mindspore 2.4.1__cp310-cp310-win_amd64.whl → 2.5.0__cp310-cp310-win_amd64.whl

mindspore/ops_generate/op_template_parser.py

@@ -0,0 +1,436 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""
+This module defines the OpTemplateParser class for parsing operator prototypes.
+
+The OpTemplateParser class is responsible for converting attributes of OpProto instances into strings
+that can be directly used to fill in code templates. It handles the parsing of argument types,
+function signatures, and generates conversion stubs for PyBoost operations.
+"""
+
+import re
+
+import pyboost_utils
+from pyboost_utils import get_input_dtype, tuple_input_to_cpp_type, get_return_type, \
+    number_input_to_cpp_type, get_const_number_convert, get_tuple_input_convert, is_optional_param
+import gen_constants as K
+from template import Template
+from op_proto import OpProto
+
+
+class OpTemplateParser:
+    """
+    Parses operator prototypes and generates template-compatible strings.
+
+    This class converts the attributes of an OpProto instance into the strings needed for code generation
+    in PyBoost operations.
+
+    Attributes:
+        op_proto (OpProto): The operator prototype containing the relevant information.
+    """
+
+    def __init__(self, op_proto: OpProto):
+        self.op_proto = op_proto
+
+    def _parse_call_args_types(self, op_args):
+        """
+        Parses the data types of the call arguments for the operator.
+
+        Args:
+            op_args (list): A list of operator arguments.
+
+        Returns:
+            list: A list of data types for the call arguments.
+        """
+        call_args_types = []
+        for op_arg in op_args:
+            is_optional = is_optional_param(op_arg)
+            call_args_types.append(get_input_dtype(op_arg.arg_dtype, is_optional))
+        return call_args_types
+
+    def parse_call_args_with_types(self):
+        """
+        Parses the original call arguments and their types for the operator.
+
+        Returns:
+            list: A list of formatted strings representing the call arguments with their types.
+        """
+        call_args = self.parse_original_call_args(self.op_proto.op_args)
+        call_args_types = self._parse_call_args_types(self.op_proto.op_args)
+        call_args_with_types = []
+        for type_name, arg_name in zip(call_args_types, call_args):
+            call_args_with_types.append("const " + type_name + " &" + arg_name)
+        return call_args_with_types
+
+    def parse_need_malloc_tensors(self):
+        """
+        Parses the operator arguments to identify which tensors require memory allocation.
+
+        Returns:
+            tuple: A tuple containing:
+                - need_malloc_tensors (list): Arguments that need memory allocation.
+                - tensor_list_convert (list): Conversions needed for tensor lists.
+                - call_args_with_tensor (list): The call arguments formatted for tensors.
+        """
+        need_malloc_tensors = []
+        tensor_list_convert = []
+        call_args_with_tensor = []
+        call_args = self.parse_original_call_args(self.op_proto.op_args)
+        for op_arg, call_arg in zip(self.op_proto.op_args, call_args):
+            if pyboost_utils.is_tensor(op_arg):
+                call_arg = op_arg.arg_name + "_tensor"
+                need_malloc_tensors.append(call_arg)
+                call_args_with_tensor.append(call_arg)
+            elif tuple_input_to_cpp_type(op_arg.arg_dtype) and pyboost_utils.is_tensor_list(op_arg):
+                need_malloc_tensors.append(call_arg + "_vector")
+                tensor_list_convert.append(get_tuple_input_convert(call_arg, op_arg.arg_dtype))
+                call_args_with_tensor.append(call_arg + "_vector")
+            else:
+                call_args_with_tensor.append(call_arg)
+        return need_malloc_tensors, tensor_list_convert, call_args_with_tensor
+
+    def parse_original_call_args(self, op_args):
+        """
+        Parses the original call arguments from the operator prototype.
+
+        Args:
+            op_args (list): A list of operator arguments.
+
+        Returns:
+            list: A list of formatted strings representing the original call arguments.
+        """
+        call_args = []
+        for op_arg in op_args:
+            if pyboost_utils.is_tensor(op_arg):
+                call_arg = op_arg.arg_name + "_tensor"
+            elif pyboost_utils.is_tensor_list(op_arg):
+                call_arg = op_arg.arg_name + "_tensor_list"
+            else:
+                call_arg = op_arg.arg_name
+            call_args.append(call_arg)
+        return call_args
+
+    def op_args_converter(self):
+        """
+        Converts operator arguments to the corresponding C++ data types.
+
+        Returns:
+            tuple: A tuple containing:
+                - call_args_after_convert (list): The converted call arguments.
+                - value_tuple_convert (list): Conversions needed for value tuples.
+                - const_number_convert (list): Conversions needed for constant numbers.
+        """
+        call_args_after_convert = []
+        value_tuple_convert = []
+        const_number_convert = []
+        call_args = self.parse_original_call_args(self.op_proto.op_args)
+        for op_arg, call_arg in zip(self.op_proto.op_args, call_args):
+            if number_input_to_cpp_type(op_arg.arg_dtype):
+                call_args_after_convert.append(call_arg + "_imm")
+                const_number_convert.append(get_const_number_convert(call_arg, op_arg))
+            elif tuple_input_to_cpp_type(op_arg.arg_dtype):
+                call_args_after_convert.append(call_arg + "_vector")
+                value_tuple_convert.append(get_tuple_input_convert(call_arg, op_arg.arg_dtype))
+            else:
+                call_args_after_convert.append(call_arg)
+        if const_number_convert:
+            const_number_convert.insert(0, '// Convert ValuePtr to c++ scalar\n')
+        if value_tuple_convert:
+            value_tuple_convert.insert(0, '// ValueTuple to std::vector\n')
+        return call_args_after_convert, value_tuple_convert, const_number_convert
+
+    def get_pyboost_func_name(self):
+        """
+        Gets the PyBoost function name based on the operator's class name.
+
+        Returns:
+            str: The generated PyBoost function name.
+        """
+        return "Pyboost_" + self.op_proto.op_class.name
+
+    def get_pyboost_name(self):
+        """
+        Gets the PyBoost name for the operator.
+
+        Returns:
+            str: The generated PyBoost name for the operator.
+        """
+        return "pyboost_" + self.op_proto.op_name
+
+    def get_op_def_name_str(self):
+        """
+        Gets the operator definition name string.
+
+        Returns:
+            str: The generated operator definition name string.
+        """
+        return "g" + self.op_proto.op_class.name
+
+    def gen_signature_same_type_table(self):
+        """
+        Generates a signature table for arguments of the same type.
+
+        Returns:
+            tuple: A tuple containing:
+                - type_num (int): The number of argument types.
+                - signature_table (str): The generated signature table as a string.
+        """
+        signature_table = ''
+        type_num = 0
+        args_signature = self.op_proto.op_args_signature
+        if args_signature is not None:
+            dtype_group = args_signature.dtype_group
+            indexes = {arg.arg_name: index for index, arg in enumerate(self.op_proto.op_args)}
+            if dtype_group is not None:
+                match = re.findall(r'\((.*?)\)', dtype_group)
+                for item in match:
+                    name_args = item.replace(' ', '').split(",")
+                    signature_table += '{'
+                    for arg in name_args:
+                        arg_index = indexes[arg]
+                        signature_table += f"""{arg_index}, """
+                    signature_table = signature_table[:-2]
+                    signature_table += '}, '
+                    type_num += 1
+                signature_table = signature_table[:-2]
+        return type_num, signature_table
+
+    def get_call_args_tensor(self):
+        """
+        Retrieves the call arguments that are of tensor type.
+
+        Returns:
+            list: A list of call arguments that are tensors.
+        """
+        call_args_tensor = []
+        call_args_types = self._parse_call_args_types(self.op_proto.op_args)
+        call_args = self.parse_original_call_args(self.op_proto.op_args)
+        for _type, arg_name in zip(call_args_types, call_args):
+            if _type in ("mindspore::tensor::BaseTensorPtr", "std::optional<mindspore::tensor::BaseTensorPtr>"):
+                call_args_tensor.append(arg_name)
+        return call_args_tensor
+
+    def has_prim_init(self):
+        """
+        Checks if any arguments require primitive initialization.
+
+        Returns:
+            bool: True if any argument requires primitive initialization, otherwise False.
+        """
+        op_args = self.op_proto.op_args
+        has_prim_init = False
+        for op_arg in op_args:
+            prim_init = op_arg.is_prim_init
+            if prim_init:
+                has_prim_init = True
+                break
+        return has_prim_init
+
+    def generate_pyboost_op_func_return_type(self):
+        """
+        Generates the C++ return type for the PyBoost operator function.
+
+        Returns:
+            str: The generated C++ return type for the function.
+
+        Raises:
+            Exception: If no valid return type is found.
+        """
+        returns_type = []
+        type_convert_to_base = {
+            'std::vector<mindspore::tensor::TensorPtr>': 'std::vector<mindspore::tensor::BaseTensorPtr>',
+            'mindspore::tensor::TensorPtr': 'mindspore::tensor::BaseTensorPtr'
+        }
+        for return_obj in self.op_proto.op_returns:
+            temp_return = get_return_type(return_obj.arg_dtype)
+            if temp_return in type_convert_to_base:
+                returns_type.append(type_convert_to_base[temp_return])
+            else:
+                raise Exception("Not return found")
+        if len(returns_type) == 1:
+            cpp_func_return = returns_type[0]
+        elif len(returns_type) > 1:
+            cpp_func_return = "std::tuple<"
+            cpp_func_return += ','.join(s for s in returns_type)
+            cpp_func_return += ">"
+        else:
+            raise Exception("Not return found")
+        return cpp_func_return
+
+    def generate_pyboost_outputs(self):
+        """
+        Generates the output variables for the PyBoost operator function.
+
+        Returns:
+            tuple: A tuple containing:
+                - op_outputs (str): The output variable representation for the operator.
+                - call_outputs (str): The call output variable representation for the operator.
+        """
+        op_outputs = ''
+        call_outputs = ''
+        returns_type = []
+        for return_obj in self.op_proto.op_returns:
+            returns_type.append(get_return_type(return_obj.arg_dtype))
+
+        if len(returns_type) == 1:
+            if returns_type[0] == 'mindspore::tensor::TensorPtr':
+                op_outputs = 'outputs[0]'
+                call_outputs = 'outputs_[0]'
+            elif returns_type[0] == "std::vector<mindspore::tensor::TensorPtr>":
+                op_outputs = 'outputs'
+                call_outputs = 'outputs_'
+            else:
+                raise Exception("Not support return type {}".format(returns_type[0]))
+        elif len(returns_type) > 1:
+            outputs_str = ''
+            for i in range(len(returns_type)):
+                outputs_str += 'outputs[{}],'.format(i)
+            op_outputs = outputs_str[:-1]
+
+            outputs_str = ''
+            for i in range(len(returns_type)):
+                outputs_str += 'outputs_[{}],'.format(i)
+            outputs_str = outputs_str[:-1]
+            call_outputs = "std::make_tuple(" + outputs_str + ")"
+
+        return op_outputs, call_outputs
+
+    def _is_input_arg(self, arg_name, op_name):
+        res = False
+        if op_name in K.INPUT_NAME_MAP and arg_name == K.INPUT_NAME_MAP[op_name]:
+            res = True
+        elif op_name not in K.INPUT_NAME_MAP and arg_name in K.INPUT_ARGS_NAME:
+            res = True
+        return res
+
+    def generate_signature_str(self, kw_only_args=None, varargs=None, *, is_tensor_api: bool) -> str:
+        """
+        Generates a single function signature string for the given operation prototype.
+
+        Args:
+            kw_only_args (list[str]): List of keyword-only argument names.
+            varargs (list[str]): List of variable args names.
+
+        Kwargs:
+            is_tensor_api (bool): Whether this function is used in the Tensor API scenario.
+
+        Returns:
+            str: Generated function signature string.
+        """
+
+        op_name = self.op_proto.op_class.name
+        args_str = f'"{op_name}('
+        first_arg = True
+        kw_args_init_flag = False
+
+        arg_index = 0
+        for arg in self.op_proto.op_args:
+            arg_name = arg.arg_name
+
+            if is_tensor_api and self._is_input_arg(arg_name, op_name):
+                continue
+
+            single_arg = ''
+            if not first_arg:
+                single_arg = ', '
+
+            arg_handler = arg.arg_handler
+            if arg_handler:
+                if arg_handler in K.ARG_HANDLER_MAP:
+                    arg_dtype = K.ARG_HANDLER_MAP[arg_handler]
+                else:
+                    raise ValueError(
+                        f"Generate failed. Check if {arg_handler} is registered in TensorFuncRegCppGenerator.")
+            else:
+                arg_dtype = arg.arg_dtype
+                for cast_type in arg.type_cast:
+                    arg_dtype += f'|{cast_type}'
+
+            # handle varargs params
+            if varargs and arg_name in varargs and arg_index == 0:
+                single_arg += f"{arg_dtype} *{arg_name}"
+            else:
+                single_arg += f"{arg_dtype} {arg_name}"
+
+            if arg.as_init_arg:
+                single_arg += f"={arg.default}"
+
+            # handle keyword-only params
+            if kw_only_args and not kw_args_init_flag and arg_name == kw_only_args[0]:
+                single_arg = ("*, " if first_arg else ", *") + single_arg
+                kw_args_init_flag = True
+
+            args_str += single_arg
+            first_arg = False
+            arg_index += 1
+
+        return args_str + ')"'
+
+    def get_arg_handler_processor(self, func_name, op_proto, *, is_tensor_api):
+        """
+        Generates argument handler processing code for the given function prototype.
+
+        Args:
+            func_name (str): The name of the function.
+            op_proto (OpProto): Operator prototype instance to generate argument processing for.
+
+        Returns:
+            str: Generated argument handler processing code.
+        """
+        tensor_arg_handler_prt_template = Template(
+            "arg_list[${idx}] = "
+            "(*pynative::${func_str}(\"${func_name}\", \"${op_arg_name}\", arg_list[${idx}]))->value();\n"
+        )
+        function_arg_handler_prt_template = Template(
+            "arg_list[${idx}] = "
+            "(*${func_str}(\"${func_name}\", \"${op_arg_name}\", arg_list[${idx}]))->value();\n"
+        )
+        arg_handler_prt_template = (
+            tensor_arg_handler_prt_template) if is_tensor_api else function_arg_handler_prt_template
+
+        arg_handler_template = Template(
+            "arg_list[${idx}] = "
+            "pynative::${func_str}(\"${func_name}\", \"${op_arg_name}\", arg_list[${idx}]);\n"
+        )
+        arg_handler_optional_template = Template(
+            'if (!py::isinstance<py::none>(arg_list[${idx}])) {\n'
+            '  ${arg_handler_str}\n'
+            '}\n'
+        )
+
+        arg_handler_processor = []
+        op_args = op_proto.op_args
+        for idx, op_arg in enumerate(op_args):
+            arg_handler = op_arg.arg_handler
+            func_str = ''.join(word.capitalize() for word in arg_handler.split('_'))
+            if arg_handler:
+                op_arg_name = op_arg.arg_name
+                if func_str in ("StrToEnum", "DtypeToTypeId"):
+                    arg_handler_str = arg_handler_prt_template.replace(func_str=func_str,
+                                                                       func_name=func_name,
+                                                                       op_arg_name=op_arg_name,
+                                                                       idx=idx)
+                else:
+                    arg_handler_str = arg_handler_template.replace(func_str=func_str,
+                                                                   func_name=func_name,
+                                                                   op_arg_name=op_arg_name,
+                                                                   idx=idx)
+
+                if op_arg.default == "None":
+                    arg_handler_str = arg_handler_optional_template.replace(idx=idx,
+                                                                            arg_handler_str=arg_handler_str)
+                arg_handler_processor.append(arg_handler_str)
+
+        return arg_handler_processor