def get_count_kernel(self, index_dtype):
index_ctype = dtype_to_ctype(index_dtype)
from pyopencl.tools import VectorArg, OtherArg
kernel_list_args = [
VectorArg(index_dtype, "plb_%s_count" % name)
for name, dtype in self.list_names_and_dtypes
if name not in self.count_sharing]
user_list_args = []
for name, dtype in self.list_names_and_dtypes:
if name in self.count_sharing:
continue
name = "plb_loc_%s_count" % name
user_list_args.append(OtherArg("%s *%s" % (
index_ctype, name), name))
kernel_name = self.name_prefix+"_count"
from pyopencl.characterize import has_double_support
src = _LIST_BUILDER_TEMPLATE.render(
is_count_stage=True,
kernel_name=kernel_name,
double_support=all(has_double_support(dev) for dev in
self.context.devices),
debug=self.debug,
do_not_vectorize=self.do_not_vectorize(),
eliminate_empty_output_lists=self.eliminate_empty_output_lists,
kernel_list_arg_decl=_get_arg_decl(kernel_list_args),
kernel_list_arg_values=_get_arg_list(user_list_args, prefix="&"),
user_list_arg_decl=_get_arg_decl(user_list_args),
user_list_args=_get_arg_list(user_list_args),
user_arg_decl_with_offset=_get_arg_decl(self.arg_decls),
user_arg_decl_no_offset=_get_arg_decl(self.arg_decls_no_offset),
user_args_no_offset=_get_arg_list(self.arg_decls_no_offset),
arg_offset_adjustment=get_arg_offset_adjuster_code(self.arg_decls),
list_names_and_dtypes=self.list_names_and_dtypes,
count_sharing=self.count_sharing,
name_prefix=self.name_prefix,
generate_template=self.generate_template,
preamble=self.preamble,
index_type=index_ctype,
)
src = str(src)
prg = cl.Program(self.context, src).build(self.options)
knl = getattr(prg, kernel_name)
from pyopencl.tools import get_arg_list_scalar_arg_dtypes
knl.set_scalar_arg_dtypes(get_arg_list_scalar_arg_dtypes(
kernel_list_args+self.arg_decls) + [index_dtype])
return knl
def get_write_kernel(self, index_dtype):
index_ctype = dtype_to_ctype(index_dtype)
from pyopencl.tools import VectorArg, OtherArg
kernel_list_args = []
kernel_list_arg_values = ""
user_list_args = []
for name, dtype in self.list_names_and_dtypes:
list_name = "plb_%s_list" % name
list_arg = VectorArg(dtype, list_name)
kernel_list_args.append(list_arg)
user_list_args.append(list_arg)
if name in self.count_sharing:
kernel_list_arg_values += "%s, " % list_name
continue
kernel_list_args.append(
VectorArg(index_dtype, "plb_%s_start_index" % name))
index_name = "plb_%s_index" % name
user_list_args.append(
OtherArg("%s *%s" % (index_ctype, index_name), index_name))
kernel_list_arg_values += "%s, &%s, " % (list_name, index_name)
kernel_name = self.name_prefix + "_write"
from pyopencl.characterize import has_double_support
src = _LIST_BUILDER_TEMPLATE.render(
is_count_stage=False,
kernel_name=kernel_name,
double_support=all(
has_double_support(dev) for dev in self.context.devices),
debug=self.debug,
do_not_vectorize=self.do_not_vectorize(),
kernel_list_arg_decl=_get_arg_decl(kernel_list_args),
kernel_list_arg_values=kernel_list_arg_values,
user_list_arg_decl=_get_arg_decl(user_list_args),
user_list_args=_get_arg_list(user_list_args),
user_arg_decl=_get_arg_decl(self.arg_decls),
user_args=_get_arg_list(self.arg_decls),
list_names_and_dtypes=self.list_names_and_dtypes,
count_sharing=self.count_sharing,
name_prefix=self.name_prefix,
generate_template=self.generate_template,
preamble=self.preamble,
index_type=index_ctype,
)
src = str(src)
prg = cl.Program(self.context, src).build(self.options)
knl = getattr(prg, kernel_name)
from pyopencl.tools import get_arg_list_scalar_arg_dtypes
knl.set_scalar_arg_dtypes(
get_arg_list_scalar_arg_dtypes(kernel_list_args + self.arg_decls) +
[index_dtype])
return knl