def generate_body(self, kernel, codegen_state):
from cgen import Block
body = Block()
# {{{ declare temporaries
body.extend(
idi.cgen_declarator
for tv in six.itervalues(kernel.temporary_variables)
for idi in tv.decl_info(kernel.target, is_written=True, index_dtype=kernel.index_dtype)
)
# }}}
from loopy.codegen.loop import set_up_hw_parallel_loops
gen_code = set_up_hw_parallel_loops(kernel, 0, codegen_state)
from cgen import Line
body.append(Line())
if isinstance(gen_code.ast, Block):
body.extend(gen_code.ast.contents)
else:
body.append(gen_code.ast)
return body, gen_code.implemented_domains
def generate_host_or_device_program(codegen_state, schedule_index):
ast_builder = codegen_state.ast_builder
temp_decls = ast_builder.get_temporary_decls(codegen_state, schedule_index)
from functools import partial
from loopy.codegen.control import build_loop_nest
if codegen_state.is_generating_device_code:
from loopy.schedule import CallKernel
assert isinstance(codegen_state.kernel.linearization[schedule_index],
CallKernel)
from loopy.codegen.loop import set_up_hw_parallel_loops
codegen_result = set_up_hw_parallel_loops(
codegen_state, schedule_index,
next_func=partial(build_loop_nest,
schedule_index=schedule_index + 1))
else:
codegen_result = build_loop_nest(codegen_state, schedule_index)
if (codegen_state.is_generating_device_code
or codegen_state.is_entrypoint):
codegen_result = merge_codegen_results(
codegen_state,
ast_builder.generate_top_of_body(codegen_state)
+ temp_decls
+ [codegen_result],
collapse=False)
cur_prog = codegen_result.current_program(codegen_state)
body_ast = cur_prog.ast
fdecl_ast = ast_builder.get_function_declaration(
codegen_state, codegen_result, schedule_index)
fdef_ast = ast_builder.get_function_definition(
codegen_state, codegen_result,
schedule_index, fdecl_ast, body_ast)
codegen_result = codegen_result.with_new_program(
codegen_state,
cur_prog.copy(
ast=ast_builder.process_ast(fdef_ast),
body_ast=ast_builder.process_ast(body_ast)))
return codegen_result
def generate_host_or_device_program(codegen_state, schedule_index):
ast_builder = codegen_state.ast_builder
temp_decls = ast_builder.get_temporary_decls(codegen_state, schedule_index)
from functools import partial
from loopy.codegen.control import build_loop_nest
if codegen_state.is_generating_device_code:
from loopy.schedule import CallKernel
assert isinstance(codegen_state.kernel.schedule[schedule_index], CallKernel)
from loopy.codegen.loop import set_up_hw_parallel_loops
codegen_result = set_up_hw_parallel_loops(
codegen_state, schedule_index,
next_func=partial(build_loop_nest,
schedule_index=schedule_index + 1))
else:
codegen_result = build_loop_nest(codegen_state, schedule_index)
codegen_result = merge_codegen_results(
codegen_state,
ast_builder.generate_top_of_body(codegen_state)
+ temp_decls
+ [codegen_result],
collapse=False)
cur_prog = codegen_result.current_program(codegen_state)
body_ast = cur_prog.ast
fdecl_ast = ast_builder.get_function_declaration(
codegen_state, codegen_result, schedule_index)
fdef_ast = ast_builder.get_function_definition(
codegen_state, codegen_result,
schedule_index, fdecl_ast, body_ast)
codegen_result = codegen_result.with_new_program(
codegen_state,
cur_prog.copy(
ast=fdef_ast,
body_ast=body_ast))
return codegen_result
def generate_body(self, kernel, codegen_state):
from cgen import Block
body = Block()
temp_decls = []
# {{{ declare temporaries
base_storage_sizes = {}
base_storage_to_is_local = {}
base_storage_to_align_bytes = {}
from cgen import ArrayOf, Pointer, Initializer, AlignedAttribute
from loopy.codegen import POD # uses the correct complex type
class ConstRestrictPointer(Pointer):
def get_decl_pair(self):
sub_tp, sub_decl = self.subdecl.get_decl_pair()
return sub_tp, ("*const restrict %s" % sub_decl)
for tv in sorted(
six.itervalues(kernel.temporary_variables),
key=lambda tv: tv.name):
decl_info = tv.decl_info(self, index_dtype=kernel.index_dtype)
if not tv.base_storage:
for idi in decl_info:
temp_var_decl = POD(self, idi.dtype, idi.name)
if idi.shape:
temp_var_decl = ArrayOf(temp_var_decl,
" * ".join(str(s) for s in idi.shape))
temp_decls.append(
self.wrap_temporary_decl(temp_var_decl, tv.is_local))
else:
offset = 0
base_storage_sizes.setdefault(tv.base_storage, []).append(
tv.nbytes)
base_storage_to_is_local.setdefault(tv.base_storage, []).append(
tv.is_local)
align_size = tv.dtype.itemsize
from loopy.kernel.array import VectorArrayDimTag
for dim_tag, axis_len in zip(tv.dim_tags, tv.shape):
if isinstance(dim_tag, VectorArrayDimTag):
align_size *= axis_len
base_storage_to_align_bytes.setdefault(tv.base_storage, []).append(
align_size)
for idi in decl_info:
cast_decl = POD(self, idi.dtype, "")
temp_var_decl = POD(self, idi.dtype, idi.name)
cast_decl = self.wrap_temporary_decl(cast_decl, tv.is_local)
temp_var_decl = self.wrap_temporary_decl(
temp_var_decl, tv.is_local)
# The 'restrict' part of this is a complete lie--of course
# all these temporaries are aliased. But we're promising to
# not use them to shovel data from one representation to the
# other. That counts, right?
cast_decl = ConstRestrictPointer(cast_decl)
temp_var_decl = ConstRestrictPointer(temp_var_decl)
cast_tp, cast_d = cast_decl.get_decl_pair()
temp_var_decl = Initializer(
temp_var_decl,
"(%s %s) (%s + %s)" % (
" ".join(cast_tp), cast_d,
tv.base_storage,
offset))
temp_decls.append(temp_var_decl)
from pytools import product
offset += (
idi.dtype.itemsize
* product(si for si in idi.shape))
for bs_name, bs_sizes in sorted(six.iteritems(base_storage_sizes)):
bs_var_decl = POD(self, np.int8, bs_name)
bs_var_decl = self.wrap_temporary_decl(
bs_var_decl, base_storage_to_is_local[bs_name])
bs_var_decl = ArrayOf(bs_var_decl, max(bs_sizes))
alignment = max(base_storage_to_align_bytes[bs_name])
bs_var_decl = AlignedAttribute(alignment, bs_var_decl)
body.append(bs_var_decl)
body.extend(temp_decls)
# }}}
from loopy.codegen.loop import set_up_hw_parallel_loops
gen_code = set_up_hw_parallel_loops(kernel, 0, codegen_state)
from cgen import Line
body.append(Line())
if isinstance(gen_code.ast, Block):
body.extend(gen_code.ast.contents)
else:
body.append(gen_code.ast)
return body, gen_code.implemented_domains
