def make_kernels(self, seq_dependencies):
result = []
for sub in self.kernels:
# {{{ figure out arguments
kernel_data = []
for arg_name in sub.arg_names:
dims = sub.dim_map.get(arg_name)
if dims is not None:
# default order is set to "F" in kernel creation below
kernel_data.append(
lp.GlobalArg(
arg_name,
dtype=sub.get_type(arg_name),
shape=sub.get_loopy_shape(arg_name),
))
else:
kernel_data.append(
lp.ValueArg(arg_name,
dtype=sub.get_type(arg_name)))
# }}}
# {{{ figure out temporary variables
for var_name in (
sub.known_names()
- set(sub.arg_names)
- sub.all_inames()):
dtype = sub.get_type(var_name, none_ok=True)
if sub.implicit_types is None and dtype is None:
continue
kernel_data.append(
lp.TemporaryVariable(
var_name, dtype=dtype,
shape=sub.get_loopy_shape(var_name)))
# }}}
knl = lp.make_kernel(
sub.index_sets,
sub.instructions,
kernel_data,
name=sub.subprogram_name,
default_order="F",
index_dtype=self.index_dtype,
target=self.target,
seq_dependencies=seq_dependencies,
)
from loopy.loop import fuse_loop_domains
knl = fuse_loop_domains(knl)
knl = lp.fold_constants(knl)
result.append(knl)
return result
def make_kernels(self):
result = []
for sub in self.kernels:
# {{{ figure out arguments
kernel_data = []
for arg_name in sub.arg_names:
dims = sub.dim_map.get(arg_name)
if dims is not None:
# default order is set to "F" in kernel creation below
kernel_data.append(
lp.GlobalArg(
arg_name,
dtype=sub.get_type(arg_name),
shape=sub.get_loopy_shape(arg_name),
))
else:
kernel_data.append(
lp.ValueArg(arg_name,
dtype=sub.get_type(arg_name)))
# }}}
# {{{ figure out temporary variables
for var_name in (
sub.known_names()
- set(sub.arg_names)
- sub.all_inames()):
dtype = sub.get_type(var_name, none_ok=True)
if sub.implicit_types is None and dtype is None:
continue
kernel_data.append(
lp.TemporaryVariable(
var_name, dtype=dtype,
shape=sub.get_loopy_shape(var_name)))
# }}}
knl = lp.make_kernel(
sub.index_sets,
sub.instructions,
kernel_data,
name=sub.subprogram_name,
default_order="F",
index_dtype=self.index_dtype,
target=self.target,
)
from loopy.loop import fuse_loop_domains
knl = fuse_loop_domains(knl)
knl = lp.fold_constants(knl)
result.append(knl)
return result
def make_kernels(self, seq_dependencies):
result = []
for sub in self.kernels:
# {{{ figure out arguments
kernel_data = []
for arg_name in sub.arg_names:
dims = sub.dim_map.get(arg_name)
if sub.data_map.get(arg_name) is not None:
raise NotImplementedError("initializer for argument %s" %
arg_name)
if dims is not None:
# default order is set to "F" in kernel creation below
kernel_data.append(
lp.GlobalArg(
arg_name,
dtype=sub.get_type(arg_name),
shape=sub.get_loopy_shape(arg_name),
))
else:
kernel_data.append(
lp.ValueArg(arg_name, dtype=sub.get_type(arg_name)))
# }}}
# {{{ figure out temporary variables
for var_name in (sub.known_names() - set(sub.arg_names) -
sub.all_inames()):
dtype = sub.get_type(var_name, none_ok=True)
if sub.implicit_types is None and dtype is None:
continue
kwargs = {}
if sub.data_map.get(var_name) is not None:
kwargs["read_only"] = True
kwargs["address_space"] = lp.AddressSpace.PRIVATE
kwargs["initializer"] = np.array(sub.data_map[var_name],
dtype=dtype)
kernel_data.append(
lp.TemporaryVariable(var_name,
dtype=dtype,
shape=sub.get_loopy_shape(var_name),
**kwargs))
# }}}
knl = lp.make_function(
sub.index_sets,
sub.instructions,
kernel_data,
name=sub.subprogram_name,
default_order="F",
index_dtype=self.index_dtype,
target=self.target,
seq_dependencies=seq_dependencies,
)
from loopy.loop import merge_loop_domains
knl = merge_loop_domains(knl)
knl = lp.fold_constants(knl)
result.append(knl)
return result