mindstudio-probe 1.1.0__py3-none-any.whl → 1.2.1__py3-none-any.whl

msprobe/mindspore/compare/ms_compare.py

@@ -1,33 +1,73 @@
+# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
+# All rights reserved.
+#
+# 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.
+
 import os
 import re
-import copy
-import sys
-from itertools import zip_longest
-
-from msprobe.core.common.utils import check_compare_param, CompareException, check_configuration_param, \
-    task_dumppath_get, struct_json_get, add_time_with_yaml
-from msprobe.core.common.file_utils import create_directory, load_yaml, load_npy, load_json, save_yaml, FileOpen
-from msprobe.core.common.const import Const, CompareConst
-from msprobe.core.common.log import logger
+from collections import defaultdict
+
+import numpy as np
+import pandas as pd
+
+from msprobe.core.common.const import CompareConst, Const
 from msprobe.core.common.exceptions import FileCheckException
-from msprobe.core.compare.acc_compare import Comparator
-from msprobe.core.compare.check import check_struct_match, fuzzy_check_op
-from msprobe.mindspore.compare.modify_mapping import modify_mapping_with_stack
-from msprobe.mindspore.compare.layer_mapping import get_layer_mapping
+from msprobe.core.common.file_utils import FileOpen, create_directory, load_json, load_npy, load_yaml
+from msprobe.core.common.log import logger
+from msprobe.core.common.utils import CompareException, check_compare_param, check_configuration_param, \
+    check_op_str_pattern_valid, get_dump_mode, set_dump_path
+from msprobe.core.compare.acc_compare import Comparator, ModeConfig
+from msprobe.core.compare.check import dtype_mapping
+from msprobe.core.compare.layer_mapping import generate_data_mapping_by_layer_mapping
+from msprobe.core.compare.utils import set_stack_json_path, reorder_op_x_list
 
-class MSComparator(Comparator):
-    def __init__(self, cell_mapping=None, api_mapping=None, data_mapping=None, is_cross_framework=False):
-        self.frame_name = MSComparator.__name__
+
+class MappingConfig:
+    def __init__(self, cell_mapping=None, api_mapping=None, data_mapping=None):
         self.cell_mapping = cell_mapping
         self.api_mapping = api_mapping
         self.data_mapping = data_mapping
-        if data_mapping:
+
+
+class MSComparator(Comparator):
+    """
+    用于mindspore动态图同框架/跨框架精度比对，支持md5/summary/all模式。
+    cell_mapping: mindspore在cell级别(L0)dump数据和pytorch的module之间的映射关系；
+    api_mapping: mindspore在api级别(L1)dump数据和pytorch的api之间的映射关系；
+    data_mapping: mindspore的cell或api的入参/出参和pytorch之间的映射关系；
+    is_cross_framework: 是否跨框架。
+    """
+    def __init__(self, mode_config, mapping_config=None, is_cross_framework=False):
+        super().__init__(mode_config)
+        self.frame_name = MSComparator.__name__
+
+        self.stack_mode = mode_config.stack_mode
+        self.auto_analyze = mode_config.auto_analyze
+        self.fuzzy_match = mode_config.fuzzy_match
+        self.dump_mode = mode_config.dump_mode
+
+        if mapping_config:
+            self.cell_mapping = mapping_config.cell_mapping
+            self.api_mapping = mapping_config.api_mapping
+            self.data_mapping = mapping_config.data_mapping
+
+        if self.data_mapping:
             self.cross_frame = is_cross_framework
         else:
-            self.cross_frame = cell_mapping is not None or api_mapping is not None
+            self.cross_frame = self.cell_mapping is not None or self.api_mapping is not None
         self.cell_mapping_dict = self.load_mapping_file(self.cell_mapping)
         self.api_mapping_dict = self.load_mapping_file(self.api_mapping)
-        if api_mapping is not None:
+        if self.api_mapping is not None:
             self.ms_to_pt_mapping = self.load_internal_api()
 
         if isinstance(self.data_mapping, str) or self.data_mapping is None:
@@ -38,9 +78,106 @@ class MSComparator(Comparator):
             raise TypeError(f"The type of parameter `data_mapping` must be dict, str or None, but got "
                             f"{type(self.data_mapping)}")
 
+    def calc_accuracy(self, result_df, header):
+        condition_no_bench = result_df[CompareConst.BENCH_NAME] == CompareConst.N_A
+        result_df[condition_no_bench] = result_df[condition_no_bench].fillna(CompareConst.N_A)
+        result_df.loc[condition_no_bench, CompareConst.ERROR_MESSAGE] = CompareConst.NO_BENCH
+
+        def calc_summary_diff(data_type: str):
+            def type_check(val):
+                check_series = pd.Series(False, index=val.index)
+                val_str = val.astype(str)
+                check_series[pd.to_numeric(val_str, errors='coerce').notna() | val_str.str.lower().eq('nan')] = True
+                return check_series
+
+            def get_number(val):
+                return pd.to_numeric(val.astype(str), errors='coerce')
+
+            ms_val = result_df['NPU ' + data_type]
+            pt_val = result_df['Bench ' + data_type]
+            diff_name = data_type.capitalize() + ' diff'
+            rel_err_name = ('norm' if data_type == 'l2norm' else data_type).capitalize() + 'RelativeErr'
+            condition_na = ~type_check(ms_val) | ~type_check(pt_val)
+            result_df.loc[condition_na, [diff_name, rel_err_name]] = CompareConst.N_A
+            result_df.loc[~(condition_no_bench | condition_na), diff_name] = get_number(ms_val) - get_number(pt_val)
+            condition_nan_diff = ~condition_no_bench & ~condition_na & result_df[diff_name].isna()
+            condition_not_nan_diff = ~condition_no_bench & ~condition_na & result_df[diff_name].notna()
+            result_df.loc[condition_nan_diff, [diff_name, rel_err_name]] = CompareConst.NAN
+            condition_pt_zero = pt_val == 0
+            result_df.loc[condition_not_nan_diff & condition_pt_zero, rel_err_name] = CompareConst.NAN
+            condition_ref_err = condition_not_nan_diff & ~condition_pt_zero
+            result_df.loc[condition_ref_err, rel_err_name] = (result_df.loc[condition_ref_err, diff_name] /
+                                                              pt_val[condition_ref_err] * 100)
+            result_df.loc[condition_ref_err, rel_err_name] = (result_df.loc[condition_ref_err, rel_err_name]
+                                                              .abs().astype(str) + '%')
+            magnitude = get_number(result_df[diff_name]).abs() / (
+                    pd.Series(np.maximum(get_number(ms_val), get_number(pt_val))).abs() + CompareConst.EPSILON)
+            return magnitude > CompareConst.MAGNITUDE
+
+        if self.dump_mode == Const.MD5:
+            condition_md5_equal = result_df[CompareConst.NPU_MD5] == result_df[CompareConst.BENCH_MD5]
+            result_df.loc[condition_md5_equal, CompareConst.RESULT] = CompareConst.PASS
+            result_df.loc[~condition_md5_equal & ~condition_no_bench, CompareConst.RESULT] = CompareConst.DIFF
+        elif self.dump_mode == Const.SUMMARY:
+            warning_list = [calc_summary_diff(data_type) for data_type in ['max', 'min', 'mean', 'l2norm']]
+            warning_flag = pd.DataFrame(warning_list).all()
+            result_df.loc[~condition_no_bench, [CompareConst.RESULT, CompareConst.ERROR_MESSAGE]] = ''
+            result_df.loc[warning_flag, CompareConst.RESULT] = CompareConst.WARNING
+            result_df.loc[warning_flag, CompareConst.ERROR_MESSAGE] = 'Need double check api accuracy.'
+        else:
+            fill_cols = [CompareConst.COSINE, CompareConst.MAX_ABS_ERR, CompareConst.MAX_RELATIVE_ERR,
+                         CompareConst.ONE_THOUSANDTH_ERR_RATIO, CompareConst.FIVE_THOUSANDTHS_ERR_RATIO,
+                         CompareConst.ERROR_MESSAGE]
+            result_df.loc[~condition_no_bench, fill_cols] = ''
+            result_df.loc[~condition_no_bench, CompareConst.ACCURACY] = CompareConst.ACCURACY_CHECK_YES
+        return result_df[header]
+
+    def make_result_df(self, result):
+        header = CompareConst.HEAD_OF_COMPARE_MODE[self.dump_mode][:]
+
+        if self.stack_mode:
+            header.append(CompareConst.STACK)
+        if self.dump_mode == Const.ALL:
+            header.append(CompareConst.DATA_NAME)
+        result.rename(columns={'op_name_x': CompareConst.NPU_NAME,
+                               'op_name_y': CompareConst.BENCH_NAME,
+                               'dtype_x': CompareConst.NPU_DTYPE,
+                               'dtype_y': CompareConst.BENCH_DTYPE,
+                               'shape_x': CompareConst.NPU_SHAPE,
+                               'shape_y': CompareConst.BENCH_SHAPE,
+                               'md5_x': CompareConst.NPU_MD5,
+                               'md5_y': CompareConst.BENCH_MD5,
+                               'data_name_x': CompareConst.DATA_NAME,
+                               'stack_info_x': CompareConst.STACK}, inplace=True)
+
+        npu_summary = [CompareConst.NPU_MAX, CompareConst.NPU_MIN, CompareConst.NPU_MEAN, CompareConst.NPU_NORM]
+        bench_summary = [CompareConst.BENCH_MAX, CompareConst.BENCH_MIN, CompareConst.BENCH_MEAN,
+                         CompareConst.BENCH_NORM]
+
+        def set_summary(summary):
+            if summary == CompareConst.N_A:
+                return [CompareConst.N_A] * 4
+            summary_list = []
+            for i in summary:
+                if i is None:
+                    summary_list.append(CompareConst.N_A)
+                elif str(i).lower() == 'nan':
+                    summary_list.append(CompareConst.NAN)
+                else:
+                    summary_list.append(i)
+            return summary_list
+
+        result[npu_summary] = result['summary_x'].apply(set_summary).tolist()
+        result[bench_summary] = result['summary_y'].apply(set_summary).tolist()
+        result_df = pd.DataFrame(columns=header)
+        for h in header:
+            if h in result.columns:
+                result_df[h] = result[h]
+        return self.calc_accuracy(result_df, header)
+
     def load_internal_api(self):
         cur_path = os.path.dirname(os.path.realpath(__file__))
-        yaml_path = os.path.join(cur_path, "ms_to_pt_api.yaml")
+        yaml_path = os.path.abspath(os.path.join(cur_path, CompareConst.INTERNAL_API_MAPPING_FILE))
         return load_yaml(yaml_path)
 
     def load_mapping_file(self, mapping_file):
@@ -51,42 +188,23 @@ class MSComparator(Comparator):
         return mapping_dict
 
     def process_cell_mapping(self, npu_op_name):
-        npu_op_name = [op_name.replace("Cell", "Module", 1) for op_name in npu_op_name]
+        if not npu_op_name:
+            return CompareConst.N_A
+        param_grad_flag = Const.PARAMS_GRAD in npu_op_name.split(Const.SEP)
+        if not param_grad_flag and not re.search(Const.REGEX_FORWARD_BACKWARD, npu_op_name):
+            return CompareConst.N_A
+        npu_op_name = npu_op_name.replace("Cell", "Module", 1)
         if self.cell_mapping_dict:
-            for index, op_name in enumerate(npu_op_name):
-                # get cell name & class name from op_name
-                # Cell.fc1.Dense.forward.0.input.0
-                cell_name = op_name.split(Const.SEP, 1)[-1].rsplit(Const.SEP, 4)[0]
-                if cell_name in self.cell_mapping_dict:
-                    npu_op_name[index] = op_name.replace(cell_name, self.cell_mapping_dict[cell_name], 1)
+            # get cell name & class name from op_name
+            # Cell.fc1.Dense.forward.0.input.0
+            cell_name = re.split(r'\.(?:forward|backward|parameters_grad)\.', npu_op_name.split(Const.SEP, 1)[-1])[0]
+            if cell_name in self.cell_mapping_dict:
+                npu_op_name = npu_op_name.replace(cell_name, self.cell_mapping_dict[cell_name], 1)
         return npu_op_name
 
-    def check_op(self, npu_dict, bench_dict, fuzzy_match):
-        npu_dict_new, bench_dict_new = copy.deepcopy(npu_dict), copy.deepcopy(bench_dict)  
-        npu_op_name, bench_op_name = npu_dict_new.get(CompareConst.OP_NAME), bench_dict_new.get(CompareConst.OP_NAME)
-        if self.cell_mapping is not None:
-            npu_op_name = self.process_cell_mapping(npu_op_name)
-        if self.api_mapping is not None:
-            npu_op_name = self.process_internal_api_mapping(npu_op_name, bench_op_name)
-            if isinstance(self.api_mapping, str):
-                npu_dict_new, bench_dict_new, target_dict = self.transform_user_mapping_api(npu_dict_new, 
-                                                                                            bench_dict_new)
-                if target_dict:
-                    bench_dict = self.reconstitution_bench_dict(npu_dict, copy.deepcopy(bench_dict_new), target_dict)
-                    npu_op_name = npu_dict_new.get(CompareConst.OP_NAME) 
-                    bench_op_name = bench_dict_new.get(CompareConst.OP_NAME)
-        struct_match = check_struct_match(npu_dict_new, bench_dict_new, cross_frame=self.cross_frame)
-        if not fuzzy_match:
-            return npu_op_name == bench_op_name and struct_match
-        is_match = True
-        try:
-            is_match = fuzzy_check_op(npu_op_name, bench_op_name)
-        except Exception as err:
-            logger.warning("%s and %s can not fuzzy match." % (npu_op_name, bench_op_name))
-            is_match = False
-        return is_match and struct_match
-    
     def read_npy_data(self, dir_path, file_name, load_pt_file=False):
+        if not file_name:
+            return None
         data_path = os.path.join(dir_path, file_name)
         if load_pt_file:
             import torch
@@ -96,35 +214,23 @@ class MSComparator(Comparator):
                 data_value = data_value.to(torch.float32)
             data_value = data_value.numpy()
         else:
-            data_value = load_npy(data_path) 
-        return data_value    
+            data_value = load_npy(data_path)
+        return data_value
 
-    def api_replace(self, npu_op_name, target, para):
-        for idx, _ in enumerate(npu_op_name):
-            npu_op_name[idx] = npu_op_name[idx].replace(target, para)
-        return npu_op_name
-    
-    def process_internal_api_mapping(self, npu_op_name, bench_op_name):
+    def process_internal_api_mapping(self, npu_op_name):
         # get api name & class name from op_name
         # Functional.addcmul.0.forward.input.0
-        npu_op_name, bench_op_name = npu_op_name.copy(), bench_op_name.copy()
-        ms_api_name = self.get_api_name(npu_op_name[0].split(Const.SEP))
-        pt_api_name = self.get_api_name(bench_op_name[0].split(Const.SEP))
+        ms_api_name = self.get_api_name(npu_op_name.split(Const.SEP))
         class_name = ms_api_name.split(Const.SEP)[0]
         if class_name == "Mint":
-            return self.api_replace(npu_op_name, "Mint", "Torch")
+            return npu_op_name.replace("Mint", "Torch")
         elif class_name == "MintFunctional":
-            return self.api_replace(npu_op_name, "MintFunctional", "Functional")
-        elif self.ms_to_pt_mapping.get(ms_api_name) == pt_api_name:
-            return self.api_replace(npu_op_name, ms_api_name, pt_api_name)
+            return npu_op_name.replace("MintFunctional", "Functional")
+        elif self.ms_to_pt_mapping.get(ms_api_name):
+            return npu_op_name.replace(ms_api_name, self.ms_to_pt_mapping.get(ms_api_name))
         else:
             return npu_op_name
-    
-    def remove_element(self, op_name, struct, summary, idx):
-        del op_name[idx]
-        del struct[idx]
-        del summary[idx]
-    
+
     def get_api_name(self, api_list):
         try:
             api_name = api_list[0] + Const.SEP + api_list[1]
@@ -132,184 +238,147 @@ class MSComparator(Comparator):
             logger.error(f'Failed to retrieve API name, please check if the dump data is reasonable')
             raise CompareException(CompareException.INDEX_OUT_OF_BOUNDS_ERROR) from error
         return api_name
-    
-    def transform_user_mapping_api(self, new_npu_dict, new_bench_dict):
-        """
-        Transform user mapping API based on new NPU and benchmark dictionaries.
-        Parameters:
-            new_npu_dict (dict): New NPU operation dictionary.
-            new_bench_dict (dict): New benchmark operation dictionary.
-        Returns:
-            tuple: Updated NPU and benchmark dictionaries, along with the target dictionary.
-        """
-        npu_op_name, bench_op_name = new_npu_dict.get(CompareConst.OP_NAME), new_bench_dict.get(CompareConst.OP_NAME)
-        npu_struct_in = new_npu_dict.get(CompareConst.INPUT_STRUCT)
-        bench_struct_in = new_bench_dict.get(CompareConst.INPUT_STRUCT)
-        npu_struct_out = new_npu_dict.get(CompareConst.OUTPUT_STRUCT)
-        bench_struct_out =  new_bench_dict.get(CompareConst.OUTPUT_STRUCT)
-        npu_summary, bench_summary = new_npu_dict.get(CompareConst.SUMMARY), new_bench_dict.get(CompareConst.SUMMARY)
-        npu_in_len, bench_in_len = len(npu_struct_in), len(bench_struct_in) 
-        npu_out_len, bench_out_len = len(npu_struct_out), len(bench_struct_out)
-        ms_api_list, pt_api_list = npu_op_name[0].split(Const.SEP), bench_op_name[0].split(Const.SEP)
-        ms_api_name = self.get_api_name(ms_api_list)
-        pt_api_name = self.get_api_name(pt_api_list)
-        target_dict = {}
-        for api_dict in self.api_mapping_dict:
-            if api_dict.get("pt_api") == pt_api_name and api_dict.get("ms_api") == ms_api_name:
-                ms_user_args_len, pt_user_args_len = len(api_dict.get("ms_args")), len(api_dict.get("pt_args"))
-                ms_user_output_len, pt_user_output_len = len(api_dict.get("ms_output")), len(api_dict.get("pt_output"))
-                if ms_user_args_len != pt_user_args_len or ms_user_output_len != pt_user_output_len:
-                    logger.warning("The user-defined mapping table is incorrect,\
-                        make sure that the number of parameters is equal")
-                    break
-                ms_out_list = api_dict.get("ms_output", [])
-                for idx in reversed(range(npu_out_len)):
-                    if idx not in ms_out_list:
-                        del npu_struct_out[idx]
-                        if idx + npu_in_len < len(npu_summary) and idx + npu_in_len < len(npu_op_name): 
-                            del npu_summary[idx + npu_in_len]
-                            del npu_op_name[idx + npu_in_len]
-                pt_out_list = api_dict.get("pt_output", [])
-                for idx in reversed(range(bench_out_len)):
-                    if idx not in pt_out_list:
-                        del bench_struct_out[idx]
-                        if idx + bench_in_len < len(bench_summary) and idx + bench_in_len < len(bench_op_name): 
-                            del bench_summary[idx + bench_in_len]
-                            del bench_op_name[idx + bench_in_len]
-                ms_para_list = api_dict.get("ms_args", []) 
-                for idx in reversed(range(npu_in_len)):
-                    if idx not in ms_para_list:
-                        self.remove_element(npu_op_name, npu_struct_in, npu_summary, idx)
-                pt_para_list = api_dict.get("pt_args", []) 
-                for idx in reversed(range(bench_in_len)):
-                    if idx not in pt_para_list:
-                        self.remove_element(bench_op_name, bench_struct_in, bench_summary, idx)
-                npu_op_name = self.api_replace(npu_op_name, ms_api_name, pt_api_name)
-                npu_op_name = self.para_sequence_update(npu_op_name, bench_op_name)
-                target_dict = api_dict
-                break
-        if target_dict:
-            new_npu_dict.update({CompareConst.OP_NAME: npu_op_name, CompareConst.INPUT_STRUCT: npu_struct_in, 
-                                 CompareConst.OUTPUT_STRUCT: npu_struct_out, CompareConst.SUMMARY: npu_summary})
-            new_bench_dict.update({CompareConst.OP_NAME: bench_op_name, CompareConst.INPUT_STRUCT: bench_struct_in,
-                                   CompareConst.OUTPUT_STRUCT: bench_struct_out, CompareConst.SUMMARY: bench_summary})
-        return new_npu_dict, new_bench_dict, target_dict  
-    
-    def para_sequence_update(self, npu_op_name, bench_op_name):
-        for idx, _ in enumerate(npu_op_name):
-            bench_op_name_list = bench_op_name[idx].rsplit(Const.SEP, 1)
-            if len(bench_op_name_list) != 0:
-                npu_op_name[idx] = npu_op_name[idx][:-1] + bench_op_name_list[-1]
-        return npu_op_name
 
-    def reconstitution_bench_dict(self, npu_dict, del_bench_dict, api_dict):
-        ms_user_args_list = api_dict.get("ms_args", [])
-        ms_user_output_list = api_dict.get("ms_output", [])
-        npu_struct_in = npu_dict.get(CompareConst.INPUT_STRUCT)
-        npu_struct_out = npu_dict.get(CompareConst.OUTPUT_STRUCT)
-        npu_in_len = len(npu_struct_in)
-        npu_out_len = len(npu_struct_out)
-        if npu_in_len == len(ms_user_args_list) and npu_out_len == len(ms_user_output_list):
-            return del_bench_dict
-        ms_input_args_list = [i for i in range(npu_in_len)]
-        input_sub_list = list(set(ms_input_args_list) - set(ms_user_args_list))
-        ms_output_args_list = [i for i in range(npu_out_len)]
-        output_sub_list = list(set(ms_output_args_list) - set(ms_user_output_list))
-        bench_op_name = del_bench_dict.get(CompareConst.OP_NAME, [])
-        bench_struct_in = del_bench_dict.get(CompareConst.INPUT_STRUCT, [])
-        bench_struct_out = del_bench_dict.get(CompareConst.OUTPUT_STRUCT, [])
-        bench_summary = del_bench_dict.get(CompareConst.SUMMARY, [])
-        for idx in input_sub_list:  # Fill in the blank value field in the pt dictionary
-            bench_op_name.insert(idx, CompareConst.N_A)
-            bench_struct_in.insert(idx, CompareConst.N_A)
-            bench_summary.insert(idx, CompareConst.N_A)
-        for idx in output_sub_list:  # Fill in the blank value field in the pt dictionary
-            bench_op_name.insert(npu_in_len + idx, CompareConst.N_A)
-            bench_struct_out.insert(idx, CompareConst.N_A)
-            bench_summary.insert(npu_in_len + idx, CompareConst.N_A)
-        del_bench_dict.update({CompareConst.OP_NAME: bench_op_name, CompareConst.INPUT_STRUCT: bench_struct_in, 
-                               CompareConst.OUTPUT_STRUCT: bench_struct_out, CompareConst.SUMMARY: bench_summary})
-        return del_bench_dict
-        
-
-def sort_by_execution_sequence(npu_data, bench_data, mapping_list, flag):
-    def generate_execution_sequence(data):
-        sequence_map = {}
-        for index, item in enumerate(data.keys()):
-            if flag in item:
-                item_split = item.split(Const.SEP)
-                item_name = Const.SEP.join(item_split[0:-2])
-                item_index = item_split[-1]
-                if item_index == 'forward' or item_index == 'backward':
-                    item_index = item_split[-2]
-                item_key = f"{item_name}.{item_index}"
-                sequence_map[item_key] = index
-        return sequence_map
-
-    npu_map = generate_execution_sequence(npu_data)
-    bench_map = generate_execution_sequence(bench_data)
-
-    def sort_by_map(item):
-        first_key = npu_map.get(item[0], sys.maxsize)
-        second_key = bench_map.get(item[1], sys.maxsize)
-        return first_key, second_key
-
-    return sorted(mapping_list, key=sort_by_map)
-
-
-def generate_kernel_data(map_value, data, flag):
-    if not map_value:
-        return [], []
-    inputs_name = []
-    outputs_name = []
-    map_split = map_value.split(Const.SEP)
-    map_name = Const.SEP.join(map_split[0:-1])
-    map_index = map_split[-1]
-    for key, value in data.items():
-        if key.find(flag) != -1 and key.find(map_name) != -1:
-            if key.split(Const.SEP)[-1] != map_index and key.split(Const.SEP)[-2] != map_index :
+    def compare_process(self, file_lists):
+        npu_json_path, bench_json_path, stack_json_path = file_lists
+        npu_json_data = load_json(npu_json_path)
+        bench_json_data = load_json(bench_json_path)
+        stack_json_data = load_json(stack_json_path) if self.stack_mode else None
+
+        npu_df = self.gen_data_df(npu_json_data, stack_json_data)
+        bench_df = self.gen_data_df(bench_json_data, stack_json_data)
+        if self.cell_mapping:
+            npu_df[CompareConst.COMPARE_KEY] = npu_df[CompareConst.OP_NAME].apply(self.process_cell_mapping)
+        elif self.api_mapping:
+            npu_df[CompareConst.COMPARE_KEY] = npu_df[CompareConst.OP_NAME].apply(self.process_internal_api_mapping)
+            if isinstance(self.api_mapping, str):
+                self.modify_compare_data_with_user_mapping(npu_df, bench_df)
+        else:
+            npu_df[CompareConst.COMPARE_KEY] = npu_df[CompareConst.OP_NAME]
+        npu_df[[Const.DTYPE, Const.SHAPE]] = npu_df[[Const.DTYPE, Const.SHAPE]].astype(str)
+        bench_df[[Const.DTYPE, Const.SHAPE]] = bench_df[[Const.DTYPE, Const.SHAPE]].astype(str)
+        npu_df[CompareConst.COMPARE_SHAPE] = npu_df[Const.SHAPE]
+        bench_df[CompareConst.COMPARE_KEY] = bench_df[CompareConst.OP_NAME]
+        bench_df[CompareConst.COMPARE_SHAPE] = bench_df[Const.SHAPE]
+        match_result = pd.merge(npu_df, bench_df, on=[CompareConst.COMPARE_KEY, CompareConst.COMPARE_SHAPE],
+                                how='outer')
+        match_result = match_result[match_result['op_name_x'].notna()].fillna(CompareConst.N_A)
+
+        def gen_dtype_condition():
+            npu_dtype = match_result['dtype_x']
+            bench_dtype = match_result['dtype_y']
+            if self.cross_frame:
+                npu_dtype = npu_dtype.map(dtype_mapping).fillna(npu_dtype)
+            return ((npu_dtype == bench_dtype) |
+                    ((npu_dtype == Const.FLOAT16) & (bench_dtype == Const.FLOAT32)) |
+                    ((npu_dtype == Const.FLOAT32) & (bench_dtype == Const.FLOAT16)) |
+                    ((npu_dtype == Const.FLOAT16) & (bench_dtype == Const.BFLOAT16)) |
+                    ((npu_dtype == Const.BFLOAT16) & (bench_dtype == Const.FLOAT16)) |
+                    ((npu_dtype == Const.TORCH_FLOAT16) & (bench_dtype == Const.TORCH_FLOAT32)) |
+                    ((npu_dtype == Const.TORCH_FLOAT32) & (bench_dtype == Const.TORCH_FLOAT16)) |
+                    ((npu_dtype == Const.TORCH_FLOAT16) & (bench_dtype == Const.TORCH_BFLOAT16)) |
+                    ((npu_dtype == Const.TORCH_BFLOAT16) & (bench_dtype == Const.TORCH_FLOAT16)))
+
+        match_result.loc[~gen_dtype_condition(), [i + '_y' for i in bench_df.columns]] = CompareConst.N_A
+        return self.make_result_df(match_result)
+
+    def modify_compare_data_with_user_mapping(self, npu_df, bench_df):
+        def get_api_indices_dict(op_name_df):
+            api_indices_dict = defaultdict(list)
+            for op_index, name in enumerate(op_name_df[CompareConst.OP_NAME]):
+                api = self.get_api_name(name.split(Const.SEP))
+                api_indices_dict[api].append(op_index)
+            return api_indices_dict
+
+        ms_api_indices_dict = get_api_indices_dict(npu_df)
+        pt_api_indices_dict = get_api_indices_dict(bench_df)
+
+        def gen_input_compare_key(pattern, term):
+            flag = True
+            for i, prefix in enumerate(mapping_dict.get(f'ms_{term}')):
+                if op_name.split(pattern)[1].startswith(str(prefix)):
+                    npu_df.loc[index, CompareConst.COMPARE_KEY] = (
+                        op_name.replace(pattern + str(prefix),
+                                        pattern + str(mapping_dict.get(f'pt_{term}')[i])))
+                    flag = False
+            return flag
+
+        for mapping_dict in self.api_mapping_dict:
+            keys_to_compare = [
+                ('ms_args', 'pt_args'),
+                ('ms_output', 'pt_output'),
+                ('ms_parameters', 'pt_parameters'),
+                ('ms_parameters_grad', 'pt_parameters_grad'),
+            ]
+            if not all(len(mapping_dict.get(k1, [])) == len(mapping_dict.get(k2, [])) for k1, k2 in keys_to_compare):
+                logger.warning('The user-defined mapping table is incorrect,\
+                                make sure that the number of parameters is equal')
+                continue
+
+            ms_api, pt_api = mapping_dict.get('ms_api'), mapping_dict.get('pt_api')
+            if ms_api not in ms_api_indices_dict or pt_api not in pt_api_indices_dict:
+                continue
+            for index in ms_api_indices_dict.get(ms_api):
+                op_name = npu_df.loc[index, CompareConst.OP_NAME].replace(ms_api, pt_api, 1)
+                if CompareConst.INPUT_PATTERN in op_name:
+                    is_abandoned = gen_input_compare_key(CompareConst.INPUT_PATTERN, 'args')
+                elif CompareConst.KWARGS_PATTERN in op_name:
+                    is_abandoned = gen_input_compare_key(CompareConst.KWARGS_PATTERN, 'args')
+                elif CompareConst.OUTPUT_PATTERN in op_name:
+                    is_abandoned = gen_input_compare_key(CompareConst.OUTPUT_PATTERN, 'output')
+                elif CompareConst.PARAMS_PATTERN in op_name:
+                    is_abandoned = gen_input_compare_key(CompareConst.PARAMS_PATTERN, 'parameters')
+                elif CompareConst.PARAMS_GRAD_PATTERN in op_name:
+                    is_abandoned = gen_input_compare_key(CompareConst.PARAMS_GRAD_PATTERN, 'parameters_grad')
+                else:
+                    logger.error(f'Excepted op_name: {op_name}')
+                    raise CompareException(CompareException.INVALID_DATA_ERROR)
+                if is_abandoned:
+                    npu_df.loc[index, CompareConst.COMPARE_KEY] = op_name + 'abandoned'
+
+    def gen_data_df(self, data_json, stack_json_data):
+        result = {
+            CompareConst.OP_NAME: [],
+            Const.DTYPE: [],
+            Const.SHAPE: [],
+            Const.SUMMARY: [],
+            'stack_info': []
+        }
+        if self.dump_mode == Const.ALL:
+            result['data_name'] = []
+        elif self.dump_mode == Const.MD5:
+            result[Const.MD5] = []
+        for data_name in data_json['data']:
+            check_op_str_pattern_valid(data_name)
+            merge_list = self.gen_merge_list(data_json, data_name, stack_json_data)
+            if not merge_list:
                 continue
-            if flag == 'forward':
-                input_args = value.get('input_args', {})
-            else:
-                input_args = value.get('input', {})
-            output_args = value.get('output', {})
-            for i in range(len(input_args)):
-                inputs_name.append(f"{key}.input.{i}")
-            for i in range(len(output_args)):
-                outputs_name.append(f"{key}.output.{i}")
-    return inputs_name, outputs_name
-
-
-def generate_file_mapping(npu_json_path, bench_json_path, mapping_list):
-
-    npu_data = load_json(npu_json_path).get("data", {})
-    bench_data = load_json(bench_json_path).get("data", {})
-
-    forward_data = []
-    mapping_list = sort_by_execution_sequence(npu_data, bench_data, mapping_list, Const.FORWARD)
-    for map_value in mapping_list:
-        npu_forward_inputs, npu_backward_outputs = generate_kernel_data(map_value[0], npu_data, "forward")
-        bench_forward_inputs, bench_backward_outputs = generate_kernel_data(map_value[1], bench_data, "forward")
-        inputs_zip = list(zip_longest(npu_forward_inputs, bench_forward_inputs))
-        outputs_zip = list(zip_longest(npu_backward_outputs, bench_backward_outputs))
-        forward_data.extend(inputs_zip)
-        forward_data.extend(outputs_zip)
-
-    backward_data = []
-    mapping_list = sort_by_execution_sequence(npu_data, bench_data, mapping_list, Const.BACKWARD)
-    for map_value in mapping_list:
-        npu_forward_inputs, npu_backward_outputs = generate_kernel_data(map_value[0], npu_data, "backward")
-        bench_forward_inputs, bench_backward_outputs = generate_kernel_data(map_value[1], bench_data, "backward")
-        inputs_zip = list(zip_longest(npu_forward_inputs, bench_forward_inputs))
-        outputs_zip = list(zip_longest(npu_backward_outputs, bench_backward_outputs))
-        backward_data.extend(inputs_zip)
-        backward_data.extend(outputs_zip)
-
-    kernel_data = forward_data + backward_data
-    result = {key: value for key, value in kernel_data if key is not None}
-
-    return result
+
+            op_name_list = merge_list.get(CompareConst.OP_NAME)
+            summary_list = merge_list.get(Const.SUMMARY)
+            data_name_list = merge_list.get('data_name')
+            op_name_reorder, summary_reorder, data_name_reorder = reorder_op_x_list(op_name_list,
+                                                                                    summary_list,
+                                                                                    data_name_list)
+            for op_name in op_name_reorder:
+                result[CompareConst.OP_NAME].append(op_name)
+                if (CompareConst.INPUT_PATTERN in op_name) or (CompareConst.KWARGS_PATTERN in op_name):
+                    struct = merge_list[CompareConst.INPUT_STRUCT].pop(0)
+                elif CompareConst.OUTPUT_PATTERN in op_name:
+                    struct = merge_list[CompareConst.OUTPUT_STRUCT].pop(0)
+                elif CompareConst.PARAMS_PATTERN in op_name:
+                    struct = merge_list[CompareConst.PARAMS_STRUCT].pop(0)
+                else:
+                    struct = merge_list[CompareConst.PARAMS_GRAD_STRUCT].pop(0)
+                result[Const.DTYPE].append(struct[0])
+                result[Const.SHAPE].append(struct[1])
+                if self.dump_mode == Const.MD5:
+                    result[Const.MD5].append(struct[2])
+                result[Const.SUMMARY].append(summary_reorder.pop(0))
+                result['stack_info'].append(merge_list['stack_info'][0] if self.stack_mode else None)
+                if self.dump_mode == Const.ALL:
+                    result['data_name'].append(data_name_reorder.pop(0))
+        return pd.DataFrame(result)
 
 
 def check_cross_framework(bench_json_path):
@@ -323,35 +392,31 @@ def check_cross_framework(bench_json_path):
 
 def ms_compare(input_param, output_path, **kwargs):
     try:
-        stack_mode = kwargs.get('stack_mode', False)
         auto_analyze = kwargs.get('auto_analyze', True)
         fuzzy_match = kwargs.get('fuzzy_match', False)
         cell_mapping = kwargs.get('cell_mapping', None)
         api_mapping = kwargs.get('api_mapping', None)
         data_mapping = kwargs.get('data_mapping', None)
         layer_mapping = kwargs.get('layer_mapping', None)
+        suffix = kwargs.get('suffix', '')
 
-        summary_compare, md5_compare = task_dumppath_get(input_param)
+        set_dump_path(input_param)
+        dump_mode = get_dump_mode(input_param)
+        if 'stack_json_path' in input_param:
+            stack_mode = kwargs.get('stack_mode', False)
+        else:
+            stack_mode = set_stack_json_path(input_param)  # set stack_mode and set "stack_json_path" in input_param
         check_configuration_param(stack_mode, auto_analyze, fuzzy_match, input_param.get('is_print_compare_log', True))
         create_directory(output_path)
-        check_compare_param(input_param, output_path, summary_compare, md5_compare)
+        check_compare_param(input_param, output_path, dump_mode, stack_mode)
     except (CompareException, FileCheckException) as error:
         logger.error('Compare failed. Please check the arguments and do it again!')
         raise CompareException(error.code) from error
     if layer_mapping:
-        pt_stack, pt_construct = struct_json_get(input_param, Const.PT_FRAMEWORK)
-        ms_stack, ms_construct = struct_json_get(input_param, Const.MS_FRAMEWORK)
-        mapping = load_yaml(layer_mapping)
-        ms_mapping_result = modify_mapping_with_stack(ms_stack, ms_construct)
-        pt_mapping_result = modify_mapping_with_stack(pt_stack, pt_construct)
-        layer_mapping = get_layer_mapping(ms_mapping_result, pt_mapping_result, mapping)
-        data_mapping = generate_file_mapping(input_param.get("npu_json_path"), input_param.get("bench_json_path"), layer_mapping)
-
-        data_mapping_name = add_time_with_yaml(f"data_mapping")
-        data_mapping_path = os.path.join(os.path.realpath(output_path), f"{data_mapping_name}")
-        save_yaml(data_mapping_path, data_mapping)
-    is_cross_framework = check_cross_framework(input_param.get("bench_json_path"))
-    ms_comparator = MSComparator(cell_mapping, api_mapping, data_mapping, is_cross_framework)
-    ms_comparator.compare_core(input_param, output_path, stack_mode=stack_mode,
-                 auto_analyze=auto_analyze, fuzzy_match=fuzzy_match, summary_compare=summary_compare,
-                 md5_compare=md5_compare)
+        data_mapping = generate_data_mapping_by_layer_mapping(input_param, layer_mapping, output_path)
+
+    mode_config = ModeConfig(stack_mode, auto_analyze, fuzzy_match, dump_mode)
+    mapping_config = MappingConfig(cell_mapping, api_mapping, data_mapping)
+    is_cross_framework = check_cross_framework(input_param.get('bench_json_path'))
+    ms_comparator = MSComparator(mode_config, mapping_config, is_cross_framework)
+    ms_comparator.compare_core(input_param, output_path, suffix=suffix)