def get_function_declaration(self, codegen_state, codegen_result,
schedule_index):
fdecl = super(OpenCLCASTBuilder,
self).get_function_declaration(codegen_state,
codegen_result,
schedule_index)
from loopy.target.c import FunctionDeclarationWrapper
assert isinstance(fdecl, FunctionDeclarationWrapper)
fdecl = fdecl.subdecl
from cgen.opencl import CLKernel, CLRequiredWorkGroupSize
fdecl = CLKernel(fdecl)
from loopy.schedule import get_insn_ids_for_block_at
_, local_sizes = codegen_state.kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(codegen_state.kernel.schedule,
schedule_index))
from loopy.symbolic import get_dependencies
if not get_dependencies(local_sizes):
# sizes can't have parameter dependencies if they are
# to be used in static WG size.
fdecl = CLRequiredWorkGroupSize(local_sizes, fdecl)
return FunctionDeclarationWrapper(fdecl)
def get_function_declaration(self, codegen_state, codegen_result,
schedule_index):
fdecl = super(CUDACASTBuilder, self).get_function_declaration(
codegen_state, codegen_result, schedule_index)
from loopy.target.c import FunctionDeclarationWrapper
assert isinstance(fdecl, FunctionDeclarationWrapper)
fdecl = fdecl.subdecl
from cgen.cuda import CudaGlobal, CudaLaunchBounds
fdecl = CudaGlobal(fdecl)
if self.target.extern_c:
from cgen import Extern
fdecl = Extern("C", fdecl)
from loopy.schedule import get_insn_ids_for_block_at
_, local_grid_size = \
codegen_state.kernel.get_grid_sizes_for_insn_ids_as_exprs(
get_insn_ids_for_block_at(
codegen_state.kernel.schedule, schedule_index))
from loopy.symbolic import get_dependencies
if not get_dependencies(local_grid_size):
# Sizes can't have parameter dependencies if they are
# to be used in static thread block size.
from pytools import product
nthreads = product(local_grid_size)
fdecl = CudaLaunchBounds(nthreads, fdecl)
return FunctionDeclarationWrapper(fdecl)
def get_function_declaration(self, codegen_state, codegen_result,
schedule_index):
name = codegen_result.current_program(codegen_state).name
from cgen import (FunctionDeclaration, Value)
from cgen.ispc import ISPCExport, ISPCTask
arg_names, arg_decls = self._arg_names_and_decls(codegen_state)
if codegen_state.is_generating_device_code:
result = ISPCTask(
FunctionDeclaration(Value("void", name), arg_decls))
else:
result = ISPCExport(
FunctionDeclaration(Value("void", name), arg_decls))
from loopy.target.c import FunctionDeclarationWrapper
return FunctionDeclarationWrapper(result)
