def simple_result_dtype_getter(vector_dtype_map, scalar_dtype_map,
const_dtypes):
from pytools import common_dtype, match_precision
result = common_dtype(vector_dtype_map.values())
scalar_dtypes = scalar_dtype_map.values() + const_dtypes
if scalar_dtypes:
prec_matched_scalar_dtype = match_precision(
common_dtype(scalar_dtypes), dtype_to_match=result)
result = common_dtype([result, prec_matched_scalar_dtype])
return result
def simple_result_dtype_getter(vector_dtype_map, scalar_dtype_map, const_dtypes):
from pytools import common_dtype, match_precision
result = common_dtype(vector_dtype_map.values())
scalar_dtypes = scalar_dtype_map.values() + const_dtypes
if scalar_dtypes:
prec_matched_scalar_dtype = match_precision(
common_dtype(scalar_dtypes),
dtype_to_match=result)
result = common_dtype([result, prec_matched_scalar_dtype])
return result
def exec_flux_batch_assign(self, insn):
from pymbolic.primitives import is_zero
class ZeroSpec:
pass
class BoundaryZeros(ZeroSpec):
pass
class VolumeZeros(ZeroSpec):
pass
def eval_arg(arg_spec):
arg_expr, is_int = arg_spec
arg = self.rec(arg_expr)
if is_zero(arg):
if insn.is_boundary and not is_int:
return BoundaryZeros()
else:
return VolumeZeros()
else:
return arg
args = [
eval_arg(arg_expr) for arg_expr in insn.flux_var_info.arg_specs
]
from pytools import common_dtype
max_dtype = common_dtype(
[a.dtype for a in args if not isinstance(a, ZeroSpec)],
self.discr.default_scalar_type)
def cast_arg(arg):
if isinstance(arg, BoundaryZeros):
return self.discr.boundary_zeros(insn.repr_op.boundary_tag,
dtype=max_dtype)
elif isinstance(arg, VolumeZeros):
return self.discr.volume_zeros(dtype=max_dtype)
elif isinstance(arg, np.ndarray):
return np.asarray(arg, dtype=max_dtype)
else:
return arg
args = [cast_arg(arg) for arg in args]
if insn.quadrature_tag is None:
if insn.is_boundary:
face_groups = self.discr.get_boundary(insn.repr_op.boundary_tag)\
.face_groups
else:
face_groups = self.discr.face_groups
else:
if insn.is_boundary:
face_groups = self.discr.get_boundary(insn.repr_op.boundary_tag)\
.get_quadrature_info(insn.quadrature_tag).face_groups
else:
face_groups = self.discr.get_quadrature_info(insn.quadrature_tag) \
.face_groups
result = []
for fg in face_groups:
# grab module
module = insn.get_module(self.discr, max_dtype)
func = module.gather_flux
# set up argument structure
arg_struct = module.ArgStruct()
for arg_name, arg in zip(insn.flux_var_info.arg_names, args):
setattr(arg_struct, arg_name, arg)
for arg_num, scalar_arg_expr in enumerate(
insn.flux_var_info.scalar_parameters):
setattr(arg_struct, "_scalar_arg_%d" % arg_num,
self.rec(scalar_arg_expr))
fof_shape = (fg.face_count * fg.face_length() *
fg.element_count(), )
all_fluxes_on_faces = [
np.zeros(fof_shape, dtype=max_dtype) for f in insn.expressions
]
for i, fof in enumerate(all_fluxes_on_faces):
setattr(arg_struct, "flux%d_on_faces" % i, fof)
# make sure everything ended up in Boost.Python attributes
# (i.e. empty __dict__)
assert not arg_struct.__dict__, arg_struct.__dict__.keys()
# perform gather
func(fg, arg_struct)
# do lift, produce output
for name, flux_bdg, fluxes_on_faces in zip(insn.names,
insn.expressions,
all_fluxes_on_faces):
if insn.quadrature_tag is None:
if flux_bdg.op.is_lift:
mat = fg.ldis_loc.lifting_matrix()
scaling = fg.local_el_inverse_jacobians
else:
mat = fg.ldis_loc.multi_face_mass_matrix()
scaling = None
else:
assert not flux_bdg.op.is_lift
mat = fg.ldis_loc_quad_info.multi_face_mass_matrix()
scaling = None
out = self.discr.volume_zeros(dtype=fluxes_on_faces.dtype)
self.executor.lift_flux(fg, mat, scaling, fluxes_on_faces, out)
if self.discr.instrumented:
from hedge.tools import lift_flops
# correct for quadrature, too.
self.discr.lift_flop_counter.add(lift_flops(fg))
result.append((name, out))
if not face_groups:
# No face groups? Still assign context variables.
for name, flux_bdg in zip(insn.names, insn.expressions):
result.append((name, self.discr.volume_zeros()))
return result, []
def make_linear_comb_kernel(scalar_dtypes, vector_dtypes):
from pytools import common_dtype
result_dtype = common_dtype(scalar_dtypes+vector_dtypes)
return make_linear_comb_kernel_with_result_dtype(
result_dtype, scalar_dtypes, vector_dtypes), result_dtype
def exec_flux_batch_assign(self, insn):
from pymbolic.primitives import is_zero
class ZeroSpec:
pass
class BoundaryZeros(ZeroSpec):
pass
class VolumeZeros(ZeroSpec):
pass
def eval_arg(arg_spec):
arg_expr, is_int = arg_spec
arg = self.rec(arg_expr)
if is_zero(arg):
if insn.is_boundary and not is_int:
return BoundaryZeros()
else:
return VolumeZeros()
else:
return arg
args = [eval_arg(arg_expr)
for arg_expr in insn.flux_var_info.arg_specs]
from pytools import common_dtype
max_dtype = common_dtype(
[a.dtype for a in args if not isinstance(a, ZeroSpec)],
self.discr.default_scalar_type)
def cast_arg(arg):
if isinstance(arg, BoundaryZeros):
return self.discr.boundary_zeros(
insn.repr_op.boundary_tag, dtype=max_dtype)
elif isinstance(arg, VolumeZeros):
return self.discr.volume_zeros(
dtype=max_dtype)
elif isinstance(arg, np.ndarray):
return np.asarray(arg, dtype=max_dtype)
else:
return arg
args = [cast_arg(arg) for arg in args]
if insn.quadrature_tag is None:
if insn.is_boundary:
face_groups = self.discr.get_boundary(insn.repr_op.boundary_tag)\
.face_groups
else:
face_groups = self.discr.face_groups
else:
if insn.is_boundary:
face_groups = self.discr.get_boundary(insn.repr_op.boundary_tag)\
.get_quadrature_info(insn.quadrature_tag).face_groups
else:
face_groups = self.discr.get_quadrature_info(insn.quadrature_tag) \
.face_groups
result = []
for fg in face_groups:
# grab module
module = insn.get_module(self.discr, max_dtype)
func = module.gather_flux
# set up argument structure
arg_struct = module.ArgStruct()
for arg_name, arg in zip(insn.flux_var_info.arg_names, args):
setattr(arg_struct, arg_name, arg)
for arg_num, scalar_arg_expr in enumerate(
insn.flux_var_info.scalar_parameters):
setattr(arg_struct,
"_scalar_arg_%d" % arg_num,
self.rec(scalar_arg_expr))
fof_shape = (fg.face_count*fg.face_length()*fg.element_count(),)
all_fluxes_on_faces = [
np.zeros(fof_shape, dtype=max_dtype)
for f in insn.expressions]
for i, fof in enumerate(all_fluxes_on_faces):
setattr(arg_struct, "flux%d_on_faces" % i, fof)
# make sure everything ended up in Boost.Python attributes
# (i.e. empty __dict__)
assert not arg_struct.__dict__, arg_struct.__dict__.keys()
# perform gather
func(fg, arg_struct)
# do lift, produce output
for name, flux_bdg, fluxes_on_faces in zip(insn.names, insn.expressions,
all_fluxes_on_faces):
if insn.quadrature_tag is None:
if flux_bdg.op.is_lift:
mat = fg.ldis_loc.lifting_matrix()
scaling = fg.local_el_inverse_jacobians
else:
mat = fg.ldis_loc.multi_face_mass_matrix()
scaling = None
else:
assert not flux_bdg.op.is_lift
mat = fg.ldis_loc_quad_info.multi_face_mass_matrix()
scaling = None
out = self.discr.volume_zeros(dtype=fluxes_on_faces.dtype)
self.executor.lift_flux(fg, mat, scaling, fluxes_on_faces, out)
if self.discr.instrumented:
from hedge.tools import lift_flops
# correct for quadrature, too.
self.discr.lift_flop_counter.add(lift_flops(fg))
result.append((name, out))
if not face_groups:
# No face groups? Still assign context variables.
for name, flux_bdg in zip(insn.names, insn.expressions):
result.append((name, self.discr.volume_zeros()))
return result, []
def make_linear_comb_kernel(scalar_dtypes, vector_dtypes):
from pytools import common_dtype
result_dtype = common_dtype(scalar_dtypes + vector_dtypes)
return make_linear_comb_kernel_with_result_dtype(
result_dtype, scalar_dtypes, vector_dtypes), result_dtype