mindspore 2.4.1__cp310-cp310-manylinux1_x86_64.whl → 2.5.0__cp310-cp310-manylinux1_x86_64.whl

mindspore/ops_generate/pyboost_op_cpp_code_generator.py

@@ -0,0 +1,851 @@
+# 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 several classes and functions for generating C++ code for PyBoost operations,
+including function headers, source files, and registration code. It handles the generation of code
+for different devices (Ascend, CPU, GPU) and manages residual files associated with operator prototypes.
+"""
+
+import os
+import re
+
+from pyboost_utils import is_cube, AclnnUtils, get_return_type, merge_strings_by_chunk_size, is_op_multi_output
+import template
+import gen_constants as K
+from gen_utils import save_file
+from op_proto import OpProto
+from op_template_parser import OpTemplateParser
+from base_generator import BaseGenerator
+
+
+class PyboostCommonOpHeaderGenerator(BaseGenerator):
+    """
+    Generates common C++ headers for PyBoost operations.
+
+    This class processes operator prototypes and generates header files containing function definitions
+    based on templates provided. It specifically generates the headers that define operations for PyBoost.
+    """
+
+    def __init__(self):
+        self.pyboost_op_header_str = template.PYBOOST_BASE_OP_DEFINE_TEMPLATE
+
+    def generate(self, work_path, op_protos):
+        """
+        Generates header files for the provided operator prototypes.
+
+        Args:
+            work_path (str): The directory path where the header files will be saved.
+            op_protos (list): A list of operator prototypes containing information about the operators.
+
+        Returns:
+            None
+        """
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            op_parser = OpTemplateParser(op_proto)
+            op_name_str = op_proto.op_class.name
+            call_args_with_type = op_parser.parse_call_args_with_types()
+            cpp_func_return = _generate_cpp_func_return(op_proto)
+            output_is_tuple = "bool output_is_tuple() const override { return true; }" \
+                if is_op_multi_output(op_proto.op_returns) else ''
+            pyboost_op_header_str = template.PYBOOST_BASE_OP_DEFINE_TEMPLATE.replace(op_name=op_name_str,
+                                                                                     op_name_upper=op_name_str.upper(),
+                                                                                     call_args=call_args_with_type,
+                                                                                     return_type=cpp_func_return,
+                                                                                     output_is_tuple=output_is_tuple)
+            save_path = os.path.join(work_path, f"{K.MS_COMMON_PYBOOST_KERNEL_PATH}/auto_generate/")
+            file_name = f"{op_proto.op_name}.h"
+            save_file(save_path, file_name, pyboost_op_header_str)
+
+
+class PyboostOpHeaderGenerator(BaseGenerator):
+    """
+    Generates device-specific C++ headers for PyBoost operations.
+
+    This class generates header files for different devices (Ascend, CPU, GPU) and defines
+    the operation functions accordingly.
+    """
+
+    def __init__(self, device):
+        """
+        Initializes the PyboostOpHeaderGenerator with the appropriate templates for the specified device.
+
+        Args:
+            device (str): The target device (ascend, gpu, or cpu).
+
+        Raises:
+            ValueError: If the device is not supported.
+        """
+        template_dict = {"ascend": template.PYBOOST_ASCEND_OP_HEADER_TEMPLATE,
+                         "gpu": template.PYBOOST_GPU_OP_HEADER_TEMPLATE,
+                         "cpu": template.PYBOOST_CPU_OP_HEADER_TEMPLATE}
+        if device not in template_dict:
+            raise ValueError(f"Device must be ascend, gpu, or cpu, {device} is not supported")
+        self.PYBOOST_OP_HEADER_TEMPLATE = template_dict[device]
+        self.code_generate_path = f"{K.MS_OPS_KERNEL_PATH}/{device}/pyboost/auto_generate/"
+        self.device = device
+
+    def generate(self, work_path, op_protos):
+        """
+        Generates header files for the provided operator prototypes based on the device.
+
+        Args:
+            work_path (str): The directory path where the header files will be saved.
+            op_protos (list): A list of operator prototypes containing information about the operators.
+
+        Returns:
+            None
+        """
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            if getattr(op_proto.op_dispatch, self.device) == 'None':
+                continue
+            op_parser = OpTemplateParser(op_proto)
+            op_name_str = op_proto.op_class.name
+            call_args_with_type = op_parser.parse_call_args_with_types()
+            cpp_func_return = _generate_cpp_func_return(op_proto)
+
+            pyboost_op_str = self.PYBOOST_OP_HEADER_TEMPLATE.replace(op_name=op_name_str,
+                                                                     op_name_upper=op_name_str.upper(),
+                                                                     operator_name=op_proto.op_name,
+                                                                     call_args_with_type=call_args_with_type,
+                                                                     return_type=cpp_func_return)
+
+            save_path = os.path.join(work_path, self.code_generate_path)
+            file_name = f"{op_proto.op_name}.h"
+            save_file(save_path, file_name, pyboost_op_str)
+
+
+class PyboostOpCppGenerator:
+    """
+    Generates C++ source files for PyBoost operations.
+
+    This class generates the implementation of operations for different devices, handling function calls
+    and registering custom kernels as necessary.
+    """
+
+    def __init__(self, device):
+        """
+        Initializes the PyboostOpCppGenerator with the appropriate templates for the specified device.
+
+        Args:
+            device (str): The target device (ascend, gpu, or cpu).
+
+        Raises:
+            ValueError: If the device is not supported.
+        """
+        if device == 'ascend':
+            PYBOOST_CUSTOMIZE_CALL_TEMPLATE = template.PYBOOST_ASCEND_CUSTOMIZE_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_ASCEND_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_ASCEND_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/ascend/pyboost/auto_generate/"
+        elif device == 'cpu':
+            PYBOOST_CUSTOMIZE_CALL_TEMPLATE = template.PYBOOST_CPU_CUSTOMIZE_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_CPU_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_CPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/cpu/pyboost/auto_generate/"
+        elif device == 'gpu':
+            PYBOOST_CUSTOMIZE_CALL_TEMPLATE = template.PYBOOST_GPU_CUSTOMIZE_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_GPU_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_GPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/gpu/pyboost/auto_generate/"
+        else:
+            raise ValueError(f"Device must be ascend, gpu, or cpu, {device} is not supported")
+        self.PYBOOST_CUSTOMIZE_CALL_TEMPLATE = PYBOOST_CUSTOMIZE_CALL_TEMPLATE
+        self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE = PYBOOST_SINGLE_OP_HEADER_TEMPLATE
+        self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = PYBOOST_SINGLE_OP_SOURCE_TEMPLATE
+        self.gen_path = gen_path
+        self.device = device
+
+    def generate_customize_op_cpp_code(self, op_protos, merge_op_header, merge_op_function):
+        """
+        Generate C++ code for PyBoost operations using the provided operation prototypes.
+
+        This method processes a list of operation prototypes, generates customized function call
+        implementations, and updates the merged headers and functions for the specified device.
+
+        Args:
+            op_protos (list): A list of operation prototypes to process. Each prototype contains
+                              metadata about the operation, including dispatch settings and arguments.
+            merge_op_header (list): A list to store the generated C++ header code for operations.
+            merge_op_function (list): A list to store the generated C++ source code for operations.
+        """
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            if getattr(op_proto.op_dispatch, self.device) == 'default':
+                continue
+            if getattr(op_proto.op_dispatch, self.device) == 'None':
+                continue
+            op_parser = OpTemplateParser(op_proto)
+            call_args = op_parser.parse_original_call_args(op_proto.op_args)
+            call_args_with_type = op_parser.parse_call_args_with_types()
+            _, call_func_outputs = op_parser.generate_pyboost_outputs()
+            operator_name = op_proto.op_name
+            op_name_str = op_proto.op_class.name
+            check_inplace_func = ''
+            for arg in op_proto.op_returns:
+                if arg.inplace != '':
+                    check_inplace_func = f'ThrowExpectionWhenInternalOverlap({arg.inplace}_tensor);'
+                    break
+            call_impl = self.PYBOOST_CUSTOMIZE_CALL_TEMPLATE.replace(
+                call_args=call_args,
+                return_values=call_func_outputs,
+                customize_func=getattr(op_proto.op_dispatch, self.device) + "Customize",
+                check_expression=check_inplace_func,
+            )
+            customize_include = \
+                f'#include "{K.MS_OPS_KERNEL_PATH}/{self.device}/pyboost/customize/{operator_name.lower()}.h"\n'
+            register_custom = self._get_register_custom_kernel(op_proto)
+            cpp_func_return = _generate_cpp_func_return(op_proto)
+            merge_op_header.append(self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE.replace(operator_name=operator_name,
+                                                                                  customize_include=customize_include))
+            merge_op_function.append(
+                self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE.replace(op_name=op_name_str,
+                                                               call_args_with_type=call_args_with_type,
+                                                               return_type=cpp_func_return, call_impl=call_impl,
+                                                               register_custom_kernel=register_custom))
+
+    def _get_register_custom_kernel(self, op_proto: OpProto):
+        """
+        Generates the registration code for custom kernels based on the device.
+
+        Args:
+            op_proto (OpProto): The operator prototype to generate registration for.
+
+        Returns:
+            str: The registration code for the custom kernel.
+        """
+        if self.device == 'ascend':
+            register_custom_kernel = ''
+        elif self.device == 'cpu':
+            register_custom_kernel = f"MS_REG_PYBOOST_CPU_CUSTOM_KERNEL({op_proto.op_class.name});"
+        elif self.device == 'gpu':
+            register_custom_kernel = f"MS_REG_PYBOOST_GPU_CUSTOM_KERNEL({op_proto.op_class.name});"
+        else:
+            raise ValueError(f"Device must be ascend, gpu, or cpu, {self.device} is not supported")
+        return register_custom_kernel
+
+
+class PyboostViewOpCppGenerator:
+    """
+    Generates C++ source files for view operations in PyBoost.
+
+    This class handles the generation of source files for view operations, which have special handling
+    compared to regular operations.
+    """
+
+    def __init__(self, device):
+        """
+        Initializes the PyboostViewOpCppGenerator with the appropriate templates for the specified device.
+
+        Args:
+            device (str): The target device (ascend, gpu, or cpu).
+
+        Raises:
+            ValueError: If the device is not supported.
+        """
+        if device == 'ascend':
+            PYBOOST_VIEW_CALL_TEMPLATE = template.PYBOOST_ASCEND_VIEW_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_ASCEND_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_ASCEND_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/ascend/pyboost/auto_generate/"
+        elif device == 'cpu':
+            PYBOOST_VIEW_CALL_TEMPLATE = template.PYBOOST_CPU_VIEW_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_CPU_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_CPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/cpu/pyboost/auto_generate/"
+        elif device == 'gpu':
+            PYBOOST_VIEW_CALL_TEMPLATE = template.PYBOOST_GPU_VIEW_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.PYBOOST_GPU_SINGLE_OP_HEADER_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_GPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/gpu/pyboost/auto_generate/"
+        else:
+            raise ValueError(f"Device must be ascend, gpu, or cpu, {device} is not supported")
+        self.PYBOOST_VIEW_CALL_TEMPLATE = PYBOOST_VIEW_CALL_TEMPLATE
+        self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE = PYBOOST_SINGLE_OP_HEADER_TEMPLATE
+        self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = PYBOOST_SINGLE_OP_SOURCE_TEMPLATE
+        self.gen_path = gen_path
+        self.device = device
+
+    def generate_view_op_cpp_code(self, op_protos, merge_op_header, merge_op_function):
+        """
+        Generate C++ code for view operations in PyBoost.
+
+        This method processes a list of operation prototypes (`op_protos`) and generates C++ code
+        for view operations where `op_view` is set to `True` and the dispatch setting for the target
+        device is `'default'`.
+
+        Args:
+            op_protos (list): A list of operation prototypes to process. Each prototype includes
+                              metadata such as dispatch settings, arguments, and view-specific attributes.
+            merge_op_header (list): A list to store the generated C++ header code for view operations.
+            merge_op_function (list): A list to store the generated C++ source code for view operations.
+        """
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            if getattr(op_proto.op_dispatch, self.device) != 'default':
+                continue
+            if getattr(op_proto.op_dispatch, self.device) == 'None':
+                continue
+            if not op_proto.op_view:
+                continue
+
+            op_parser = OpTemplateParser(op_proto)
+            call_args_tensor = op_parser.get_call_args_tensor()
+            call_args = op_parser.parse_original_call_args(op_proto.op_args)
+            call_args_with_type = op_parser.parse_call_args_with_types()
+            _, call_func_outputs = op_parser.generate_pyboost_outputs()
+            call_impl = self.PYBOOST_VIEW_CALL_TEMPLATE.replace(op_name=op_proto.op_class.name,
+                                                                call_args=call_args,
+                                                                call_tensors=call_args_tensor,
+                                                                return_values=call_func_outputs,
+                                                                input=call_args[0])
+            customize_include = f'#include "{K.MS_OPS_VIEW_PATH}/{op_proto.op_name}_strides_calc.h"\n'
+            cpp_func_return = _generate_cpp_func_return(op_proto)
+            merge_op_header.append(self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE.replace(operator_name=op_proto.op_name,
+                                                                                  customize_include=customize_include))
+
+            merge_op_function.append(
+                self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE.replace(op_name=op_proto.op_class.name,
+                                                               call_args_with_type=call_args_with_type,
+                                                               return_type=cpp_func_return,
+                                                               call_impl=call_impl,
+                                                               register_custom_kernel=""))
+
+
+class AclnnOpCppCodeGenerator:
+    """
+    Generates C++ source files for ACLNN operations in PyBoost.
+
+    This class handles the generation of source files for operations that utilize the ACLNN framework,
+    including customized calls and tensor management.
+
+    Attributes:
+        PYBOOST_CALL_TEMPLATE (Template): Template for generating ACLNN operation calls.
+        PYBOOST_OP_SOURCE_TEMPLATE (Template): Template for generating operation source files.
+        gen_path (str): Path for saving the generated C++ source files.
+        device (str): The target device (ascend, cpu, or gpu).
+    """
+
+    def __init__(self, device):
+        """
+        Initializes the AclnnOpCppCodeGenerator with the appropriate templates for the specified device.
+
+        Args:
+            device (str): The target device (ascend, gpu, or cpu).
+
+        Raises:
+            ValueError: If the device is not supported.
+        """
+        if device == 'ascend':
+            PYBOOST_CALL_TEMPLATE = template.PYBOOST_ASCEND_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_ASCEND_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/ascend/pyboost/auto_generate/"
+        elif device == 'cpu':
+            PYBOOST_CALL_TEMPLATE = template.PYBOOST_CPU_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_CPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/cpu/pyboost/auto_generate/"
+        elif device == 'gpu':
+            PYBOOST_CALL_TEMPLATE = template.PYBOOST_GPU_CALL_TEMPLATE
+            PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = template.PYBOOST_GPU_SINGLE_OP_SOURCE_TEMPLATE
+            gen_path = f"{K.MS_OPS_KERNEL_PATH}/gpu/pyboost/auto_generate/"
+        else:
+            raise ValueError(f"Device must be ascend, gpu, or cpu, {device} is not supported")
+        self.PYBOOST_CALL_TEMPLATE = PYBOOST_CALL_TEMPLATE
+        self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE = template.Template(
+            '#include "kernel/${device}/pyboost/auto_generate/${operator_name}.h"\n'
+        )
+        self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE = PYBOOST_SINGLE_OP_SOURCE_TEMPLATE
+        self.gen_path = gen_path
+        self.device = device
+
+    def generate_aclnn_op_cpp_code(self, op_protos, merge_op_header, merge_op_function):
+        """
+        Generate C++ code for ACLNN operations in PyBoost.
+
+        This method processes a list of operation prototypes (`op_protos`) and generates C++ code
+        for aclnn operations. The method filters the operation
+        prototypes based on their dispatch and view settings, and then uses templates and metadata
+        to generate the necessary implementation and header files.
+
+        Args:
+            op_protos (list): A list of operation prototypes. Each prototype includes metadata
+                              such as operation name, dispatch settings, view attributes, and arguments.
+            merge_op_header (list): A list to store the generated C++ header code for ACLNN operations.
+            merge_op_function (list): A list to store the generated C++ source code for ACLNN operations.
+        """
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            if getattr(op_proto.op_dispatch, self.device) != 'default':
+                continue
+            if getattr(op_proto.op_dispatch, self.device) == 'None':
+                continue
+            if op_proto.op_view:
+                continue
+
+            op_parser = OpTemplateParser(op_proto)
+            aclnn_name = AclnnUtils.get_aclnn_interface(op_proto.op_class.name)
+
+            call_args_tensor = op_parser.get_call_args_tensor()
+            create_input_address = self._generate_create_input_address(op_parser)
+            malloc_inputs = self._generate_malloc_input(op_parser)
+            op_outputs, call_func_outputs = op_parser.generate_pyboost_outputs()
+            get_inputs_kernel_tensors = self._generate_get_inputs_kernel_tensors(op_parser)
+
+            cube_math_type, get_cube_math_type = '', ''
+            if self.device == 'ascend' and is_cube(op_proto.op_class.name):
+                get_cube_math_type = f'// cubeMathType: 0 - KEEP_DTYPE, 1 - ALLOW_FP32_DOWN_PRECISION\n'
+                get_cube_math_type += "auto cube_math_type = GetCubeMathType();"
+                cube_math_type = ', cube_math_type'
+
+            real_output = ', ' + op_outputs \
+                if _generate_inplace_process_cpp_code(op_proto) == '' else ''
+
+            cast_input_code, real_call_args_tensor = self._generate_tensor_cpu_cast_input_code(
+                op_parser)
+            cpp_func_return = _generate_cpp_func_return(op_proto)
+            _, tensor_list_convert, call_args_with_tensor = op_parser.parse_need_malloc_tensors()
+            call_args_after_convert, value_tuple_convert, const_number_convert = op_parser.op_args_converter()
+            call_args = op_parser.parse_original_call_args(op_proto.op_args)
+            call_args_with_type = op_parser.parse_call_args_with_types()
+            inplace_process = _generate_inplace_process_cpp_code(op_proto)
+            call_impl = self.PYBOOST_CALL_TEMPLATE.replace(aclnn_name=aclnn_name,
+                                                           call_args=call_args,
+                                                           call_tensors=call_args_tensor,
+                                                           value_tuple_convert=value_tuple_convert,
+                                                           const_number_convert=const_number_convert,
+                                                           create_input_address=create_input_address,
+                                                           tensor_list_convert=tensor_list_convert,
+                                                           call_args_with_tensor=call_args_with_tensor,
+                                                           malloc_inputs=malloc_inputs,
+                                                           get_inputs_kernel_tensors=get_inputs_kernel_tensors,
+                                                           get_cube_math_type=get_cube_math_type,
+                                                           cube_math_type=cube_math_type,
+                                                           real_call_args=call_args_after_convert,
+                                                           return_values=call_func_outputs,
+                                                           outputs=real_output,
+                                                           inplace_process=inplace_process,
+                                                           cast_input_code=cast_input_code,
+                                                           real_call_args_tensor=real_call_args_tensor,
+                                                           class_name=op_proto.op_class.name,
+                                                           op_name_str=op_proto.op_class.name)
+
+            merge_op_header.append(self.PYBOOST_SINGLE_OP_HEADER_TEMPLATE.replace(operator_name=op_proto.op_name,
+                                                                                  device=self.device))
+
+            merge_op_function.append(
+                self.PYBOOST_SINGLE_OP_SOURCE_TEMPLATE.replace(op_name=op_proto.op_class.name,
+                                                               call_args_with_type=call_args_with_type,
+                                                               return_type=cpp_func_return,
+                                                               call_impl=call_impl,
+                                                               register_custom_kernel=''))
+
+    def _generate_tensor_cpu_cast_input_code(self, op_parser: OpTemplateParser):
+        """
+        Generates the input casting code for CPU tensor operations.
+
+        Args:
+            op_parser (OpTemplateParser): The parser object for the operation prototype.
+
+        Returns:
+            tuple: A tuple containing the casting code and the updated tensor call arguments.
+        """
+        _, _, call_args_with_tensor = op_parser.parse_need_malloc_tensors()
+        call_tensors = op_parser.get_call_args_tensor()
+        cast_input = ""
+        real_call_args_tensor = call_args_with_tensor.copy()
+        for i, tensor in enumerate(call_args_with_tensor):
+            is_tuple_tensor = real_call_args_tensor[i].endswith("_vector")
+            is_tensor = real_call_args_tensor[i] in call_tensors
+            if is_tensor:
+                cast_input += f'const auto &real_{tensor} = PyBoostUtils::CastTensor({tensor}, ' \
+                              f'select_kernel.input_type()[{i}].dtype, "CPU");\n'
+                real_call_args_tensor[i] = "real_" + real_call_args_tensor[i]
+            if is_tuple_tensor:
+                cast_input += f'const auto &real_{tensor} = PyBoostUtils::CastTensor({tensor}, ' \
+                              f'select_kernel.input_type()[{i}].dtype, "CPU");\n'
+                real_call_args_tensor[i] = "PyBoostUtils::ConvertTensorVectorToTuple(real_" + real_call_args_tensor[
+                    i] + ")"
+        if cast_input != "":
+            cast_input = "auto &select_kernel = kernel_attr_pair.second;\n" + cast_input
+        return cast_input, real_call_args_tensor
+
+    def _generate_create_input_address(self, op_parser: OpTemplateParser):
+        need_malloc_tensors, _, _ = op_parser.parse_need_malloc_tensors()
+        create_input_address = ''
+        args_list = ''
+        for item in need_malloc_tensors:
+            args_list += f'{item}, '
+        args_list = args_list[:-2]
+        if args_list:
+            create_input_address = f'PyBoostUtils::PrepareOpInputs(device_context_, op->stream_id(), {args_list});\n'
+        return create_input_address
+
+    def _generate_malloc_input(self, op_parser: OpTemplateParser):
+        """
+        Generates the code for creating input addresses for tensors that need to be allocated.
+
+        Args:
+            op_parser (OpTemplateParser): The parser object for the operation prototype.
+
+        Returns:
+            str: The generated code for creating input addresses.
+        """
+        need_malloc_tensors, _, _ = op_parser.parse_need_malloc_tensors()
+        malloc_inputs = ''
+        args_list = ''
+        for item in need_malloc_tensors:
+            args_list += f'{item}, '
+        args_list = args_list[:-2]
+        if args_list:
+            malloc_inputs += f'PyBoostUtils::MallocOpInputs(device_context, {args_list});\n'
+        return malloc_inputs
+
+    def _generate_get_inputs_kernel_tensors(self, op_parser: OpTemplateParser):
+        """
+        Generates the code for retrieving input kernel tensors.
+
+        Args:
+            op_parser (OpTemplateParser): The parser object for the operation prototype.
+
+        Returns:
+            str: The generated code for retrieving input kernel tensors.
+        """
+        _, _, call_args_with_tensor = op_parser.parse_need_malloc_tensors()
+        inputs_kernel_tensors = ''
+        args_list = ''
+        for item in call_args_with_tensor:
+            args_list += f'{item}, '
+        args_list = args_list[:-2]
+        if args_list:
+            inputs_kernel_tensors += f'const auto &input_address_info = PyBoostUtils::GetAddressInfo(' \
+                                     f'device_context, op->stream_id(), op->input_abs(), {args_list});\n'
+        return inputs_kernel_tensors
+
+
+class PyboostOpFunctionGenerator(BaseGenerator):
+    """
+    Generates C++ source files for ACLNN operations in PyBoost.
+
+    This class handles the generation of source files for operations that utilize the ACLNN framework,
+    including customized calls and tensor management.
+
+    Attributes:
+        PYBOOST_CALL_TEMPLATE (Template): Template for generating ACLNN operation calls.
+        PYBOOST_OP_SOURCE_TEMPLATE (Template): Template for generating operation source files.
+        gen_path (str): Path for saving the generated C++ source files.
+        device (str): The target device (ascend, cpu, or gpu).
+    """
+
+    def __init__(self):
+        self.ascend_op_cpp_generator = PyboostOpCppGenerator('ascend')
+        self.ascend_view_op_cpp_generator = PyboostViewOpCppGenerator('ascend')
+        self.ascend_aclnn_cpp_generator = AclnnOpCppCodeGenerator('ascend')
+
+        self.cpu_op_cpp_generator = PyboostOpCppGenerator('cpu')
+        self.cpu_view_op_cpp_generator = PyboostViewOpCppGenerator('cpu')
+        self.cpu_aclnn_cpp_generator = AclnnOpCppCodeGenerator('cpu')
+
+        self.gpu_op_cpp_generator = PyboostOpCppGenerator('gpu')
+        self.gpu_view_op_cpp_generator = PyboostViewOpCppGenerator('gpu')
+        self.gpu_aclnn_cpp_generator = AclnnOpCppCodeGenerator('gpu')
+
+        self.PYBOOST_ASCEND_OP_SOURCE_TEMPLATE = template.PYBOOST_ASCEND_OP_SOURCE_TEMPLATE
+        self.PYBOOST_CPU_OP_SOURCE_TEMPLATE = template.PYBOOST_CPU_OP_SOURCE_TEMPLATE
+        self.PYBOOST_GPU_OP_SOURCE_TEMPLATE = template.PYBOOST_GPU_OP_SOURCE_TEMPLATE
+        self.ascend_gen_path = f"{K.MS_OPS_KERNEL_PATH}/ascend/pyboost/auto_generate/"
+        self.cpu_gen_path = f"{K.MS_OPS_KERNEL_PATH}/cpu/pyboost/auto_generate/"
+        self.gpu_gen_path = f"{K.MS_OPS_KERNEL_PATH}/gpu/pyboost/auto_generate/"
+
+    def generate(self, work_path, op_protos):
+        """
+        Generate and save C++ source code for PyBoost operations across different devices.
+
+        This method generates C++ source files for operations (`op_protos`) tailored to Ascend, CPU,
+        and GPU devices. It combines headers and function implementations for each device, and then
+        saves the final source files to the appropriate paths.
+
+        Args:
+            op_protos (list): A list of operation prototypes containing metadata such as
+                              operation name, dispatch settings, arguments, and view attributes.
+            work_path (str): The base working directory where the generated files will be saved.
+
+        Generated Files:
+            - Ascend: `pyboost_ascend_ops.cc`
+            - CPU: `pyboost_cpu_ops.cc`
+            - GPU: `pyboost_gpu_ops.cc`
+        """
+        self._generate_pyboost_ascend_ops(work_path, op_protos)
+        self._generate_pyboost_cpu_ops(work_path, op_protos)
+        self._generate_pyboost_gpu_ops(work_path, op_protos)
+
+    def _generate_pyboost_ascend_ops(self, work_path, op_protos):
+        """
+        Generates Ascend PyBoost ops functions source files after being merged into specific chunk sizes.
+
+        Args:
+            work_path (str): The directory path where the generated C++ source files will be saved.
+            op_protos (list): A list of operation prototypes that define the operations for which
+                              the C++ code will be generated.
+        """
+        ascend_merge_op_header = []
+        ascend_merge_op_function = []
+        self.ascend_op_cpp_generator.generate_customize_op_cpp_code(op_protos, ascend_merge_op_header,
+                                                                    ascend_merge_op_function)
+        self.ascend_view_op_cpp_generator.generate_view_op_cpp_code(op_protos, ascend_merge_op_header,
+                                                                    ascend_merge_op_function)
+        self.ascend_aclnn_cpp_generator.generate_aclnn_op_cpp_code(op_protos, ascend_merge_op_header,
+                                                                   ascend_merge_op_function)
+
+        ascend_op_header_merge_by_chunk_size = merge_strings_by_chunk_size(ascend_merge_op_header, chunk_size=120)
+        ascend_op_function_merge_by_chunk_size = merge_strings_by_chunk_size(ascend_merge_op_function, chunk_size=120)
+
+        new_gen_num = len(ascend_op_header_merge_by_chunk_size)
+        self._delete_residual_merged_ops_files(os.path.join(work_path, self.ascend_gen_path), new_gen_num)
+
+        for i, op_header, op_function in zip(range(len(ascend_op_header_merge_by_chunk_size)),
+                                             ascend_op_header_merge_by_chunk_size,
+                                             ascend_op_function_merge_by_chunk_size):
+            ascend_pyboost_op_source = self.PYBOOST_ASCEND_OP_SOURCE_TEMPLATE.replace(
+                merge_op_header=op_header, merge_op_function=op_function)
+            save_file(os.path.join(work_path, self.ascend_gen_path), f"pyboost_ascend_ops_{i}.cc",
+                      ascend_pyboost_op_source)
+
+    def _generate_pyboost_cpu_ops(self, work_path, op_protos):
+        """
+        Generates CPU PyBoost ops functions source files after being merged into specific chunk sizes.
+
+        Args:
+            work_path (str): The directory path where the generated C++ source files will be saved.
+            op_protos (list): A list of operation prototypes that define the operations for which
+                              the C++ code will be generated.
+        """
+        cpu_merge_op_header = []
+        cpu_merge_op_function = []
+        self.cpu_op_cpp_generator.generate_customize_op_cpp_code(op_protos, cpu_merge_op_header, cpu_merge_op_function)
+        self.cpu_view_op_cpp_generator.generate_view_op_cpp_code(op_protos, cpu_merge_op_header, cpu_merge_op_function)
+        self.cpu_aclnn_cpp_generator.generate_aclnn_op_cpp_code(op_protos, cpu_merge_op_header, cpu_merge_op_function)
+        cpu_op_header_merge_by_chunk_size = merge_strings_by_chunk_size(cpu_merge_op_header, chunk_size=120)
+        cpu_op_function_merge_by_chunk_size = merge_strings_by_chunk_size(cpu_merge_op_function, chunk_size=120)
+
+        new_gen_num = len(cpu_op_header_merge_by_chunk_size)
+        self._delete_residual_merged_ops_files(os.path.join(work_path, self.cpu_gen_path), new_gen_num)
+
+        for i, op_header, op_function in zip(range(len(cpu_op_header_merge_by_chunk_size)),
+                                             cpu_op_header_merge_by_chunk_size,
+                                             cpu_op_function_merge_by_chunk_size):
+            cpu_pyboost_op_source = self.PYBOOST_CPU_OP_SOURCE_TEMPLATE.replace(
+                merge_op_header=op_header, merge_op_function=op_function)
+            save_file(os.path.join(work_path, self.cpu_gen_path), f"pyboost_cpu_ops_{i}.cc",
+                      cpu_pyboost_op_source)
+
+    def _generate_pyboost_gpu_ops(self, work_path, op_protos):
+        """
+        Generates GPU PyBoost ops functions source files after being merged into specific chunk sizes.
+
+        Args:
+            work_path (str): The directory path where the generated C++ source files will be saved.
+            op_protos (list): A list of operation prototypes that define the operations for which
+                              the C++ code will be generated.
+        """
+        gpu_merge_op_header = []
+        gpu_merge_op_function = []
+        self.gpu_op_cpp_generator.generate_customize_op_cpp_code(op_protos, gpu_merge_op_header, gpu_merge_op_function)
+        self.gpu_view_op_cpp_generator.generate_view_op_cpp_code(op_protos, gpu_merge_op_header, gpu_merge_op_function)
+        self.gpu_aclnn_cpp_generator.generate_aclnn_op_cpp_code(op_protos, gpu_merge_op_header, gpu_merge_op_function)
+        gpu_op_header_merge_by_chunk_size = merge_strings_by_chunk_size(gpu_merge_op_header, chunk_size=120)
+        gpu_op_function_merge_by_chunk_size = merge_strings_by_chunk_size(gpu_merge_op_function, chunk_size=120)
+
+        new_gen_num = len(gpu_op_header_merge_by_chunk_size)
+        self._delete_residual_merged_ops_files(os.path.join(work_path, self.gpu_gen_path), new_gen_num)
+
+        for i, op_header, op_function in zip(range(len(gpu_op_header_merge_by_chunk_size)),
+                                             gpu_op_header_merge_by_chunk_size,
+                                             gpu_op_function_merge_by_chunk_size):
+            gpu_pyboost_op_source = self.PYBOOST_GPU_OP_SOURCE_TEMPLATE.replace(
+                merge_op_header=op_header, merge_op_function=op_function)
+            save_file(os.path.join(work_path, self.gpu_gen_path), f"pyboost_gpu_ops_{i}.cc",
+                      gpu_pyboost_op_source)
+
+    def _delete_residual_merged_ops_files(self, files_path, new_gen_num):
+        """
+        Deletes residual merged operation files in the specified directory if the number of
+        newly generated files does not match the number of existing ones.
+
+        This method first lists all files in the specified directory, then filters out the files
+        that match the pattern `pyboost_.*_ops_.*.cc` (i.e., files related to pyboost ops). It compares
+        the number of such files (`old_files_num`) with the `new_gen_num` argument, which represents
+        the expected number of new pyboost ops files. If the counts do not match, the method will
+        delete all the existing pyboost ops files in the directory before any new ones can be generated.
+
+        Args:
+            files_path (str): The path to the directory containing the files to be checked and deleted.
+            new_gen_num (int): The number of newly generated pyboost ops files expected to be in the directory.
+
+        Returns:
+            None
+        """
+        all_files = os.listdir(files_path)
+        old_pyboost_ops_files = [file for file in all_files if re.match(r'pyboost_.*_ops_.*\.cc', file)]
+        old_files_num = len(old_pyboost_ops_files)
+        if new_gen_num != old_files_num:
+            for file in old_pyboost_ops_files:
+                os.remove(os.path.join(files_path, file))
+
+
+def _generate_cpp_func_return(op_proto):
+    """Generates the C++ return type for the given operator prototype.
+
+    Args:
+        op_proto (OpProto): The operator prototype containing return information.
+
+    Returns:
+        str: The C++ return type for the function based on the operator prototype.
+
+    Raises:
+        Exception: If no 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 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_inplace_process_cpp_code(op_proto):
+    """Generates C++ code for updating outputs by input tensors for inplace processing.
+
+    Args:
+        op_proto (OpProto): The operator prototype containing return information.
+
+    Returns:
+        str: The C++ code for inplace processing, or an empty string if no inplace processing is needed.
+    """
+    inplace_process = f'// RefOps update output by input tensor\n'
+    has_ref = False
+    for index, return_obj in enumerate(op_proto.op_returns):
+        if return_obj.inplace != '':
+            inplace_process += f'outputs_[{index}]->set_device_address(' \
+                               f'{return_obj.inplace}_tensor->device_address()); '
+            has_ref = True
+            break
+    if has_ref:
+        return inplace_process
+    return ''
+
+
+def delete_residual_files(work_path, op_protos):
+    """
+    Deletes residual files generated for operator prototypes that are no longer needed.
+
+    Args:
+        work_path (str): The base directory path where generated files are located.
+        op_protos (list): A list of operator prototypes that are currently valid.
+
+    Returns:
+        None
+    """
+    all_operator_name = []
+    for op_proto in op_protos:
+        all_operator_name.append(op_proto.op_name)
+    code_generate_path_list = [f"{K.MS_OPS_KERNEL_PATH}/{device}/pyboost/auto_generate/" for device in
+                               ["ascend", "gpu", "cpu"]]
+    code_generate_path_list.append(f"{K.MS_COMMON_PYBOOST_KERNEL_PATH}/auto_generate/")
+    for code_generate_path in code_generate_path_list:
+        filter_files = []
+        code_generate_path = os.path.join(work_path, code_generate_path)
+        if os.path.exists(code_generate_path):
+            all_files = os.listdir(code_generate_path)
+            # No need to delete pyboost_.*_ops_.*.cc files and op_register.cc.
+            # These residual files will be deleted before new files generate.
+            filter_files = [file for file in all_files if
+                            not re.match(r'pyboost_.*_ops_.*\.cc', file) and file != "op_register.cc"]
+        registered_op_name = set(item.split(".")[0] for item in filter_files)
+        need_clean_op = registered_op_name - set(all_operator_name)
+
+        for file in filter_files:
+            file_name = file.split(".")[0]
+            if file_name in need_clean_op:
+                file_path = os.path.join(code_generate_path, file)
+                if os.path.exists(file_path):
+                    os.remove(file_path)
+
+
+class PyboostOpRegisterCppCodeGenerator:
+    """
+    Generates registration C++ code for PyBoost operations.
+
+    This class is responsible for creating a registration source file that includes
+    all the necessary headers and template instantiations for the registered operations.
+
+    Attributes:
+        PYBOOST_OP_REGISTER_TEMPLATE (Template): Template for generating the operation registration code.
+    """
+
+    def __init__(self):
+        self.PYBOOST_OP_REGISTER_TEMPLATE = template.PYBOOST_OP_REGISTER_TEMPLATE
+
+    def generate(self, work_path, op_protos):
+        """
+        Generates a C++ source file for registering all PyBoost operations.
+
+        Args:
+            work_path (str): The directory path where the registration file will be saved.
+            op_protos (list): A list of operator prototypes containing information about the operations.
+
+        Returns:
+            None
+        """
+        all_op_names = []
+        all_functional_names = []
+        for op_proto in op_protos:
+            if op_proto.op_dispatch is None:
+                continue
+            functional_name = op_proto.op_name
+            op_name_str = op_proto.op_class.name
+            all_op_names.append(op_name_str)
+            all_functional_names.append(functional_name)
+
+        include_str = ''
+        factory_str = ''
+        for op_name in all_op_names:
+            factory_str += "template class OpFactory<{0}>;\n".format(op_name)
+        for operator_name in all_functional_names:
+            include_str += f'#include "{K.MS_COMMON_PYBOOST_KERNEL_PATH}/auto_generate/{operator_name}.h"\n'
+        op_register_file_str = self.PYBOOST_OP_REGISTER_TEMPLATE.replace(op_includes=include_str,
+                                                                         op_factory_templates=factory_str)
+        save_path = os.path.join(work_path, f"{K.MS_COMMON_PYBOOST_KERNEL_PATH}/auto_generate/")
+        file_name = "op_register.cc"
+        save_file(save_path, file_name, op_register_file_str)