def get_function_definition(self, codegen_state, codegen_result,
schedule_index,
function_decl, function_body):
kernel = codegen_state.kernel
from cgen import (
FunctionBody,
# Post-mid-2016 cgens have 'Collection', too.
Module as Collection,
Initializer,
Line)
result = []
from loopy.kernel.data import AddressSpace
from loopy.schedule import CallKernel
# We only need to write declarations for global variables with
# the first device program. `is_first_dev_prog` determines
# whether this is the first device program in the schedule.
is_first_dev_prog = codegen_state.is_generating_device_code
for i in range(schedule_index):
if isinstance(kernel.schedule[i], CallKernel):
is_first_dev_prog = False
break
if is_first_dev_prog:
for tv in sorted(
six.itervalues(kernel.temporary_variables),
key=lambda tv: tv.name):
if tv.address_space == AddressSpace.GLOBAL and (
tv.initializer is not None):
assert tv.read_only
decl_info, = tv.decl_info(self.target,
index_dtype=kernel.index_dtype)
decl = self.wrap_global_constant(
self.get_temporary_decl(
codegen_state, schedule_index, tv,
decl_info))
if tv.initializer is not None:
decl = Initializer(decl, generate_array_literal(
codegen_state, tv, tv.initializer))
result.append(decl)
fbody = FunctionBody(function_decl, function_body)
if not result:
return fbody
else:
return Collection(result+[Line(), fbody])
def get_function_definition(self, codegen_state, codegen_result,
schedule_index,
function_decl, function_body):
kernel = codegen_state.kernel
from cgen import (
FunctionBody,
# Post-mid-2016 cgens have 'Collection', too.
Module as Collection,
Initializer,
Line)
result = []
from loopy.kernel.data import temp_var_scope
for tv in sorted(
six.itervalues(kernel.temporary_variables),
key=lambda tv: tv.name):
if tv.scope == temp_var_scope.GLOBAL and tv.initializer is not None:
assert tv.read_only
decl_info, = tv.decl_info(self.target,
index_dtype=kernel.index_dtype)
decl = self.wrap_global_constant(
self.get_temporary_decl(
codegen_state, schedule_index, tv,
decl_info))
if tv.initializer is not None:
decl = Initializer(decl, generate_array_literal(
codegen_state, tv, tv.initializer))
result.append(decl)
fbody = FunctionBody(function_decl, function_body)
if not result:
return fbody
else:
return Collection(result+[Line(), fbody])