools import single_valued
ilist_length = single_valued(len(il) for il in ilists)
assert ilist_length == self.plan.dofs_per_face
from cgen import Typedef, POD
from pytools import flatten
flat_ilists_uncast = numpy.array(list(flatten(ilists)))
if numpy.max(flat_ilists_uncast) >= 256:
tp = numpy.uint16
else:
tp = numpy.uint8
flat_ilists = numpy.asarray(flat_ilists_uncast, dtype=tp)
assert (flat_ilists == flat_ilists_uncast).all()
return GPUIndexLists(
type=tp,
code=[Typedef(POD(tp, "index_list_entry_t"))],
device_memory=cuda.to_device(flat_ilists),
bytes=flat_ilists.size * flat_ilists.itemsize,
)
def static_extremum_of_pw_aff(pw_aff, constants_only, set_method, what, context):
if context is not None:
context = isl.align_spaces(context, pw_aff.get_domain_space(),
obj_bigger_ok=True).params()
pw_aff = pw_aff.gist(context)
pieces = pw_aff.get_pieces()
if len(pieces) == 1:
(_, result), = pieces
if constants_only and not result.is_cst():
raise ValueError("a numeric %s was not found for PwAff '%s'"
% (what, pw_aff))
return result
from pytools import memoize, flatten
@memoize
def is_bounded(set):
assert set.dim(dim_type.set) == 0
return (set
.move_dims(dim_type.set, 0,
dim_type.param, 0, set.dim(dim_type.param))
.is_bounded())
# put constant bounds with unbounded validity first
order = [
(True, False), # constant, unbounded validity
(False, False), # nonconstant, unbounded validity
(True, True), # constant, bounded validity
(False, True), # nonconstant, bounded validity
]
pieces = flatten([
[(set, aff) for set, aff in pieces
if aff.is_cst() == want_is_constant
and is_bounded(set) == want_is_bounded]
for want_is_constant, want_is_bounded in order])
reference = pw_aff.get_aggregate_domain()
if context is not None:
reference = reference.intersect(context)
# {{{ find bounds that are also global bounds
for set, candidate_aff in pieces:
# gist can be time-consuming, try without first
for use_gist in [False, True]:
if use_gist:
candidate_aff = candidate_aff.gist(set)
if constants_only and not candidate_aff.is_cst():
continue
if reference <= set_method(pw_aff, candidate_aff):
return candidate_aff
# }}}
raise StaticValueFindingError("a static %s was not found for PwAff '%s'"
% (what, pw_aff))
def get_next_step(self, available_names, done_insns):
from pytools import all, argmax2
available_insns = [(insn, insn.priority) for insn in self.instructions
if insn not in done_insns and all(
dep.name in available_names
for dep in insn.get_dependencies())]
if not available_insns:
raise self.NoInstructionAvailable
from pytools import flatten
discardable_vars = set(available_names) - set(
flatten([dep.name for dep in insn.get_dependencies()]
for insn in self.instructions if insn not in done_insns))
# {{{ make sure results do not get discarded
dm = mappers.DependencyMapper(composite_leaves=False)
def remove_result_variable(result_expr):
# The extra dependency mapper run is necessary
# because, for instance, subscripts can make it
# into the result expression, which then does
# not consist of just variables.
for var in dm(result_expr):
assert isinstance(var, Variable)
discardable_vars.discard(var.name)
obj_array_vectorize(remove_result_variable, self.result)
# }}}
return argmax2(available_insns), discardable_vars
def static_extremum_of_pw_aff(pw_aff, constants_only, set_method, what, context):
if context is not None:
context = isl.align_spaces(context, pw_aff.get_domain_space(),
obj_bigger_ok=True).params()
pw_aff = pw_aff.gist(context)
pieces = pw_aff.get_pieces()
if len(pieces) == 1:
(_, result), = pieces
if constants_only and not result.is_cst():
raise ValueError("a numeric %s was not found for PwAff '%s'"
% (what, pw_aff))
return result
from pytools import memoize, flatten
@memoize
def is_bounded(set):
assert set.dim(dim_type.set) == 0
return (set
.move_dims(dim_type.set, 0,
dim_type.param, 0, set.dim(dim_type.param))
.is_bounded())
# put constant bounds with unbounded validity first
order = [
(True, False), # constant, unbounded validity
(False, False), # nonconstant, unbounded validity
(True, True), # constant, bounded validity
(False, True), # nonconstant, bounded validity
]
pieces = flatten([
[(set, aff) for set, aff in pieces
if aff.is_cst() == want_is_constant
and is_bounded(set) == want_is_bounded]
for want_is_constant, want_is_bounded in order])
reference = pw_aff.get_aggregate_domain()
if context is not None:
reference = reference.intersect(context)
# {{{ find bounds that are also global bounds
for set, candidate_aff in pieces:
# gist can be time-consuming, try without first
for use_gist in [False, True]:
if use_gist:
candidate_aff = candidate_aff.gist(set)
if constants_only and not candidate_aff.is_cst():
continue
if reference <= set_method(pw_aff, candidate_aff):
return candidate_aff
# }}}
raise StaticValueFindingError("a static %s was not found for PwAff '%s'"
% (what, pw_aff))
def trace(self):
tensor = np.array([1, 0, 0, 1])
trace_argument = pytools.flatten([[tensor, [i]]
for i in range(self.no_qubits)])
return np.einsum(self.dm,
list(range(self.no_qubits)),
*trace_argument,
optimize=True)
def map_num_reference_derivative(self, expr):
discr = self.bound_expr.get_discretization(expr.dofdesc)
from pytools import flatten
ref_axes = flatten([axis] * mult for axis, mult in expr.ref_axes)
return discr.num_reference_derivative(
self.queue,
ref_axes, self.rec(expr.operand)) \
.with_queue(self.queue)
def map_num_reference_derivative(self, expr):
discr = self.bound_expr.get_discretization(expr.where)
from pytools import flatten
ref_axes = flatten([axis] * mult for axis, mult in expr.ref_axes)
return discr.num_reference_derivative(
self.queue,
ref_axes, self.rec(expr.operand)) \
.with_queue(self.queue)
def get_dependencies(self):
deps = set()
for wdflux in self.expressions:
deps |= set(wdflux.interior_deps)
deps |= set(wdflux.boundary_deps)
dep_mapper = self.dep_mapper_factory()
from pytools import flatten
return set(flatten(dep_mapper(dep) for dep in deps))
def make_linear_comb_kernel_with_result_dtype(
result_dtype, scalar_dtypes, vector_dtypes):
comp_count = len(vector_dtypes)
from pytools import flatten
return ElementwiseKernel([VectorArg(result_dtype, "result")] + list(flatten(
(ScalarArg(scalar_dtypes[i], "a%d_fac" % i),
VectorArg(vector_dtypes[i], "a%d" % i))
for i in range(comp_count))),
"result[i] = " + " + ".join("a%d_fac*a%d[i]" % (i, i)
for i in range(comp_count)))
def get_dependencies(self):
deps = set()
for wdflux in self.expressions:
deps |= set(wdflux.interior_deps)
deps |= set(wdflux.boundary_deps)
dep_mapper = self.dep_mapper_factory()
from pytools import flatten
return set(flatten(dep_mapper(dep) for dep in deps))
def make_linear_comb_kernel_with_result_dtype(result_dtype, scalar_dtypes,
vector_dtypes):
comp_count = len(vector_dtypes)
from pytools import flatten
return ElementwiseKernel([VectorArg(result_dtype, "result")] + list(
flatten((ScalarArg(scalar_dtypes[i], "a%d_fac" % i),
VectorArg(vector_dtypes[i], "a%d" % i))
for i in range(comp_count))), "result[i] = " +
" + ".join("a%d_fac*a%d[i]" % (i, i)
for i in range(comp_count)))
def forward_metric_nth_derivative(xyz_axis, ref_axes, dd=None):
r"""
Pointwise metric derivatives representing repeated derivatives
.. math::
\frac{\partial^n x_{\mathrm{xyz\_axis}} }{\partial r_{\mathrm{ref\_axes}}}
where *ref_axes* is a multi-index description.
:arg ref_axes: a :class:`tuple` of tuples indicating indices of
coordinate axes of the reference element to the number of derivatives
which will be taken. For example, the value ``((0, 2), (1, 1))``
indicates taking the second derivative with respect to the first
axis and the first derivative with respect to the second
axis. Each axis must occur only once and the tuple must be sorted
by the axis index.
May also be a singile integer *i*, which is viewed as equivalent
to ``((i, 1),)``.
"""
if isinstance(ref_axes, int):
ref_axes = ((ref_axes, 1),)
if not isinstance(ref_axes, tuple):
raise ValueError("ref_axes must be a tuple")
if tuple(sorted(ref_axes)) != ref_axes:
raise ValueError("ref_axes must be sorted")
if len(dict(ref_axes)) != len(ref_axes):
raise ValueError("ref_axes must not contain an axis more than once")
if dd is None:
dd = DD_VOLUME
inner_dd = dd.with_qtag(QTAG_NONE)
from pytools import flatten
flat_ref_axes = flatten([rst_axis] * n for rst_axis, n in ref_axes)
from grudge.symbolic.operators import RefDiffOperator
result = NodeCoordinateComponent(xyz_axis, inner_dd)
for rst_axis in flat_ref_axes:
result = RefDiffOperator(rst_axis, inner_dd)(result)
if dd.uses_quadrature():
from grudge.symbolic.operators import project
result = project(inner_dd, dd)(result)
prefix = "dx%d_%s" % (
xyz_axis,
"_".join("%sr%d" % ("d" * n, rst_axis) for rst_axis, n in ref_axes))
return cse(result, prefix, cse_scope.DISCRETIZATION)
def scalar_loop_args(self):
from loopy.kernel.data import ValueArg
if self.args is None:
return []
else:
from pytools import flatten
loop_arg_names = list(flatten(dom.get_var_names(dim_type.param)
for dom in self.domains))
return [arg.name for arg in self.args if isinstance(arg, ValueArg)
if arg.name in loop_arg_names]
def scalar_loop_args(self):
from loopy.kernel.data import ValueArg
if self.args is None:
return []
else:
from pytools import flatten
loop_arg_names = list(flatten(dom.get_var_names(dim_type.param)
for dom in self.domains))
return [arg.name for arg in self.args if isinstance(arg, ValueArg)
if arg.name in loop_arg_names]
def flatten(self, ary):
# Return a flat version of *ary*. The returned value is suitable for
# use with solvers whose API expects a one-dimensional array.
if not self._operator_uses_obj_array:
ary = [ary]
from arraycontext import flatten
result = self.array_context.empty(self.total_dofs, self.dtype)
for res_i, (start, end) in zip(ary, self.starts_and_ends):
result[start:end] = flatten(res_i, self.array_context)
return result
def second_fundamental_form(actx: ArrayContext,
dcoll: DiscretizationCollection,
dd=None) -> np.ndarray:
r"""Computes the second fundamental form:
.. math::
S(x) = \begin{bmatrix}
\partial_{uu} x\cdot n & \partial_{uv} x\cdot n \\
\partial_{uv} x\cdot n & \partial_{vv} x\cdot n
\end{bmatrix}
where :math:`n` is the surface normal, :math:`x(u, v)` defines a parameterized
surface, and :math:`u,v` are coordinates on the parameterized surface.
:arg dd: a :class:`~grudge.dof_desc.DOFDesc`, or a value convertible to one.
:returns: a rank-2 object array describing second fundamental form.
"""
if dd is None:
dd = DD_VOLUME
dim = dcoll.discr_from_dd(dd).dim
normal = rel_mv_normal(actx, dcoll, dd=dd).as_vector(dtype=object)
if dim == 1:
second_ref_axes = [((0, 2), )]
elif dim == 2:
second_ref_axes = [((0, 2), ), ((0, 1), (1, 1)), ((1, 2), )]
else:
raise ValueError("%dD surfaces not supported" % dim)
from pytools import flatten
form2 = np.empty((dim, dim), dtype=object)
for ref_axes in second_ref_axes:
i, j = flatten([rst_axis] * n for rst_axis, n in ref_axes)
ruv = make_obj_array([
forward_metric_nth_derivative(actx,
dcoll,
xyz_axis,
ref_axes,
dd=dd)
for xyz_axis in range(dcoll.ambient_dim)
])
form2[i, j] = form2[j, i] = normal.dot(ruv)
return form2
def get_dependencies(self, source_files):
from codepy.tools import join_continued_lines
result, stdout, stderr = call_capture_output(
[self.cc] + ["-M"] + ["-D%s" % define for define in self.defines] +
["-U%s" % undefine for undefine in self.undefines] +
["-I%s" % idir
for idir in self.include_dirs] + self.cflags + source_files)
if result != 0:
raise CompileError("getting dependencies failed: " + stderr)
lines = join_continued_lines(stdout.split("\n"))
from pytools import flatten
return set(flatten(line.split()[2:] for line in lines))
def generate_icosahedron(
r: float,
order: int,
*,
node_vertex_consistency_tolerance: Optional[Union[float, bool]] = None,
unit_nodes: Optional[np.ndarray] = None):
# https://en.wikipedia.org/w/index.php?title=Icosahedron&oldid=387737307
phi = (1 + 5**(1 / 2)) / 2
from pytools import flatten
vertices = np.array(
sorted(
flatten([(0, pm1 * 1, pm2 * phi), (pm1 * 1, pm2 * phi,
0), (pm1 * phi, 0, pm2 * 1)]
for pm1 in [-1, 1] for pm2 in [-1, 1]))).T.copy()
top_ring = [11, 7, 1, 2, 8]
bottom_ring = [10, 9, 3, 0, 4]
bottom_point = 6
top_point = 5
tris = []
m = len(top_ring)
for i in range(m):
tris.append([top_ring[i], top_ring[(i + 1) % m], top_point])
tris.append([
bottom_ring[i],
bottom_point,
bottom_ring[(i + 1) % m],
])
tris.append([bottom_ring[i], bottom_ring[(i + 1) % m], top_ring[i]])
tris.append(
[top_ring[i], bottom_ring[(i + 1) % m], top_ring[(i + 1) % m]])
vertices *= r / la.norm(vertices[:, 0])
vertex_indices = np.array(tris, dtype=np.int32)
grp = make_group_from_vertices(vertices,
vertex_indices,
order,
unit_nodes=unit_nodes)
from meshmode.mesh import Mesh
return Mesh(
vertices, [grp],
node_vertex_consistency_tolerance=node_vertex_consistency_tolerance,
is_conforming=True)
def get_dependencies(self):
deps = set()
from hedge.tools import setify_field as setify
from hedge.optemplate import OperatorBinding, BoundaryPair
for f in self.expressions:
assert isinstance(f, OperatorBinding)
if isinstance(f.field, BoundaryPair):
deps |= setify(f.field.field) | setify(f.field.bfield)
else:
deps |= setify(f.field)
dep_mapper = self.dep_mapper_factory()
from pytools import flatten
return set(flatten(dep_mapper(dep) for dep in deps))
def get_dependencies(self):
deps = set()
from hedge.tools import setify_field as setify
from hedge.optemplate import OperatorBinding, BoundaryPair
for f in self.expressions:
assert isinstance(f, OperatorBinding)
if isinstance(f.field, BoundaryPair):
deps |= setify(f.field.field) | setify(f.field.bfield)
else:
deps |= setify(f.field)
dep_mapper = self.dep_mapper_factory()
from pytools import flatten
return set(flatten(dep_mapper(dep) for dep in deps))
def uninstall_from_make_output(text):
lines = re.split(r'(\n|\&\&)', text)
lines = [
re.sub(r'(^\s+|\s+$)', '', line)
for line in lines
if re.search(r'\/usr\/bin\/install', line) != None
]
lines = [
line for line in lines
if re.search(r'\-d', line) == None
]
paths = flatten(map(install_paths, lines))
for path in paths:
print(path)
print(os.path.exists(path))
print(os.path.isfile(path))
if os.path.isfile(path):
os.unlink(path)
def to_array(self):
single_tensor = ptm.single_tensor
in_indices = list(reversed(range(self.no_qubits)))
idx = [[i, 2 * self.no_qubits - i, 3 * self.no_qubits - i]
for i in in_indices]
transformation_tensors = list(
zip([single_tensor] * self.no_qubits, idx))
transformation_tensors = pytools.flatten(transformation_tensors)
density_matrix = np.einsum(self.dm,
in_indices,
*transformation_tensors,
optimize=True)
density_matrix = density_matrix.reshape(
(2**self.no_qubits, 2**self.no_qubits))
return density_matrix
def get_dependencies(self, source_files):
from codepy.tools import join_continued_lines
from pytools.prefork import call_capture_output
result, stdout, stderr = call_capture_output(
[self.cc]
+ ["-M"]
+ ["-D%s" % define for define in self.defines]
+ ["-U%s" % undefine for undefine in self.defines]
+ ["-I%s" % idir for idir in self.include_dirs]
+ source_files
)
if result != 0:
raise CompileError("getting dependencies failed: "+stderr)
lines = join_continued_lines(stdout.split("\n"))
from pytools import flatten
return set(flatten(
line.split()[2:] for line in lines))
def get_dependencies(self, source_files):
from codepy.tools import join_continued_lines
result, stdout, stderr = call_capture_output(
[self.cc]
+ ["-M"]
+ [f"-D{define}" for define in self.defines]
+ [f"-U{undefine}" for undefine in self.undefines]
+ [f"-I{idir}" for idir in self.include_dirs]
+ self.cflags
+ source_files
)
if result != 0:
raise CompileError(f"getting dependencies failed: {stderr}")
lines = join_continued_lines(stdout.split("\n"))
lines = [line for line in lines
if not (line.strip() and line.strip()[0] == "#")]
from pytools import flatten
return set(flatten(
line.split()[2:] for line in lines))
def volume_to_face_up_interpolation_matrix(self):
"""Generate a matrix that maps volume nodal values to
a vector of face nodal values on the quadrature grid, with
faces immediately concatenated, i.e::
[face 1 nodal data][face 2 nodal data]...
"""
ldis = self.ldis
face_maps = ldis.face_affine_maps()
from pytools import flatten
face_nodes = list(
flatten([face_map(qnode) for qnode in self.face_nodes]
for face_map in face_maps))
from hedge.polynomial import generic_vandermonde
vdm = generic_vandermonde(face_nodes, list(ldis.basis_functions()))
from hedge.tools.linalg import leftsolve
return leftsolve(self.ldis.vandermonde(), vdm)
def face_affine_maps(self):
"""Return an affine map for each face that maps the (n-1)-dimensional
face unit coordinates to their volume coordintes.
"""
face_vertex_node_index_lists = \
self.geometry.face_vertices(self.vertex_indices())
from pytools import flatten, one
vertex_node_indices = set(flatten(face_vertex_node_index_lists))
def find_missing_node(face_vertex_node_indices):
return unit_nodes[one(vertex_node_indices -
set(face_vertex_node_indices))]
unit_nodes = self.unit_nodes()
sets_of_to_points = [
[unit_nodes[fvni] for fvni in face_vertex_node_indices] +
[find_missing_node(face_vertex_node_indices)]
for face_vertex_node_indices in face_vertex_node_index_lists
]
from_points = sets_of_to_points[0]
# Construct an affine map that promotes face nodes into volume
# by appending -1, this should end up on the first face
dim = self.dimensions
from hedge.tools.affine import AffineMap
from hedge.tools.linalg import unit_vector
to_face_1 = AffineMap(
numpy.vstack([
numpy.eye(dim - 1, dtype=numpy.float64),
numpy.zeros(dim - 1)
]), -unit_vector(dim, dim - 1, dtype=numpy.float64))
def finish_affine_map(amap):
return amap.post_compose(to_face_1)
from hedge.tools.affine import identify_affine_map
return [
finish_affine_map(identify_affine_map(from_points, to_points))
for to_points in sets_of_to_points
]
def face_affine_maps(self):
"""Return an affine map for each face that maps the (n-1)-dimensional
face unit coordinates to their volume coordintes.
"""
face_vertex_node_index_lists = \
self.geometry.face_vertices(self.vertex_indices())
from pytools import flatten, one
vertex_node_indices = set(flatten(face_vertex_node_index_lists))
def find_missing_node(face_vertex_node_indices):
return unit_nodes[one(
vertex_node_indices - set(face_vertex_node_indices))]
unit_nodes = self.unit_nodes()
sets_of_to_points = [[unit_nodes[fvni]
for fvni in face_vertex_node_indices]
+ [find_missing_node(face_vertex_node_indices)]
for face_vertex_node_indices in face_vertex_node_index_lists]
from_points = sets_of_to_points[0]
# Construct an affine map that promotes face nodes into volume
# by appending -1, this should end up on the first face
dim = self.dimensions
from hedge.tools.affine import AffineMap
from hedge.tools.linalg import unit_vector
to_face_1 = AffineMap(
numpy.vstack([
numpy.eye(dim-1, dtype=numpy.float64),
numpy.zeros(dim-1)]),
-unit_vector(dim, dim-1, dtype=numpy.float64))
def finish_affine_map(amap):
return amap.post_compose(to_face_1)
from hedge.tools.affine import identify_affine_map
return [
finish_affine_map(
identify_affine_map(from_points, to_points))
for to_points in sets_of_to_points]
def volume_to_face_up_interpolation_matrix(self):
"""Generate a matrix that maps volume nodal values to
a vector of face nodal values on the quadrature grid, with
faces immediately concatenated, i.e::
[face 1 nodal data][face 2 nodal data]...
"""
ldis = self.ldis
face_maps = ldis.face_affine_maps()
from pytools import flatten
face_nodes = list(flatten([face_map(qnode) for qnode in self.face_nodes] for face_map in face_maps))
from hedge.polynomial import generic_vandermonde
vdm = generic_vandermonde(face_nodes, list(ldis.basis_functions()))
from hedge.tools.linalg import leftsolve
return leftsolve(self.ldis.vandermonde(), vdm)
def get_next_step(self, available_names, done_insns):
from pytools import all, argmax2
available_insns = [
(insn, insn.priority) for insn in self.instructions
if insn not in done_insns
and all(dep.name in available_names
for dep in insn.get_dependencies())]
if not available_insns:
raise self.NoInstructionAvailable
from pytools import flatten
discardable_vars = set(available_names) - set(flatten(
[dep.name for dep in insn.get_dependencies()]
for insn in self.instructions
if insn not in done_insns))
# {{{ make sure results do not get discarded
from hedge.tools import with_object_array_or_scalar
from hedge.optemplate.mappers import DependencyMapper
dm = DependencyMapper(composite_leaves=False)
def remove_result_variable(result_expr):
# The extra dependency mapper run is necessary
# because, for instance, subscripts can make it
# into the result expression, which then does
# not consist of just variables.
for var in dm(result_expr):
from pymbolic.primitives import Variable
assert isinstance(var, Variable)
discardable_vars.discard(var.name)
with_object_array_or_scalar(remove_result_variable, self.result)
# }}}
return argmax2(available_insns), discardable_vars
def second_fundamental_form(ambient_dim, dim=None, dd=None):
if dim is None:
dim = ambient_dim - 1
normal = surface_normal(ambient_dim, dim=dim, dd=dd).as_vector()
if dim == 1:
second_ref_axes = [((0, 2),)]
elif dim == 2:
second_ref_axes = [((0, 2),), ((0, 1), (1, 1)), ((1, 2),)]
else:
raise ValueError("%dD surfaces not supported" % dim)
from pytools import flatten
form2 = np.empty((dim, dim), dtype=object)
for ref_axes in second_ref_axes:
i, j = flatten([rst_axis] * n for rst_axis, n in ref_axes)
ruv = np.array([
forward_metric_nth_derivative(xyz_axis, ref_axes, dd=dd)
for xyz_axis in range(ambient_dim)])
form2[i, j] = form2[j, i] = normal.dot(ruv)
return cse(form2, "form2_mat", cse_scope.DISCRETIZATION)
def __init__(self, no_qubits, data=None):
if no_qubits > 15:
raise ValueError(
"no_qubits=%d is way too many qubits, are you sure?" %
no_qubits)
self.no_qubits = no_qubits
self.shape = [4] * no_qubits
if isinstance(data, np.ndarray):
single_tensor = ptm.single_tensor
assert data.size == 4**self.no_qubits
data = data.reshape((2, 2) * self.no_qubits)
in_indices = list(
reversed(range(self.no_qubits, 3 * self.no_qubits)))
contraction_indices = [(i, i + self.no_qubits,
i + 2 * self.no_qubits)
for i in range(self.no_qubits)]
out_indices = list(reversed(range(self.no_qubits)))
transformation_tensors = list(
zip([single_tensor] * self.no_qubits, contraction_indices))
transformation_tensors = pytools.flatten(transformation_tensors)
self.dm = np.einsum(data,
in_indices,
*transformation_tensors,
out_indices,
optimize=True).real
elif data is None:
self.dm = np.zeros(self.shape)
self.dm[tuple([0] * self.no_qubits)] = 1
else:
raise ValueError("type of data not understood")
def generate_icosahedron(r, order):
# http://en.wikipedia.org/w/index.php?title=Icosahedron&oldid=387737307
phi = (1+5**(1/2))/2
from pytools import flatten
vertices = np.array(sorted(flatten([
(0, pm1*1, pm2*phi),
(pm1*1, pm2*phi, 0),
(pm1*phi, 0, pm2*1)]
for pm1 in [-1, 1]
for pm2 in [-1, 1]))).T.copy()
top_ring = [11, 7, 1, 2, 8]
bottom_ring = [10, 9, 3, 0, 4]
bottom_point = 6
top_point = 5
tris = []
l = len(top_ring)
for i in range(l):
tris.append([top_ring[i], top_ring[(i+1) % l], top_point])
tris.append([bottom_ring[i], bottom_point, bottom_ring[(i+1) % l], ])
tris.append([bottom_ring[i], bottom_ring[(i+1) % l], top_ring[i]])
tris.append([top_ring[i], bottom_ring[(i+1) % l], top_ring[(i+1) % l]])
vertices *= r/la.norm(vertices[:, 0])
vertex_indices = np.array(tris, dtype=np.int32)
grp = make_group_from_vertices(vertices, vertex_indices, order)
from meshmode.mesh import Mesh
return Mesh(
vertices, [grp],
nodal_adjacency=None,
facial_adjacency_groups=None)
def _compute_nodal_adjacency_from_vertices(mesh):
# FIXME Native code would make this faster
_, nvertices = mesh.vertices.shape
vertex_to_element = [[] for i in range(nvertices)]
for grp in mesh.groups:
iel_base = grp.element_nr_base
for iel_grp in range(grp.nelements):
for ivertex in grp.vertex_indices[iel_grp]:
vertex_to_element[ivertex].append(iel_base + iel_grp)
element_to_element = [set() for i in range(mesh.nelements)]
for grp in mesh.groups:
iel_base = grp.element_nr_base
for iel_grp in range(grp.nelements):
for ivertex in grp.vertex_indices[iel_grp]:
element_to_element[iel_base + iel_grp].update(
vertex_to_element[ivertex])
for iel, neighbors in enumerate(element_to_element):
neighbors.remove(iel)
lengths = [len(el_list) for el_list in element_to_element]
neighbors_starts = np.cumsum(
np.array([0] + lengths, dtype=mesh.element_id_dtype))
from pytools import flatten
neighbors = np.array(
list(flatten(element_to_element)),
dtype=mesh.element_id_dtype)
assert neighbors_starts[-1] == len(neighbors)
return NodalAdjacency(
neighbors_starts=neighbors_starts,
neighbors=neighbors)
def index_list_backend(self, ilists):
from pytools import single_valued
ilist_length = single_valued(len(il) for il in ilists)
assert ilist_length == self.plan.dofs_per_face
from cgen import Typedef, POD
from pytools import flatten
flat_ilists_uncast = numpy.array(list(flatten(ilists)))
if numpy.max(flat_ilists_uncast) >= 256:
tp = numpy.uint16
else:
tp = numpy.uint8
flat_ilists = numpy.asarray(flat_ilists_uncast, dtype=tp)
assert (flat_ilists == flat_ilists_uncast).all()
return GPUIndexLists(
type=tp,
code=[Typedef(POD(tp, "index_list_entry_t"))],
device_memory=cuda.to_device(flat_ilists),
bytes=flat_ilists.size * flat_ilists.itemsize,
)
def _compute_nodal_adjacency_from_vertices(mesh):
# FIXME Native code would make this faster
_, nvertices = mesh.vertices.shape
vertex_to_element = [[] for i in range(nvertices)]
for grp in mesh.groups:
iel_base = grp.element_nr_base
for iel_grp in range(grp.nelements):
for ivertex in grp.vertex_indices[iel_grp]:
vertex_to_element[ivertex].append(iel_base + iel_grp)
element_to_element = [set() for i in range(mesh.nelements)]
for grp in mesh.groups:
iel_base = grp.element_nr_base
for iel_grp in range(grp.nelements):
for ivertex in grp.vertex_indices[iel_grp]:
element_to_element[iel_base + iel_grp].update(
vertex_to_element[ivertex])
for iel, neighbors in enumerate(element_to_element):
neighbors.remove(iel)
lengths = [len(el_list) for el_list in element_to_element]
neighbors_starts = np.cumsum(
np.array([0] + lengths, dtype=mesh.element_id_dtype))
from pytools import flatten
neighbors = np.array(
list(flatten(element_to_element)),
dtype=mesh.element_id_dtype)
assert neighbors_starts[-1] == len(neighbors)
return NodalAdjacency(
neighbors_starts=neighbors_starts,
neighbors=neighbors)
def generate_nodal_adjacency(self, nelements, nvertices, groups):
# medium-term FIXME: make this an incremental update
# rather than build-from-scratch
vertex_to_element = [[] for i in range(nvertices)]
element_index = 0
for grp in groups:
for iel_grp in range(len(grp)):
for ivertex in grp[iel_grp]:
vertex_to_element[ivertex].append(element_index)
element_index += 1
element_to_element = [set() for i in range(nelements)]
element_index = 0
if self.lazy:
for grp in groups:
for iel_grp in range(len(grp)):
for i in range(len(grp[iel_grp])):
for j in range(i+1, len(grp[iel_grp])):
vertex_pair = (min(grp[iel_grp][i], grp[iel_grp][j]), max(grp[iel_grp][i], grp[iel_grp][j]))
#print 'iel:', iel_grp, 'pair:', vertex_pair
if vertex_pair not in self.seen_tuple:
self.propagate_tree(self.get_root(self.pair_map[vertex_pair]), self.hanging_vertex_element, element_to_element)
#print self.pair_map[vertex_pair].left_vertex, self.pair_map[vertex_pair].right_vertex, self.pair_map[vertex_pair].adjacent_elements, self.hanging_vertex_element[self.pair_map[vertex_pair].left_vertex], self.hanging_vertex_element[self.pair_map[vertex_pair].right_vertex]
else:
for grp in groups:
for iel_grp in range(len(grp)):
for ivertex in grp[iel_grp]:
element_to_element[element_index].update(
vertex_to_element[ivertex])
if self.hanging_vertex_element[ivertex]:
for hanging_element in self.hanging_vertex_element[ivertex]:
if element_index != hanging_element:
element_to_element[element_index].update([hanging_element])
element_to_element[hanging_element].update([element_index])
for i in range(len(grp[iel_grp])):
for j in range(i+1, len(grp[iel_grp])):
vertex_pair = (min(grp[iel_grp][i], grp[iel_grp][j]),
max(grp[iel_grp][i], grp[iel_grp][j]))
#element_to_element[element_index].update(
#self.pair_map[vertex_pair].adjacent_elements)
queue = [self.pair_map[vertex_pair]]
while queue:
vertex = queue.pop(0)
#if leaf node
if vertex.left is None and vertex.right is None:
assert(element_index in vertex.adjacent_elements)
element_to_element[element_index].update(
vertex.adjacent_elements)
else:
queue.append(vertex.left)
queue.append(vertex.right)
'''
if self.hanging_vertex_element[ivertex] and element_index != self.hanging_vertex_element[ivertex][0]:
element_to_element[element_index].update([self.hanging_vertex_element[ivertex][0]])
element_to_element[self.hanging_vertex_element[ivertex][0]].update([element_index])
'''
element_index += 1
logger.debug("number of new elements: %d" % len(element_to_element))
for iel, neighbors in enumerate(element_to_element):
if iel in neighbors:
neighbors.remove(iel)
#print(self.ray_elements)
'''
for ray in self.rays:
curnode = ray.first
while curnode is not None:
if len(curnode.value.elements) >= 2:
if curnode.value.elements[0] is not None:
element_to_element[curnode.value.elements[0]].update(curnode.value.elements)
if curnode.value.elements[1] is not None:
element_to_element[curnode.value.elements[1]].update(curnode.value.elements)
if len(curnode.value.velements) >= 2:
if curnode.value.velements[0] is not None:
element_to_element[curnode.value.velements[0]].update(curnode.value.velements)
if curnode.value.velements[1] is not None:
element_to_element[curnode.value.velements[1]].update(curnode.value.velements)
curnode = curnode.next
'''
'''
for i in self.ray_elements:
for j in i:
#print j[0], j[1]
element_to_element[j[0]].update(j)
element_to_element[j[1]].update(j)
'''
#print element_to_element
lengths = [len(el_list) for el_list in element_to_element]
neighbors_starts = np.cumsum(
np.array([0] + lengths, dtype=self.last_mesh.element_id_dtype),
# cumsum silently widens integer types
dtype=self.last_mesh.element_id_dtype)
from pytools import flatten
neighbors = np.array(
list(flatten(element_to_element)),
dtype=self.last_mesh.element_id_dtype)
assert neighbors_starts[-1] == len(neighbors)
from meshmode.mesh import NodalAdjacency
return NodalAdjacency(neighbors_starts=neighbors_starts, neighbors=neighbors)
def combine(self, values):
from pytools import flatten
return set(flatten(values))
def get_boundary_flux_mod(fluxes, fvi, discr, dtype):
from cgen import \
FunctionDeclaration, FunctionBody, Typedef, Struct, \
Const, Reference, Value, POD, MaybeUnused, \
Statement, Include, Line, Block, Initializer, Assign, \
CustomLoop, For
from pytools import to_uncomplex_dtype, flatten
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
mod.add_to_preamble([
Include("cstdlib"),
Include("algorithm"),
Line(),
Include("boost/foreach.hpp"),
Line(),
Include("hedge/face_operators.hpp"),
])
S = Statement
mod.add_to_module([
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
arg_struct = Struct("arg_struct", [
Value("numpy_array<value_type>", "flux%d_on_faces" % i)
for i in range(len(fluxes))
]+[
Value("numpy_array<value_type>", arg_name)
for arg_name in fvi.arg_names
])
mod.add_struct(arg_struct, "ArgStruct")
mod.add_to_module([Line()])
fdecl = FunctionDeclaration(
Value("void", "gather_flux"),
[
Const(Reference(Value("face_group<face_pair<straight_face> >" , "fg"))),
Reference(Value("arg_struct", "args"))
])
from pymbolic.mapper.stringifier import PREC_PRODUCT
def gen_flux_code():
f2cm = FluxToCodeMapper()
result = [
Assign("fof%d_it[loc_fof_base+i]" % flux_idx,
"uncomplex_type(fp.int_side.face_jacobian) * " +
flux_to_code(f2cm, False, flux_idx, fvi, flux.op.flux, PREC_PRODUCT))
for flux_idx, flux in enumerate(fluxes)
]
return [
Initializer(Value("value_type", cse_name), cse_str)
for cse_name, cse_str in f2cm.cse_name_list] + result
fbody = Block([
Initializer(
Const(Value("numpy_array<value_type>::iterator", "fof%d_it" % i)),
"args.flux%d_on_faces.begin()" % i)
for i in range(len(fluxes))
]+[
Initializer(
Const(Value("numpy_array<value_type>::const_iterator",
"%s_it" % arg_name)),
"args.%s.begin()" % arg_name)
for arg_name in fvi.arg_names
]+[
Line(),
CustomLoop("BOOST_FOREACH(const face_pair<straight_face> &fp, fg.face_pairs)", Block(
list(flatten([
Initializer(Value("node_number_t", "%s_ebi" % where),
"fp.%s.el_base_index" % where),
Initializer(Value("index_lists_t::const_iterator", "%s_idx_list" % where),
"fg.index_list(fp.%s.face_index_list_number)" % where),
Line(),
]
for where in ["int_side", "ext_side"]
))+[
Line(),
Initializer(Value("node_number_t", "loc_fof_base"),
"fg.face_length()*(fp.%(where)s.local_el_number*fg.face_count"
" + fp.%(where)s.face_id)" % {"where": "int_side"}),
Line(),
For(
"unsigned i = 0",
"i < fg.face_length()",
"++i",
Block(
[
Initializer(MaybeUnused(
Value("node_number_t", "%s_idx" % where)),
"%(where)s_ebi + %(where)s_idx_list[i]"
% {"where": where})
for where in ["int_side", "ext_side"]
]+gen_flux_code()
)
)
]))
])
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print mod.generate()
#raw_input("[Enter]")
return mod.compile(get_flux_toolchain(discr, fluxes))
def _setup_neighbor_connections(self):
comm = self.context.communicator
# Why is this barrier needed? Some of our ranks may arrive at this
# point early and start sending packets to ranks that are still stuck
# in previous wildcard-recv loops. These receivers will then be very
# confused by packets they didn't expect, and, once they reach their
# recv bit in *this* subroutine, will wait for packets that will never
# arrive. This same argument does not apply to other recv()s in this
# file because they are targeted and thus benefit from MPI's
# non-overtaking rule.
#
# Parallel programming is fun.
comm.Barrier()
if self.neighbor_ranks:
# send interface information to neighboring ranks -----------------
from pytools import reverse_dictionary
local2global_vertex_indices = \
reverse_dictionary(self.global2local_vertex_indices)
send_requests = []
for rank in self.neighbor_ranks:
bdry_tag = hedge.mesh.TAG_RANK_BOUNDARY(rank)
rank_bdry = self.subdiscr.mesh.tag_to_boundary[bdry_tag]
rank_discr_boundary = self.subdiscr.get_boundary(bdry_tag)
# a list of global vertex numbers for each face
my_vertices_global = [
tuple(local2global_vertex_indices[vi]
for vi in el.faces[face_nr])
for el, face_nr in rank_bdry]
# a list of node coordinates, indicating the order
# in which nodal values will be sent, this is for
# testing only and could (potentially) be omitted
my_node_coords = []
for el, face_nr in rank_bdry:
eslice, ldis = self.subdiscr.find_el_data(el.id)
findices = ldis.face_indices()[face_nr]
my_node_coords.append(
[self.nodes[eslice.start+i] for i in findices])
# compile a list of FluxFace.h values for unification
# across the rank boundary
my_h_values = [rank_discr_boundary.find_facepair_side(el_face).h
for el_face in rank_bdry]
packet = (my_vertices_global, my_node_coords, my_h_values)
send_requests.append(comm.isend(packet, dest=rank, tag=0))
received_packets = {}
while len(received_packets) < len(self.neighbor_ranks):
status = mpi.Status()
received_packet = comm.recv(tag=0, source=mpi.ANY_SOURCE, status=status)
received_packets[status.source] = received_packet
mpi.Request.Waitall(send_requests)
# process received packets ----------------------------------------
from pytools import flatten
# nb_ stands for neighbor_
self.from_neighbor_maps = {}
for rank, (nb_all_facevertices_global, nb_node_coords, nb_h_values) in \
received_packets.iteritems():
bdry_tag = hedge.mesh.TAG_RANK_BOUNDARY(rank)
rank_bdry = self.subdiscr.mesh.tag_to_boundary[bdry_tag]
rank_discr_boundary = self.subdiscr.get_boundary(bdry_tag)
flat_nb_node_coords = list(flatten(nb_node_coords))
# step 1: find start node indices for each
# of the neighbor's elements
nb_face_starts = [0]
for node_coords in nb_node_coords[:-1]:
nb_face_starts.append(
nb_face_starts[-1]+len(node_coords))
# step 2: match faces by matching vertices
nb_face_order = dict(
(frozenset(vertices), i)
for i, vertices in enumerate(nb_all_facevertices_global))
# step 3: make a list of indices into the data we
# receive from our neighbor that'll tell us how
# to reshuffle them to match our node order
from_indices = []
shuffled_indices_cache = {}
def get_shuffled_indices(face_node_count, shuffle_op):
try:
return shuffled_indices_cache[shuffle_op]
except KeyError:
unshuffled_indices = range(face_node_count)
result = shuffled_indices_cache[shuffle_op] = \
shuffle_op(unshuffled_indices)
return result
for el, face_nr in rank_bdry:
eslice, ldis = self.subdiscr.find_el_data(el.id)
my_vertices = el.faces[face_nr]
my_global_vertices = tuple(local2global_vertex_indices[vi]
for vi in my_vertices)
face_node_count = ldis.face_node_count()
try:
nb_face_idx = nb_face_order[frozenset(my_global_vertices)]
# continue below in else part
except KeyError:
# this happens if my_global_vertices is not a permutation
# of the neighbor's face vertices. Periodicity is the only
# reason why that would be so.
my_vertices_there, axis = self.global_periodic_opposite_faces[
my_global_vertices]
nb_face_idx = nb_face_order[frozenset(my_vertices_there)]
his_vertices_here, axis2 = self.global_periodic_opposite_faces[
nb_all_facevertices_global[nb_face_idx]]
assert axis == axis2
nb_face_start = nb_face_starts[nb_face_idx]
shuffle_op = \
ldis.get_face_index_shuffle_to_match(
my_global_vertices,
his_vertices_here)
shuffled_other_node_indices = [nb_face_start+i
for i in get_shuffled_indices(
face_node_count, shuffle_op)]
from_indices.extend(shuffled_other_node_indices)
# check if the nodes really match up
if "parallel_setup" in self.debug:
my_node_indices = [eslice.start+i for i in ldis.face_indices()[face_nr]]
for my_i, other_i in zip(my_node_indices, shuffled_other_node_indices):
dist = self.nodes[my_i]-flat_nb_node_coords[other_i]
dist[axis] = 0
assert la.norm(dist) < 1e-14
else:
# continue handling of nonperiodic case
nb_global_vertices = nb_all_facevertices_global[nb_face_idx]
nb_face_start = nb_face_starts[nb_face_idx]
shuffle_op = \
ldis.get_face_index_shuffle_to_match(
my_global_vertices,
nb_global_vertices)
shuffled_other_node_indices = [nb_face_start+i
for i in get_shuffled_indices(
face_node_count, shuffle_op)]
from_indices.extend(shuffled_other_node_indices)
# check if the nodes really match up
if "parallel_setup" in self.debug:
my_node_indices = [eslice.start+i
for i in ldis.face_indices()[face_nr]]
for my_i, other_i in zip(my_node_indices, shuffled_other_node_indices):
dist = self.nodes[my_i]-flat_nb_node_coords[other_i]
assert la.norm(dist) < 1e-14
# finally, unify FluxFace.h values across boundary
nb_h = nb_h_values[nb_face_idx]
flux_face = rank_discr_boundary.find_facepair_side((el, face_nr))
flux_face.h = max(nb_h, flux_face.h)
if "parallel_setup" in self.debug:
assert len(from_indices) == len(flat_nb_node_coords)
# construct from_neighbor_map
self.from_neighbor_maps[rank] = \
self.subdiscr.prepare_from_neighbor_map(from_indices)
def combine(self, values):
from pytools import flatten
return set(flatten(values))
def make_method_matrix(stepper, rhss, f_size, s_size):
from pymbolic import var
from pytools.obj_array import make_obj_array
f = make_obj_array([var("f%d" % i) for i in range(f_size)])
s = make_obj_array([var("s%d" % i) for i in range(s_size)])
from hedge.timestep.multirate_ab.methods import (HIST_F2F, HIST_S2F,
HIST_F2S, HIST_S2S,
HIST_NAMES)
hist_sizes = {
HIST_F2F: f_size,
HIST_S2F: f_size,
HIST_S2S: s_size,
HIST_F2S: s_size
}
orig_histories = {}
for hn in HIST_NAMES:
my_size = hist_sizes[hn]
my_length = stepper.orders[hn]
hist_name_str = hn.__name__[5:].lower()
hist = [
make_obj_array([
var("h_%s_%d_%d" % (hist_name_str, age, i))
for i in range(s_size)
]) for age in range(my_length)
]
stepper.histories[hn] = hist
orig_histories[hn] = hist[:]
stepper.startup_stepper = None
del stepper.startup_history
f_post_step, s_post_step = stepper([f, s], 0, rhss)
def matrix_from_expressions(row_exprs, column_exprs):
row_exprs = [fold_constants(expand(expr)) for expr in row_exprs]
result = numpy.zeros((len(row_exprs), len(column_exprs)), dtype=object)
for i, row_expr in enumerate(row_exprs):
for j, col_expr in enumerate(column_exprs):
result[i, j] = differentiate(row_expr, col_expr)
return result
from pytools import flatten
row_exprs = list(
flatten([
f_post_step,
s_post_step,
] + list(flatten([stepper.histories[hn] for hn in HIST_NAMES]))))
column_exprs = list(
flatten([f, s] +
list(flatten([orig_histories[hn] for hn in HIST_NAMES]))))
return matrix_from_expressions(row_exprs, column_exprs), \
column_exprs
def partition_mesh(mesh, partition, part_bdry_tag_factory):
"""*partition* is a mapping that maps element id to
integers that represent different pieces of the mesh.
For historical reasons, the values in partition are called
'parts'.
"""
# Find parts to which we need to distribute.
all_parts = list(set(partition[el.id] for el in mesh.elements))
# Prepare a mapping of elements to tags to speed up
# copy_el_tagger, below.
el2tags = {}
for tag, elements in mesh.tag_to_elements.iteritems():
if tag == hedge.mesh.TAG_ALL:
continue
for el in elements:
el2tags.setdefault(el, []).append(tag)
# prepare a mapping of (el, face_nr) to boundary_tags
# to speed up partition_bdry_tagger, below
elface2tags = {}
for tag, elfaces in mesh.tag_to_boundary.iteritems():
if tag == hedge.mesh.TAG_ALL:
continue
for el, fn in elfaces:
elface2tags.setdefault((el, fn), []).append(tag)
# prepare a mapping from (el, face_nr) to the part
# at the other end of the interface, if different from
# current. concurrently, prepare a mapping
# part -> set([parts that border me])
elface2part = {}
neighboring_parts = {}
for elface1, elface2 in mesh.interfaces:
e1, f1 = elface1
e2, f2 = elface2
r1 = partition[e1.id]
r2 = partition[e2.id]
if r1 != r2:
neighboring_parts.setdefault(r1, set()).add(r2)
neighboring_parts.setdefault(r2, set()).add(r1)
elface2part[elface1] = r2
elface2part[elface2] = r1
# prepare a new mesh for each part and send it
from hedge.mesh import TAG_NO_BOUNDARY
for part in all_parts:
part_global_elements = [
el for el in mesh.elements if partition[el.id] == part
]
# pick out this part's vertices
from pytools import flatten
part_global_vertex_indices = set(
flatten(el.vertex_indices for el in part_global_elements))
part_local_vertices = [
mesh.points[vi] for vi in part_global_vertex_indices
]
# find global-to-local maps
part_global2local_vertex_indices = dict(
(gvi, lvi) for lvi, gvi in enumerate(part_global_vertex_indices))
part_global2local_elements = dict(
(el.id, i) for i, el in enumerate(part_global_elements))
# find elements in local numbering
part_local_elements = [[
part_global2local_vertex_indices[vi] for vi in el.vertex_indices
] for el in part_global_elements]
# make new local Mesh object, including
# boundary and element tagging
def partition_bdry_tagger(fvi, local_el, fn, all_vertices):
el = part_global_elements[local_el.id]
result = elface2tags.get((el, fn), [])
try:
opp_part = elface2part[el, fn]
result.append(part_bdry_tag_factory(opp_part))
# keeps this part of the boundary from falling
# under TAG_ALL.
result.append(TAG_NO_BOUNDARY)
except KeyError:
pass
return result
def copy_el_tagger(local_el, all_vertices):
return el2tags.get(part_global_elements[local_el.id], [])
def is_partbdry_face((local_el, face_nr)):
return (part_global_elements[local_el.id], face_nr) in elface2part
from hedge.mesh import make_conformal_mesh
part_mesh = make_conformal_mesh(part_local_vertices,
part_local_elements,
partition_bdry_tagger, copy_el_tagger,
mesh.periodicity, is_partbdry_face)
# assemble per-part data
my_nb_parts = neighboring_parts.get(part, [])
yield PartitionData(part,
part_mesh,
part_global2local_elements,
part_global2local_vertex_indices,
my_nb_parts,
mesh.periodic_opposite_faces,
part_boundary_tags=dict(
(nb_part, part_bdry_tag_factory(nb_part))
for nb_part in my_nb_parts),
tag_to_elements=part_mesh.tag_to_elements)
def partition_mesh(mesh, partition, part_bdry_tag_factory):
"""*partition* is a mapping that maps element id to
integers that represent different pieces of the mesh.
For historical reasons, the values in partition are called
'parts'.
"""
# Find parts to which we need to distribute.
all_parts = list(set(
partition[el.id] for el in mesh.elements))
# Prepare a mapping of elements to tags to speed up
# copy_el_tagger, below.
el2tags = {}
for tag, elements in mesh.tag_to_elements.iteritems():
if tag == hedge.mesh.TAG_ALL:
continue
for el in elements:
el2tags.setdefault(el, []).append(tag)
# prepare a mapping of (el, face_nr) to boundary_tags
# to speed up partition_bdry_tagger, below
elface2tags = {}
for tag, elfaces in mesh.tag_to_boundary.iteritems():
if tag == hedge.mesh.TAG_ALL:
continue
for el, fn in elfaces:
elface2tags.setdefault((el, fn), []).append(tag)
# prepare a mapping from (el, face_nr) to the part
# at the other end of the interface, if different from
# current. concurrently, prepare a mapping
# part -> set([parts that border me])
elface2part = {}
neighboring_parts = {}
for elface1, elface2 in mesh.interfaces:
e1, f1 = elface1
e2, f2 = elface2
r1 = partition[e1.id]
r2 = partition[e2.id]
if r1 != r2:
neighboring_parts.setdefault(r1, set()).add(r2)
neighboring_parts.setdefault(r2, set()).add(r1)
elface2part[elface1] = r2
elface2part[elface2] = r1
# prepare a new mesh for each part and send it
from hedge.mesh import TAG_NO_BOUNDARY
for part in all_parts:
part_global_elements = [el
for el in mesh.elements
if partition[el.id] == part]
# pick out this part's vertices
from pytools import flatten
part_global_vertex_indices = set(flatten(
el.vertex_indices for el in part_global_elements))
part_local_vertices = [mesh.points[vi]
for vi in part_global_vertex_indices]
# find global-to-local maps
part_global2local_vertex_indices = dict(
(gvi, lvi) for lvi, gvi in
enumerate(part_global_vertex_indices))
part_global2local_elements = dict(
(el.id, i) for i, el in
enumerate(part_global_elements))
# find elements in local numbering
part_local_elements = [
[part_global2local_vertex_indices[vi]
for vi in el.vertex_indices]
for el in part_global_elements]
# make new local Mesh object, including
# boundary and element tagging
def partition_bdry_tagger(fvi, local_el, fn, all_vertices):
el = part_global_elements[local_el.id]
result = elface2tags.get((el, fn), [])
try:
opp_part = elface2part[el, fn]
result.append(part_bdry_tag_factory(opp_part))
# keeps this part of the boundary from falling
# under TAG_ALL.
result.append(TAG_NO_BOUNDARY)
except KeyError:
pass
return result
def copy_el_tagger(local_el, all_vertices):
return el2tags.get(part_global_elements[local_el.id], [])
def is_partbdry_face((local_el, face_nr)):
return (part_global_elements[local_el.id], face_nr) in elface2part
from hedge.mesh import make_conformal_mesh
part_mesh = make_conformal_mesh(
part_local_vertices,
part_local_elements,
partition_bdry_tagger,
copy_el_tagger,
mesh.periodicity,
is_partbdry_face)
# assemble per-part data
my_nb_parts = neighboring_parts.get(part, [])
yield PartitionData(
part,
part_mesh,
part_global2local_elements,
part_global2local_vertex_indices,
my_nb_parts,
mesh.periodic_opposite_faces,
part_boundary_tags=dict(
(nb_part, part_bdry_tag_factory(nb_part))
for nb_part in my_nb_parts),
tag_to_elements = part_mesh.tag_to_elements
)
def diagonal(self, *, get_data=True, target_array=None, flatten=True):
"""Obtain the diagonal of the density matrix.
Parameters
----------
target_array : None or pycuda.gpuarray.array
An already-allocated GPU array to which the data will be copied.
If `None`, make a new GPU array.
get_data : boolean
Whether the data should be copied from the GPU.
flatten : boolean
TODO docstring
"""
diag_bases = [pb.computational_subbasis() for pb in self.bases]
diag_shape = [db.dim_pauli for db in diag_bases]
diag_size = pytools.product(diag_shape)
if target_array is None:
if self._work_data.gpudata.size < diag_size * 8:
self._work_data.gpudata.free()
self._work_data = ga.empty(diag_shape, np.float64)
self._work_data.gpudata.size = self._work_data.nbytes
target_array = self._work_data
else:
if target_array.size < diag_size:
raise ValueError(
"Size of `target_gpu_array` is too small ({}).\n"
"Should be at least {}."
.format(target_array.size, diag_size))
idx = [[pb.computational_basis_indices[i]
for i in range(pb.dim_hilbert)
if pb.computational_basis_indices[i] is not None]
for pb in self.bases]
idx_j = np.array(list(pytools.flatten(idx))).astype(np.uint32)
idx_i = np.cumsum([0] + [len(i) for i in idx][:-1]).astype(np.uint32)
xshape = np.array(self._data.shape, np.uint32)
yshape = np.array(diag_shape, np.uint32)
xshape_gpu = self._cached_gpuarray(xshape)
yshape_gpu = self._cached_gpuarray(yshape)
idx_i_gpu = self._cached_gpuarray(idx_i)
idx_j_gpu = self._cached_gpuarray(idx_j)
block = (2 ** 8, 1, 1)
grid = (max(1, (diag_size - 1) // 2 ** 8 + 1), 1, 1)
if len(yshape) == 0:
# brain-dead case, but should be handled according to exp.
target_array.set(self._data.get())
else:
_multitake.prepared_call(
grid, block, self._data.gpudata, target_array.gpudata,
idx_i_gpu.gpudata, idx_j_gpu.gpudata,
xshape_gpu.gpudata, yshape_gpu.gpudata,
np.uint32(len(yshape))
)
if get_data:
if flatten:
return target_array.get().ravel()[:diag_size]
else:
return (target_array.get().ravel()[:diag_size]
.reshape(diag_shape))
else:
return ga.GPUArray(shape=diag_shape,
gpudata=target_array.gpudata,
dtype=np.float64)
def make_boundary_restriction(queue, discr, group_factory):
"""
:return: a tuple ``(bdry_mesh, bdry_discr, connection)``
"""
logger.info("building boundary connection: start")
# {{{ build face_map
# maps (igrp, el_grp, face_id) to a frozenset of vertex IDs
face_map = {}
for igrp, mgrp in enumerate(discr.mesh.groups):
grp_face_vertex_indices = mgrp.face_vertex_indices()
for iel_grp in range(mgrp.nelements):
for fid, loc_face_vertices in enumerate(grp_face_vertex_indices):
face_vertices = frozenset(
mgrp.vertex_indices[iel_grp, fvi]
for fvi in loc_face_vertices
)
face_map.setdefault(face_vertices, []).append(
(igrp, iel_grp, fid))
del face_vertices
# }}}
boundary_faces = [
face_ids[0]
for face_vertices, face_ids in six.iteritems(face_map)
if len(face_ids) == 1]
from pytools import flatten
bdry_vertex_vol_nrs = sorted(set(flatten(six.iterkeys(face_map))))
vol_to_bdry_vertices = np.empty(
discr.mesh.vertices.shape[-1],
discr.mesh.vertices.dtype)
vol_to_bdry_vertices.fill(-1)
vol_to_bdry_vertices[bdry_vertex_vol_nrs] = np.arange(
len(bdry_vertex_vol_nrs))
bdry_vertices = discr.mesh.vertices[:, bdry_vertex_vol_nrs]
from meshmode.mesh import Mesh, SimplexElementGroup
bdry_mesh_groups = []
connection_data = {}
for igrp, grp in enumerate(discr.groups):
mgrp = grp.mesh_el_group
group_boundary_faces = [
(ibface_el, ibface_face)
for ibface_group, ibface_el, ibface_face in boundary_faces
if ibface_group == igrp]
if not isinstance(mgrp, SimplexElementGroup):
raise NotImplementedError("can only take boundary of "
"SimplexElementGroup-based meshes")
# {{{ Preallocate arrays for mesh group
ngroup_bdry_elements = len(group_boundary_faces)
vertex_indices = np.empty(
(ngroup_bdry_elements, mgrp.dim+1-1),
mgrp.vertex_indices.dtype)
bdry_unit_nodes = mp.warp_and_blend_nodes(mgrp.dim-1, mgrp.order)
bdry_unit_nodes_01 = (bdry_unit_nodes + 1)*0.5
vol_basis = mp.simplex_onb(mgrp.dim, mgrp.order)
nbdry_unit_nodes = bdry_unit_nodes_01.shape[-1]
nodes = np.empty(
(discr.ambient_dim, ngroup_bdry_elements, nbdry_unit_nodes),
dtype=np.float64)
# }}}
grp_face_vertex_indices = mgrp.face_vertex_indices()
grp_vertex_unit_coordinates = mgrp.vertex_unit_coordinates()
# batch by face_id
batch_base = 0
for face_id in range(len(grp_face_vertex_indices)):
batch_boundary_el_numbers_in_grp = np.array(
[
ibface_el
for ibface_el, ibface_face in group_boundary_faces
if ibface_face == face_id],
dtype=np.intp)
new_el_numbers = np.arange(
batch_base,
batch_base + len(batch_boundary_el_numbers_in_grp))
# {{{ no per-element axes in these computations
# Find boundary vertex indices
loc_face_vertices = list(grp_face_vertex_indices[face_id])
# Find unit nodes for boundary element
face_vertex_unit_coordinates = \
grp_vertex_unit_coordinates[loc_face_vertices]
# Find A, b such that A [e_1 e_2] + b = [r_1 r_2]
# (Notation assumes that the volume is 3D and the face is 2D.
# Code does not.)
b = face_vertex_unit_coordinates[0]
A = (
face_vertex_unit_coordinates[1:]
- face_vertex_unit_coordinates[0]).T
face_unit_nodes = (np.dot(A, bdry_unit_nodes_01).T + b).T
resampling_mat = mp.resampling_matrix(
vol_basis,
face_unit_nodes, mgrp.unit_nodes)
# }}}
# {{{ build information for mesh element group
# Find vertex_indices
glob_face_vertices = mgrp.vertex_indices[
batch_boundary_el_numbers_in_grp][:, loc_face_vertices]
vertex_indices[new_el_numbers] = \
vol_to_bdry_vertices[glob_face_vertices]
# Find nodes
nodes[:, new_el_numbers, :] = np.einsum(
"ij,dej->dei",
resampling_mat,
mgrp.nodes[:, batch_boundary_el_numbers_in_grp, :])
# }}}
connection_data[igrp, face_id] = _ConnectionBatchData(
group_source_element_indices=batch_boundary_el_numbers_in_grp,
group_target_element_indices=new_el_numbers,
A=A,
b=b,
)
batch_base += len(batch_boundary_el_numbers_in_grp)
bdry_mesh_group = SimplexElementGroup(
mgrp.order, vertex_indices, nodes, unit_nodes=bdry_unit_nodes)
bdry_mesh_groups.append(bdry_mesh_group)
bdry_mesh = Mesh(bdry_vertices, bdry_mesh_groups)
from meshmode.discretization import Discretization
bdry_discr = Discretization(
discr.cl_context, bdry_mesh, group_factory)
connection = _build_boundary_connection(
queue, discr, bdry_discr, connection_data)
logger.info("building boundary connection: done")
return bdry_mesh, bdry_discr, connection
def test_mesh_multiple_groups(actx_factory, ambient_dim, visualize=False):
actx = actx_factory()
order = 4
mesh = mgen.generate_regular_rect_mesh(a=(-0.5, ) * ambient_dim,
b=(0.5, ) * ambient_dim,
nelements_per_axis=(8, ) *
ambient_dim,
order=order)
assert len(mesh.groups) == 1
from meshmode.mesh.processing import split_mesh_groups
element_flags = np.any(
mesh.vertices[0, mesh.groups[0].vertex_indices] < 0.0,
axis=1).astype(np.int64)
mesh = split_mesh_groups(mesh, element_flags)
assert len(mesh.groups) == 2 # pylint: disable=no-member
assert mesh.facial_adjacency_groups
assert mesh.nodal_adjacency
if visualize and ambient_dim == 2:
from meshmode.mesh.visualization import draw_2d_mesh
draw_2d_mesh(mesh,
draw_vertex_numbers=False,
draw_element_numbers=True,
draw_face_numbers=False,
set_bounding_box=True)
import matplotlib.pyplot as plt
plt.savefig("test_mesh_multiple_groups_2d_elements.png", dpi=300)
from meshmode.discretization import Discretization
discr = Discretization(actx, mesh,
PolynomialWarpAndBlendGroupFactory(order))
if visualize:
group_id = discr.empty(actx, dtype=np.int32)
for igrp, vec in enumerate(group_id):
vec.fill(igrp)
from meshmode.discretization.visualization import make_visualizer
vis = make_visualizer(actx, discr, vis_order=order)
vis.write_vtk_file("mesh_multiple_groups.vtu",
[("group_id", group_id)],
overwrite=True)
# check face restrictions
from meshmode.discretization.connection import (
make_face_restriction, make_face_to_all_faces_embedding,
make_opposite_face_connection, check_connection)
for boundary_tag in [BTAG_ALL, FACE_RESTR_INTERIOR, FACE_RESTR_ALL]:
conn = make_face_restriction(
actx,
discr,
group_factory=PolynomialWarpAndBlendGroupFactory(order),
boundary_tag=boundary_tag,
per_face_groups=False)
check_connection(actx, conn)
bdry_f = conn.to_discr.zeros(actx) + 1
if boundary_tag == FACE_RESTR_INTERIOR:
opposite = make_opposite_face_connection(actx, conn)
check_connection(actx, opposite)
op_bdry_f = opposite(bdry_f)
error = flat_norm(bdry_f - op_bdry_f, np.inf)
assert error < 1.0e-11, error
if boundary_tag == FACE_RESTR_ALL:
embedding = make_face_to_all_faces_embedding(
actx, conn, conn.to_discr)
check_connection(actx, embedding)
em_bdry_f = embedding(bdry_f)
error = flat_norm(bdry_f - em_bdry_f)
assert error < 1.0e-11, error
# check some derivatives (nb: flatten is a generator)
import pytools
ref_axes = pytools.flatten([[i] for i in range(ambient_dim)])
from meshmode.discretization import num_reference_derivative
x = thaw(discr.nodes(), actx)
num_reference_derivative(discr, ref_axes, x[0])
def _setup_neighbor_connections(self):
comm = self.context.communicator
# Why is this barrier needed? Some of our ranks may arrive at this
# point early and start sending packets to ranks that are still stuck
# in previous wildcard-recv loops. These receivers will then be very
# confused by packets they didn't expect, and, once they reach their
# recv bit in *this* subroutine, will wait for packets that will never
# arrive. This same argument does not apply to other recv()s in this
# file because they are targeted and thus benefit from MPI's
# non-overtaking rule.
#
# Parallel programming is fun.
comm.Barrier()
if self.neighbor_ranks:
# send interface information to neighboring ranks -----------------
from pytools import reverse_dictionary
local2global_vertex_indices = \
reverse_dictionary(self.global2local_vertex_indices)
send_requests = []
for rank in self.neighbor_ranks:
bdry_tag = hedge.mesh.TAG_RANK_BOUNDARY(rank)
rank_bdry = self.subdiscr.mesh.tag_to_boundary[bdry_tag]
rank_discr_boundary = self.subdiscr.get_boundary(bdry_tag)
# a list of global vertex numbers for each face
my_vertices_global = [
tuple(local2global_vertex_indices[vi]
for vi in el.faces[face_nr])
for el, face_nr in rank_bdry
]
# a list of node coordinates, indicating the order
# in which nodal values will be sent, this is for
# testing only and could (potentially) be omitted
my_node_coords = []
for el, face_nr in rank_bdry:
eslice, ldis = self.subdiscr.find_el_data(el.id)
findices = ldis.face_indices()[face_nr]
my_node_coords.append(
[self.nodes[eslice.start + i] for i in findices])
# compile a list of FluxFace.h values for unification
# across the rank boundary
my_h_values = [
rank_discr_boundary.find_facepair_side(el_face).h
for el_face in rank_bdry
]
packet = (my_vertices_global, my_node_coords, my_h_values)
send_requests.append(comm.isend(packet, dest=rank, tag=0))
received_packets = {}
while len(received_packets) < len(self.neighbor_ranks):
status = mpi.Status()
received_packet = comm.recv(tag=0,
source=mpi.ANY_SOURCE,
status=status)
received_packets[status.source] = received_packet
mpi.Request.Waitall(send_requests)
# process received packets ----------------------------------------
from pytools import flatten
# nb_ stands for neighbor_
self.from_neighbor_maps = {}
for rank, (nb_all_facevertices_global, nb_node_coords, nb_h_values) in \
received_packets.iteritems():
bdry_tag = hedge.mesh.TAG_RANK_BOUNDARY(rank)
rank_bdry = self.subdiscr.mesh.tag_to_boundary[bdry_tag]
rank_discr_boundary = self.subdiscr.get_boundary(bdry_tag)
flat_nb_node_coords = list(flatten(nb_node_coords))
# step 1: find start node indices for each
# of the neighbor's elements
nb_face_starts = [0]
for node_coords in nb_node_coords[:-1]:
nb_face_starts.append(nb_face_starts[-1] +
len(node_coords))
# step 2: match faces by matching vertices
nb_face_order = dict(
(frozenset(vertices), i)
for i, vertices in enumerate(nb_all_facevertices_global))
# step 3: make a list of indices into the data we
# receive from our neighbor that'll tell us how
# to reshuffle them to match our node order
from_indices = []
shuffled_indices_cache = {}
def get_shuffled_indices(face_node_count, shuffle_op):
try:
return shuffled_indices_cache[shuffle_op]
except KeyError:
unshuffled_indices = range(face_node_count)
result = shuffled_indices_cache[shuffle_op] = \
shuffle_op(unshuffled_indices)
return result
for el, face_nr in rank_bdry:
eslice, ldis = self.subdiscr.find_el_data(el.id)
my_vertices = el.faces[face_nr]
my_global_vertices = tuple(local2global_vertex_indices[vi]
for vi in my_vertices)
face_node_count = ldis.face_node_count()
try:
nb_face_idx = nb_face_order[frozenset(
my_global_vertices)]
# continue below in else part
except KeyError:
# this happens if my_global_vertices is not a permutation
# of the neighbor's face vertices. Periodicity is the only
# reason why that would be so.
my_vertices_there, axis = \
self.global_periodic_opposite_faces[
my_global_vertices]
nb_face_idx = nb_face_order[frozenset(
my_vertices_there)]
his_vertices_here, axis2 = \
self.global_periodic_opposite_faces[
nb_all_facevertices_global[nb_face_idx]]
assert axis == axis2
nb_face_start = nb_face_starts[nb_face_idx]
shuffle_op = \
ldis.get_face_index_shuffle_to_match(
my_global_vertices,
his_vertices_here)
shuffled_other_node_indices = [
nb_face_start + i for i in get_shuffled_indices(
face_node_count, shuffle_op)
]
from_indices.extend(shuffled_other_node_indices)
# check if the nodes really match up
if "parallel_setup" in self.debug:
my_node_indices = [
eslice.start + i
for i in ldis.face_indices()[face_nr]
]
for my_i, other_i in zip(
my_node_indices,
shuffled_other_node_indices):
dist = self.nodes[my_i] - flat_nb_node_coords[
other_i]
dist[axis] = 0
assert la.norm(dist) < 1e-14
else:
# continue handling of nonperiodic case
nb_global_vertices = nb_all_facevertices_global[
nb_face_idx]
nb_face_start = nb_face_starts[nb_face_idx]
shuffle_op = \
ldis.get_face_index_shuffle_to_match(
my_global_vertices,
nb_global_vertices)
shuffled_other_node_indices = [
nb_face_start + i for i in get_shuffled_indices(
face_node_count, shuffle_op)
]
from_indices.extend(shuffled_other_node_indices)
# check if the nodes really match up
if "parallel_setup" in self.debug:
my_node_indices = [
eslice.start + i
for i in ldis.face_indices()[face_nr]
]
for my_i, other_i in zip(
my_node_indices,
shuffled_other_node_indices):
dist = self.nodes[my_i] - flat_nb_node_coords[
other_i]
assert la.norm(dist) < 1e-14
# finally, unify FluxFace.h values across boundary
nb_h = nb_h_values[nb_face_idx]
flux_face = rank_discr_boundary.find_facepair_side(
(el, face_nr))
flux_face.h = max(nb_h, flux_face.h)
if "parallel_setup" in self.debug:
assert len(from_indices) == len(flat_nb_node_coords)
# construct from_neighbor_map
self.from_neighbor_maps[rank] = \
self.subdiscr.prepare_from_neighbor_map(from_indices)
def initialize(self, method):
Depositor.initialize(self, method)
backend_class = getattr(_internal, "GridFindDepositor"
+ method.get_dimensionality_suffix())
backend = self.backend = backend_class(method.mesh_data)
discr = method.discretization
grid_node_num_to_nodes = {}
if self.brick_generator is None:
bbox_min, bbox_max = discr.mesh.bounding_box()
max_bbox_size = max(bbox_max-bbox_min)
self.brick_generator = SingleBrickGenerator(
mesh_margin=1e-3*max_bbox_size,
overresolve=0.2)
from pyrticle._internal import Brick
for i, (stepwidths, origin, dims) in enumerate(
self.brick_generator(discr)):
backend.bricks.append(
Brick(i, backend.grid_node_count(), stepwidths, origin, dims))
from pyrticle._internal import BoxFloat
for eg in discr.element_groups:
ldis = eg.local_discretization
for el in eg.members:
el_bbox = BoxFloat(*el.bounding_box(discr.mesh.points))
el_slice = discr.find_el_range(el.id)
for brk in backend.bricks:
if brk.bounding_box().intersect(el_bbox).is_empty():
continue
for node_num in range(el_slice.start, el_slice.stop):
try:
cell_number = brk.which_cell(
discr.nodes[node_num])
except ValueError:
pass
else:
grid_node_num_to_nodes.setdefault(
brk.index(cell_number), []).append(node_num)
from pytools import flatten
unassigned_nodes = (set(xrange(len(discr)))
- set(flatten(grid_node_num_to_nodes.itervalues())))
if unassigned_nodes:
raise RuntimeError("dep_grid_find: unassigned mesh nodes found. "
"you should specify a mesh_margin when generating "
"bricks")
usecounts = numpy.zeros(
(backend.grid_node_count(),))
for gnn in xrange(backend.grid_node_count()):
grid_nodes = grid_node_num_to_nodes.get(gnn, [])
usecounts[gnn] = len(grid_nodes)
backend.node_number_list_starts.append(len(backend.node_number_lists))
backend.node_number_lists.extend(grid_nodes)
backend.node_number_list_starts.append(len(backend.node_number_lists))
if "depositor" in method.debug:
from hedge.visualization import SiloVisualizer
vis = SiloVisualizer(discr)
visf = vis.make_file("grid-find-debug")
vis.add_data(visf, [])
self.visualize_grid_quantities(visf, [
("usecounts", usecounts)
])
visf.close()
if "interactive" in cloud.debug:
from matplotlib.pylab import hist, show
hist(usecounts, bins=20)
show()
def make_conformal_mesh_ext(points, elements,
boundary_tagger=None,
volume_tagger=None,
periodicity=None,
allow_internal_boundaries=False,
_is_rankbdry_face=None,
):
"""Construct a simplicial mesh.
Face indices follow the convention for the respective element,
such as Triangle or Tetrahedron, in this module.
:param points: an array of vertex coordinates, given as vectors.
:param elements: an iterable of :class:`hedge.mesh.element.Element`
instances.
:param boundary_tagger: A function of *(fvi, el, fn, all_v)*
that returns a list of boundary tags for a face identified
by the parameters.
*fvi* is the set of vertex indices of the face
in question, *el* is an :class:`Element` instance,
*fn* is the face number within *el*, and *all_v* is
a list of all vertices.
:param volume_tagger: A function of *(el, all_v)*
returning a list of volume tags for the element identified
by the parameters.
*el* is an :class:`Element` instance and *all_v* is a list of
all vertex coordinates.
:param periodicity: either None or is a list of tuples
just like the one documented for the `periodicity`
member of class :class:`Mesh`.
:param allow_internal_boundaries: Calls the boundary tagger
for element-element interfaces as well. If the tagger returns
an empty list of tags for an internal interface, it remains
internal.
:param _is_rankbdry_face: an implementation detail,
should not be used from user code. It is a function
returning whether a given face identified by
*(element instance, face_nr)* is cut by a parallel
mesh partition.
"""
# input validation
if (not isinstance(points, numpy.ndarray)
or not points.dtype == numpy.float64):
raise TypeError("points must be a float64 array")
if boundary_tagger is None:
def boundary_tagger(fvi, el, fn, all_v):
return []
if volume_tagger is None:
def volume_tagger(el, all_v):
return []
if _is_rankbdry_face is None:
def _is_rankbdry_face(el_face):
return False
dim = max(el.dimensions for el in elements)
if periodicity is None:
periodicity = dim*[None]
assert len(periodicity) == dim
# tag elements
tag_to_elements = {TAG_NONE: [], TAG_ALL: []}
for el in elements:
for el_tag in volume_tagger(el, points):
tag_to_elements.setdefault(el_tag, []).append(el)
tag_to_elements[TAG_ALL].append(el)
# create face_map, which is a mapping of
# (vertices on a face) ->
# [(element, face_idx) for elements bordering that face]
face_map = {}
for el in elements:
for fid, face_vertices in enumerate(el.faces):
face_map.setdefault(frozenset(face_vertices), []).append((el, fid))
# build non-periodic connectivity structures
interfaces = []
tag_to_boundary = {
TAG_NONE: [],
TAG_ALL: [],
TAG_REALLY_ALL: [],
}
for face_vertices, els_faces in face_map.iteritems():
boundary_el_faces_tags = []
if len(els_faces) == 2:
if allow_internal_boundaries:
el_face_a, el_face_b = els_faces
el_a, face_a = el_face_a
el_b, face_b = el_face_b
tags_a = boundary_tagger(face_vertices, el_a, face_a, points)
tags_b = boundary_tagger(face_vertices, el_b, face_b, points)
if not tags_a and not tags_b:
interfaces.append(els_faces)
elif tags_a and tags_b:
boundary_el_faces_tags.append((el_face_a, tags_a))
boundary_el_faces_tags.append((el_face_b, tags_b))
else:
raise RuntimeError("boundary tagger is inconsistent "
"about boundary-ness of interior interface")
else:
interfaces.append(els_faces)
elif len(els_faces) == 1:
el_face = el, face = els_faces[0]
tags = boundary_tagger(face_vertices, el, face, points)
boundary_el_faces_tags.append((el_face, tags))
else:
raise RuntimeError("face can at most border two elements")
for el_face, tags in boundary_el_faces_tags:
el, face = el_face
tags = set(tags) - MESH_CREATION_TAGS
assert not isinstance(tags, str), \
RuntimeError("Received string as tag list")
assert TAG_ALL not in tags
assert TAG_REALLY_ALL not in tags
for btag in tags:
tag_to_boundary.setdefault(btag, []) \
.append(el_face)
if TAG_NO_BOUNDARY not in tags:
# TAG_NO_BOUNDARY is used to mark rank interfaces
# as not being part of the boundary
tag_to_boundary[TAG_ALL].append(el_face)
tag_to_boundary[TAG_REALLY_ALL].append(el_face)
# add periodicity-induced connectivity
from pytools import flatten, reverse_dictionary
periodic_opposite_faces = {}
periodic_opposite_vertices = {}
for tag_bdries in tag_to_boundary.itervalues():
assert len(set(tag_bdries)) == len(tag_bdries)
for axis, axis_periodicity in enumerate(periodicity):
if axis_periodicity is not None:
# find faces on +-axis boundaries
minus_tag, plus_tag = axis_periodicity
minus_faces = tag_to_boundary.get(minus_tag, [])
plus_faces = tag_to_boundary.get(plus_tag, [])
# find vertex indices and points on these faces
minus_vertex_indices = list(set(flatten(el.faces[face]
for el, face in minus_faces)))
plus_vertex_indices = list(set(flatten(el.faces[face]
for el, face in plus_faces)))
minus_z_points = [points[pi] for pi in minus_vertex_indices]
plus_z_points = [points[pi] for pi in plus_vertex_indices]
# find a mapping from -axis to +axis vertices
minus_to_plus, not_found = find_matching_vertices_along_axis(
axis, minus_z_points, plus_z_points,
minus_vertex_indices, plus_vertex_indices)
plus_to_minus = reverse_dictionary(minus_to_plus)
for a, b in minus_to_plus.iteritems():
periodic_opposite_vertices.setdefault(a, []).append((b, axis))
periodic_opposite_vertices.setdefault(b, []).append((a, axis))
# establish face connectivity
for minus_face in minus_faces:
minus_el, minus_fi = minus_face
minus_fvi = minus_el.faces[minus_fi]
try:
mapped_plus_fvi = tuple(minus_to_plus[i] for i in minus_fvi)
plus_faces = face_map[frozenset(mapped_plus_fvi)]
assert len(plus_faces) == 1
except KeyError:
# is our periodic counterpart is in a different mesh clump?
if _is_rankbdry_face(minus_face):
# if so, cool. parallel handler will take care of it.
continue
else:
# if not, bad.
raise
plus_face = plus_faces[0]
interfaces.append([minus_face, plus_face])
plus_el, plus_fi = plus_face
plus_fvi = plus_el.faces[plus_fi]
mapped_minus_fvi = tuple(plus_to_minus[i] for i in plus_fvi)
# periodic_opposite_faces maps face vertex tuples from
# one end of the periodic domain to the other, while
# correspondence between each entry
periodic_opposite_faces[minus_fvi] = mapped_plus_fvi, axis
periodic_opposite_faces[plus_fvi] = mapped_minus_fvi, axis
tag_to_boundary[TAG_ALL].remove(plus_face)
tag_to_boundary[TAG_ALL].remove(minus_face)
tag_to_boundary[TAG_REALLY_ALL].remove(plus_face)
tag_to_boundary[TAG_REALLY_ALL].remove(minus_face)
return ConformalMesh(
points=points,
elements=elements,
interfaces=interfaces,
tag_to_boundary=tag_to_boundary,
tag_to_elements=tag_to_elements,
periodicity=periodicity,
periodic_opposite_faces=periodic_opposite_faces,
periodic_opposite_vertices=periodic_opposite_vertices,
has_internal_boundaries=allow_internal_boundaries,
)