def get_value_arg_decl(self, name, shape, dtype, is_written):
result = super().get_value_arg_decl(name, shape, dtype, is_written)
from cgen import Reference, Const
was_const = isinstance(result, Const)
if was_const:
result = result.subdecl
if self.target.occa_mode:
result = Reference(result)
if was_const:
result = Const(result)
from cgen.ispc import ISPCUniform
return ISPCUniform(result)
def make_lift(self, fgroup, with_scale, dtype):
discr = self.discr
from cgen import (FunctionDeclaration, FunctionBody, Typedef, Const,
Reference, Value, POD, Statement, Include, Line,
Block, Initializer, Assign, For, If, Define)
from pytools import to_uncomplex_dtype
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
S = Statement
mod.add_to_preamble([
Include("hedge/face_operators.hpp"),
Include("hedge/volume_operators.hpp"),
Include("boost/foreach.hpp"),
])
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Define("DOFS_PER_EL", fgroup.ldis_loc.node_count()),
Define("FACES_PER_EL", fgroup.ldis_loc.face_count()),
Define("DIMENSIONS", discr.dimensions),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
def if_(cond, result, else_=None):
if cond:
return [result]
else:
if else_ is None:
return []
else:
return [else_]
fdecl = FunctionDeclaration(Value("void", "lift"), [
Const(
Reference(Value("face_group<face_pair<straight_face> >",
"fg"))),
Value("ublas::matrix<uncomplex_type>", "matrix"),
Value("numpy_array<value_type>", "field"),
Value("numpy_array<value_type>", "result")
] + if_(
with_scale,
Const(
Reference(Value("numpy_array<double>",
"elwise_post_scaling")))))
def make_it(name, is_const=True, tpname="value_type"):
if is_const:
const = "const_"
else:
const = ""
return Initializer(
Value("numpy_array<%s>::%siterator" % (tpname, const),
name + "_it"), "%s.begin()" % name)
fbody = Block([
make_it("field"),
make_it("result", is_const=False),
] + if_(with_scale, make_it("elwise_post_scaling", tpname="double")) + [
Line(),
For(
"unsigned fg_el_nr = 0", "fg_el_nr < fg.element_count()",
"++fg_el_nr",
Block([
Initializer(Value("node_number_t", "dest_el_base"),
"fg.local_el_write_base[fg_el_nr]"),
Initializer(Value("node_number_t", "src_el_base"),
"FACES_PER_EL*fg.face_length()*fg_el_nr"),
Line(),
For(
"unsigned i = 0", "i < DOFS_PER_EL", "++i",
Block([
Initializer(Value("value_type", "tmp"), 0),
Line(),
For(
"unsigned j = 0",
"j < FACES_PER_EL*fg.face_length()", "++j",
S("tmp += matrix(i, j)*field_it[src_el_base+j]"
)),
Line(),
] + if_(
with_scale,
Assign(
"result_it[dest_el_base+i]",
"tmp * value_type(*elwise_post_scaling_it)"),
Assign("result_it[dest_el_base+i]", "tmp")))),
] + if_(with_scale, S("elwise_post_scaling_it++"))))
])
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print FunctionBody(fdecl, fbody)
#raw_input()
return mod.compile(self.discr.toolchain).lift
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 make_diff(self, elgroup, dtype, shape):
"""
:param shape: If non-square, the resulting code takes two element_ranges
arguments and supports non-square matrices.
"""
from hedge._internal import UniformElementRanges
assert isinstance(elgroup.ranges, UniformElementRanges)
ldis = elgroup.local_discretization
discr = self.discr
from cgen import (
FunctionDeclaration, FunctionBody, Typedef,
Const, Reference, Value, POD,
Statement, Include, Line, Block, Initializer, Assign,
For, If,
Define)
from pytools import to_uncomplex_dtype
from codepy.bpl import BoostPythonModule
mod = BoostPythonModule()
# {{{ preamble
S = Statement
mod.add_to_preamble([
Include("hedge/volume_operators.hpp"),
Include("boost/foreach.hpp"),
])
mod.add_to_module([
S("namespace ublas = boost::numeric::ublas"),
S("using namespace hedge"),
S("using namespace pyublas"),
Line(),
Define("ROW_COUNT", shape[0]),
Define("COL_COUNT", shape[1]),
Define("DIMENSIONS", discr.dimensions),
Line(),
Typedef(POD(dtype, "value_type")),
Typedef(POD(to_uncomplex_dtype(dtype), "uncomplex_type")),
])
fdecl = FunctionDeclaration(
Value("void", "diff"),
[
Const(Reference(Value("uniform_element_ranges", "from_ers"))),
Const(Reference(Value("uniform_element_ranges", "to_ers"))),
Value("numpy_array<value_type>", "field")
]+[
Value("ublas::matrix<uncomplex_type>", "diffmat_rst%d" % rst)
for rst in range(discr.dimensions)
]+[
Value("numpy_array<value_type>", "result%d" % i)
for i in range(discr.dimensions)
]
)
# }}}
# {{{ set-up
def make_it(name, is_const=True, tpname="value_type"):
if is_const:
const = "const_"
else:
const = ""
return Initializer(
Value("numpy_array<%s>::%siterator" % (tpname, const), name+"_it"),
"%s.begin()" % name)
fbody = Block([
If("ROW_COUNT != diffmat_rst%d.size1()" % i,
S('throw(std::runtime_error("unexpected matrix size"))'))
for i in range(discr.dimensions)
] + [
If("COL_COUNT != diffmat_rst%d.size2()" % i,
S('throw(std::runtime_error("unexpected matrix size"))'))
for i in range(discr.dimensions)
]+[
If("ROW_COUNT != to_ers.el_size()",
S('throw(std::runtime_error("unsupported image element size"))')),
If("COL_COUNT != from_ers.el_size()",
S('throw(std::runtime_error("unsupported preimage element size"))')),
If("from_ers.size() != to_ers.size()",
S('throw(std::runtime_error("image and preimage element groups '
'do nothave the same element count"))')),
Line(),
make_it("field"),
]+[
make_it("result%d" % i, is_const=False)
for i in range(discr.dimensions)
]+[
Line(),
# }}}
# {{{ computation
For("element_number_t eg_el_nr = 0",
"eg_el_nr < to_ers.size()",
"++eg_el_nr",
Block([
Initializer(
Value("node_number_t", "from_el_base"),
"from_ers.start() + eg_el_nr*COL_COUNT"),
Initializer(
Value("node_number_t", "to_el_base"),
"to_ers.start() + eg_el_nr*ROW_COUNT"),
Line(),
For("unsigned i = 0",
"i < ROW_COUNT",
"++i",
Block([
Initializer(Value("value_type", "drst_%d" % rst), 0)
for rst in range(discr.dimensions)
]+[
Line(),
]+[
For("unsigned j = 0",
"j < COL_COUNT",
"++j",
Block([
S("drst_%(rst)d += "
"diffmat_rst%(rst)d(i, j)*field_it[from_el_base+j]"
% {"rst":rst})
for rst in range(discr.dimensions)
])
),
Line(),
]+[
Assign("result%d_it[to_el_base+i]" % rst,
"drst_%d" % rst)
for rst in range(discr.dimensions)
])
)
])
)
])
# }}}
# {{{ compilation
mod.add_function(FunctionBody(fdecl, fbody))
#print "----------------------------------------------------------------"
#print mod.generate()
#raw_input()
compiled_func = mod.compile(self.discr.toolchain).diff
if self.discr.instrumented:
from hedge.tools import time_count_flop
compiled_func = time_count_flop(compiled_func,
discr.diff_timer, discr.diff_counter,
discr.diff_flop_counter,
flops=discr.dimensions*(
2 # mul+add
* ldis.node_count() * len(elgroup.members)
* ldis.node_count()
+
2 * discr.dimensions
* len(elgroup.members) * ldis.node_count()),
increment=discr.dimensions)
return compiled_func