mindstudio-probe 1.1.0__py3-none-any.whl → 1.1.1__py3-none-any.whl

msprobe/core/compare/acc_compare.py

@@ -15,41 +15,48 @@
 
 import multiprocessing
 import os
+import re
+from copy import deepcopy
+
 import pandas as pd
-from tqdm import tqdm
-from msprobe.core.common.file_utils import load_json
+from msprobe.core.advisor.advisor import Advisor
 from msprobe.core.common.const import CompareConst, Const
 from msprobe.core.common.exceptions import FileCheckException
-from msprobe.core.common.log import logger
-from msprobe.core.common.utils import add_time_with_xlsx, CompareException, check_op_str_pattern_valid
+from msprobe.core.common.file_utils import load_json
 from msprobe.core.common.file_utils import remove_path
+from msprobe.core.common.log import logger
+from msprobe.core.common.utils import add_time_with_xlsx, CompareException, check_op_str_pattern_valid, safe_get_value
 from msprobe.core.compare.check import check_graph_mode, check_struct_match, fuzzy_check_op, check_dump_json_str, \
-                                        check_stack_json_str
+    check_stack_json_str
 from msprobe.core.compare.highlight import find_compare_result_error_rows, highlight_rows_xlsx
-from msprobe.core.compare.utils import read_op, merge_tensor, get_un_match_accuracy, get_accuracy
 from msprobe.core.compare.multiprocessing_compute import _handle_multi_process, ComparisonResult, _save_cmp_result
 from msprobe.core.compare.npy_compare import compare_ops_apply, get_error_type, reshape_value, get_relative_err, \
     get_error_message
-from msprobe.core.advisor.advisor import Advisor
+from msprobe.core.compare.utils import read_op, merge_tensor, get_un_match_accuracy, get_accuracy, \
+    get_rela_diff_summary_mode, print_compare_ends_info
+from tqdm import tqdm
 
 
 class Comparator:
-    
+
     def __init__(self):
         pass
 
     @staticmethod
     def get_result_md5_compare(ms_op_name, bench_op_name, npu_ops_all, bench_ops_all, *args):
-        result_item = [ms_op_name, bench_op_name, npu_ops_all.get(ms_op_name).get('struct')[0],
-                       bench_ops_all.get(bench_op_name).get('struct')[0],
-                       npu_ops_all.get(ms_op_name).get('struct')[1],
-                       bench_ops_all.get(bench_op_name).get('struct')[1],
-                       npu_ops_all.get(ms_op_name).get('struct')[2],
-                       bench_ops_all.get(bench_op_name).get('struct')[2],
-                       CompareConst.PASS if npu_ops_all.get(ms_op_name).get('struct')[2]
-                                            == bench_ops_all.get(bench_op_name).get('struct')[2]
-                       else CompareConst.DIFF]
-        if args[0]:
+        npu_struct = npu_ops_all.get(ms_op_name).get('struct', [])
+        bench_struct = bench_ops_all.get(bench_op_name).get('struct', [])
+
+        if len(npu_struct) < 3 or len(bench_struct) < 3:
+            logger.error(f"The length of npu_struct and bench_struct must be >= 3, "
+                         f"but got npu_struct={len(npu_struct)} and bench_struct={len(bench_struct)}. Please check!")
+            raise CompareException(CompareException.INDEX_OUT_OF_BOUNDS_ERROR)
+
+        result_item = [ms_op_name, bench_op_name, npu_struct[0], bench_struct[0],
+                       npu_struct[1], bench_struct[1], npu_struct[2], bench_struct[2],
+                       CompareConst.PASS if npu_struct[2] == bench_struct[2] else CompareConst.DIFF]
+
+        if len(args) >= 2 and args[0]:
             result_item.extend(args[1])
         else:
             result_item.append(CompareConst.NONE)
@@ -58,59 +65,47 @@ class Comparator:
     @staticmethod
     def calculate_summary_data(npu_summary_data, bench_summary_data, result_item):
         err_msg = ""
-        start_idx = CompareConst.SUMMARY_COMPARE_RESULT_HEADER.index(CompareConst.MAX_DIFF)
-        warning_flag = False
-        for i, (npu_val, bench_val) in enumerate(zip(npu_summary_data, bench_summary_data)):
-            if isinstance(npu_val, (float, int)) and isinstance(bench_val, (float, int)):
-                diff = npu_val - bench_val
-                if bench_val != 0:
-                    relative = str(abs((diff / bench_val) * 100)) + '%'
-                else:
-                    relative = "N/A"
-                result_item[start_idx + i] = diff
-                result_item[start_idx + i + 4] = relative
-                magnitude_diff = abs(diff) / (max(abs(npu_val), abs(bench_val)) + 1e-10)
-                if magnitude_diff > 0.5:
-                    warning_flag = True
-            else:
-                result_item[start_idx + i] = CompareConst.NONE
-        accuracy_check = CompareConst.WARNING if warning_flag else ""
-        err_msg += "Need double check api accuracy." if warning_flag else ""
-        for i in range(start_idx, len(result_item)):
-            if str(result_item[i]) in ('inf', '-inf', 'nan'):
-                result_item[i] = f'{result_item[i]}\t'
+        result_item, accuracy_check, err_msg = get_rela_diff_summary_mode(result_item, npu_summary_data,
+                                                                          bench_summary_data, err_msg)
         result_item.append(accuracy_check)
         result_item.append(err_msg)
-    
+
+    @staticmethod
+    def _generate_na_data(ops_all):
+        if not ops_all:
+            return {}
+        key = next(iter(ops_all))
+        value = deepcopy(ops_all[key])
+        for k, v in value.items():
+            if isinstance(v, tuple):
+                value[k] = tuple(CompareConst.N_A for _ in range(len(v)))
+            elif isinstance(v, list):
+                value[k] = [CompareConst.N_A] * len(v)
+            else:
+                value[k] = CompareConst.N_A
+        return value
+
     @classmethod
-    def make_result_table(cls, result, md5_compare, summary_compare, stack_mode):
-        if md5_compare:
-            header = CompareConst.MD5_COMPARE_RESULT_HEADER[:]
-        elif summary_compare:
-            header = CompareConst.SUMMARY_COMPARE_RESULT_HEADER[:]
-        else:
-            header = CompareConst.COMPARE_RESULT_HEADER[:]
+    def make_result_table(cls, result, stack_mode, dump_mode):
+        header = CompareConst.HEAD_OF_COMPARE_MODE[dump_mode][:]
 
-        all_mode_bool = not (summary_compare or md5_compare)
         if stack_mode:
-            if all_mode_bool:
-                header.append(CompareConst.STACK)
+            header.append(CompareConst.STACK)
+            if dump_mode == Const.ALL:
                 header.append(CompareConst.DATA_NAME)
-            else:
-                header.append(CompareConst.STACK)
         else:
-            if all_mode_bool:
+            if dump_mode == Const.ALL:
                 for row in result:
-                    del row[-2]
+                    del row[-2]  # 输出结果不要堆栈信息时，删除中间结果result中的stack info，真实数据时为倒数第2列
                 header.append(CompareConst.DATA_NAME)
             else:
                 for row in result:
-                    del row[-1]
+                    del row[-1]  # 输出结果不要堆栈信息时，删除中间结果result中的stack info，非真实数据时为倒数第1列
         result_df = pd.DataFrame(result, columns=header, dtype='object')
-        return result_df   
-    
+        return result_df
+
     @classmethod
-    def gen_merge_list(cls, json_data, op_name, stack_json_data, summary_compare, md5_compare):
+    def gen_merge_list(cls, json_data, op_name, stack_json_data, dump_mode):
         op_data = json_data['data'][op_name]
         check_dump_json_str(op_data, op_name)
         op_parsed_list = read_op(op_data, op_name)
@@ -122,31 +117,32 @@ class Comparator:
             'full_op_name': op_name,
             'full_info': stack_info
         })
-            
-        merge_list = merge_tensor(op_parsed_list, summary_compare, md5_compare)
+
+        merge_list = merge_tensor(op_parsed_list, dump_mode)
         return merge_list
-    
+
     def check_op(self, npu_dict, bench_dict, fuzzy_match):
-        a_op_name = npu_dict["op_name"]
-        b_op_name = bench_dict["op_name"]
-        graph_mode = check_graph_mode(a_op_name[0], b_op_name[0])
-        
+        npu_op_name = npu_dict[CompareConst.OP_NAME]
+        bench_op_name = bench_dict[CompareConst.OP_NAME]
+        graph_mode = check_graph_mode(safe_get_value(npu_op_name, 0, "npu_op_name"),
+                                      safe_get_value(bench_op_name, 0, "bench_op_name"))
+
         frame_name = getattr(self, "frame_name")
         if frame_name == "PTComparator":
             from msprobe.pytorch.compare.match import graph_mapping
             if graph_mode:
-                return graph_mapping.match(a_op_name[0], b_op_name[0])
+                return graph_mapping.match(npu_op_name[0], bench_op_name[0])
         struct_match = check_struct_match(npu_dict, bench_dict)
         if not fuzzy_match:
-            return a_op_name == b_op_name and struct_match
+            return npu_op_name == bench_op_name and struct_match
         is_match = True
         try:
-            is_match = fuzzy_check_op(a_op_name, b_op_name)
+            is_match = fuzzy_check_op(npu_op_name, bench_op_name)
         except Exception as err:
-            logger.warning("%s and %s can not fuzzy match." % (a_op_name, b_op_name))
+            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 match_op(self, npu_queue, bench_queue, fuzzy_match):
         for b_index, b_op in enumerate(bench_queue[0: -1]):
             if self.check_op(npu_queue[-1], b_op, fuzzy_match):
@@ -157,8 +153,8 @@ class Comparator:
             if self.check_op(n_op, bench_queue[-1], fuzzy_match):
                 return n_index, len(bench_queue) - 1
         return -1, -1
-    
-    def compare_process(self, file_lists, stack_mode, fuzzy_match, summary_compare=False, md5_compare=False):
+
+    def compare_process(self, file_lists, stack_mode, fuzzy_match, dump_mode):
         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)
@@ -189,8 +185,7 @@ class Comparator:
                 op_name_npu = next(ops_npu_iter)
                 check_op_str_pattern_valid(op_name_npu)
                 read_err_npu = True
-                npu_merge_list = self.gen_merge_list(npu_json_data, op_name_npu, stack_json_data,
-                                                     summary_compare, md5_compare)
+                npu_merge_list = self.gen_merge_list(npu_json_data, op_name_npu, stack_json_data, dump_mode)
                 if npu_merge_list:
                     npu_ops_queue.append(npu_merge_list)
             except StopIteration:
@@ -199,8 +194,7 @@ class Comparator:
                 last_bench_ops_len = len(bench_ops_queue)
                 op_name_bench = next(ops_bench_iter)
                 check_op_str_pattern_valid(op_name_bench)
-                bench_merge_list = self.gen_merge_list(bench_json_data, op_name_bench, stack_json_data,
-                                                       summary_compare, md5_compare)
+                bench_merge_list = self.gen_merge_list(bench_json_data, op_name_bench, stack_json_data, dump_mode)
                 if bench_merge_list:
                     bench_ops_queue.append(bench_merge_list)
             except StopIteration:
@@ -226,71 +220,93 @@ class Comparator:
             b_match_data = bench_ops_queue[b_match_point]
             un_match_data = npu_ops_queue[0: n_match_point]
             for npu_data in un_match_data:
-                get_un_match_accuracy(result, npu_data, md5_compare, summary_compare)
-            get_accuracy(result, n_match_data, b_match_data, summary_compare, md5_compare)
+                get_un_match_accuracy(result, npu_data, dump_mode)
+            get_accuracy(result, n_match_data, b_match_data, dump_mode)
             del npu_ops_queue[0: n_match_point + 1]
             del bench_ops_queue[0: b_match_point + 1]
+        progress_bar.close()
         if npu_ops_queue:
             for npu_data in npu_ops_queue:
-                get_un_match_accuracy(result, npu_data, md5_compare, summary_compare)
-                
-        result_df = self.make_result_table(result, md5_compare, summary_compare, stack_mode)
+                get_un_match_accuracy(result, npu_data, dump_mode)
+
+        result_df = self.make_result_table(result, stack_mode, dump_mode)
         return result_df
 
-    def merge_data(self, json_data, stack_json_data, summary_compare, md5_compare):
+    def merge_data(self, json_data, stack_json_data, dump_mode):
         ops_all = {}
         for op_name in json_data.get('data', {}):
-            merge_list = self.gen_merge_list(json_data, op_name, stack_json_data, summary_compare,
-                                             md5_compare)
+            merge_list = self.gen_merge_list(json_data, op_name, stack_json_data, dump_mode)
             if merge_list:
                 input_index, output_index = 0, 0
-                for index, input_or_output in enumerate(merge_list['op_name']):
+                for index, input_or_output in enumerate(merge_list[CompareConst.OP_NAME]):
                     input_or_output_list = input_or_output.split(Const.SEP)
                     data_name = merge_list.get('data_name')
                     data_name = data_name[index] if data_name else None
                     if Const.INPUT in input_or_output_list or Const.KWARGS in input_or_output_list:
-                        ops_all[input_or_output] = {'struct': merge_list.get('input_struct')[input_index],
-                                                    'summary': merge_list.get('summary')[index],
-                                                    'data_name': data_name,
-                                                    'stack_info': merge_list.get('stack_info')}
+                        ops_all[input_or_output] = {
+                            CompareConst.STRUCT: safe_get_value(merge_list, input_index, "merge_list",
+                                                                key=CompareConst.INPUT_STRUCT),
+                            CompareConst.SUMMARY: safe_get_value(merge_list, index, "merge_list",
+                                                                 key=CompareConst.SUMMARY),
+                            'data_name': data_name,
+                            'stack_info': merge_list.get('stack_info')
+                        }
                         input_index += 1
 
                     elif Const.OUTPUT in input_or_output_list:
-                        ops_all[input_or_output] = {'struct': merge_list.get('output_struct')[output_index],
-                                                    'summary': merge_list.get('summary')[index],
-                                                    'data_name': data_name,
-                                                    'stack_info': merge_list.get('stack_info')}
+                        ops_all[input_or_output] = {
+                            CompareConst.STRUCT: safe_get_value(merge_list, output_index, "merge_list",
+                                                                key=CompareConst.OUTPUT_STRUCT),
+                            CompareConst.SUMMARY: safe_get_value(merge_list, index, "merge_list",
+                                                                 key=CompareConst.SUMMARY),
+                            'data_name': data_name,
+                            'stack_info': merge_list.get('stack_info')
+                        }
                         output_index += 1
         return ops_all
 
-    def get_accuracy(self, npu_ops_all, bench_ops_all, summary_compare, md5_compare):
+    def get_accuracy(self, npu_ops_all, bench_ops_all, dump_mode):
         result = []
+        bench_ops_all[CompareConst.N_A] = self._generate_na_data(bench_ops_all)
         for ms_op_name, bench_op_name in self.data_mapping_dict.items():
             if ms_op_name in npu_ops_all and bench_op_name in bench_ops_all:
                 npu_stack_info = npu_ops_all.get(ms_op_name).get("stack_info", None)
                 bench_stack_info = bench_ops_all.get(bench_op_name).get("stack_info", None)
                 has_stack = npu_stack_info and bench_stack_info
-                if md5_compare:
+                if dump_mode == Const.MD5:
                     result.append(self.get_result_md5_compare(ms_op_name, bench_op_name, npu_ops_all,
                                                               bench_ops_all, has_stack, npu_stack_info))
                     continue
-                if summary_compare:
-                    result_item = [ms_op_name, bench_op_name, npu_ops_all.get(ms_op_name).get('struct')[0],
-                                   bench_ops_all.get(bench_op_name).get('struct')[0],
-                                   npu_ops_all.get(ms_op_name).get('struct')[1],
-                                   bench_ops_all.get(bench_op_name).get('struct')[1],
-                                   " ", " ", " ", " ", " ", " ", " ", " "]
+
+                npu_struct = npu_ops_all.get(ms_op_name).get('struct', [])
+                bench_struct = bench_ops_all.get(bench_op_name).get('struct', [])
+
+                if len(npu_struct) < 2 or len(bench_struct) < 2:
+                    logger.error(
+                        f"The length of npu_struct and bench_struct must be >= 2, "
+                        f"but got npu_struct={len(npu_struct)} and bench_struct={len(bench_struct)}. "
+                        f"Please check!"
+                    )
+                    raise CompareException(CompareException.INDEX_OUT_OF_BOUNDS_ERROR)
+
+                base_result_item = [
+                    ms_op_name, bench_op_name,
+                    npu_struct[0],
+                    bench_struct[0],
+                    npu_struct[1],
+                    bench_struct[1]
+                ]
+
+                if dump_mode == Const.SUMMARY:
+                    result_item = base_result_item + [" "] * 8
                 else:
-                    result_item = [ms_op_name, bench_op_name, npu_ops_all.get(ms_op_name).get('struct')[0],
-                                   bench_ops_all.get(bench_op_name).get('struct')[0],
-                                   npu_ops_all.get(ms_op_name).get('struct')[1],
-                                   bench_ops_all.get(bench_op_name).get('struct')[1],
-                                   " ", " ", " ", " ", " "]
+                    result_item = base_result_item + [" "] * 5
+
                 npu_summary_data = npu_ops_all.get(ms_op_name).get("summary")
                 result_item.extend(npu_summary_data)
                 bench_summary_data = bench_ops_all.get(bench_op_name).get("summary")
                 result_item.extend(bench_summary_data)
-                if summary_compare:
+                if dump_mode == Const.SUMMARY:
                     self.calculate_summary_data(npu_summary_data, bench_summary_data, result_item)
                 else:
                     result_item.append(CompareConst.ACCURACY_CHECK_YES)
@@ -299,7 +315,7 @@ class Comparator:
                     result_item.extend(npu_stack_info)
                 else:
                     result_item.append(CompareConst.NONE)
-                if not (summary_compare or md5_compare):
+                if dump_mode == Const.ALL:
                     result_item.append(npu_ops_all.get(ms_op_name).get("data_name", None))
                 result.append(result_item)
             elif ms_op_name not in npu_ops_all:
@@ -308,26 +324,40 @@ class Comparator:
                 logger.warning(f'Can not find bench op name : `{bench_op_name}` in bench dump json file.')
         return result
 
-    def compare_process_custom(self, file_lists, stack_mode, summary_compare=False, md5_compare=False):
+    def compare_process_custom(self, file_lists, stack_mode, dump_mode):
         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)
 
-        npu_ops_all = self.merge_data(npu_json_data, stack_json_data, summary_compare, md5_compare)
-        bench_ops_all = self.merge_data(bench_json_data, stack_json_data, summary_compare, md5_compare)
+        npu_ops_all = self.merge_data(npu_json_data, stack_json_data, dump_mode)
+        bench_ops_all = self.merge_data(bench_json_data, stack_json_data, dump_mode)
 
-        result = self.get_accuracy(npu_ops_all, bench_ops_all, summary_compare, md5_compare)
-        result_df = self.make_result_table(result, md5_compare, summary_compare, stack_mode)
+        result = self.get_accuracy(npu_ops_all, bench_ops_all, dump_mode)
+        result_df = self.make_result_table(result, stack_mode, dump_mode)
         return result_df
 
-    def compare_by_op(self, npu_op_name, bench_op_name, op_name_mapping_dict, input_param):
+    def compare_by_op(self, npu_op_name, bench_op_name, op_name_mapping_dict, input_param, bench_data):
+        """
+        :param npu_op_name: excel中的NPU_Name，例如：MintFunctional.conv2d.0.forward.input.3.0
+        :param bench_op_name: excel中的Bench_Name，例如：Functional.conv2d.0.forward.input.3.0
+        :param op_name_mapping_dict: op_name和npy或pt文件的映射关系
+        :param input_param: npu_json_path/bench_json_path/stack_json_path等参数
+        :param bench_data: bench的dump数据中"data"字段
+        :return: result_list，包含余弦相似度、最大绝对误差、最大相对误差、千分之一误差率、千分之五误差率和错误信息
+        用于读取excel中的NPU_Name和Bench_Name，根据映射关系找到npy或pt文件，然后读取文件中的数据进行比较，计算余弦相似度、
+        最大绝对误差、最大相对误差、千分之一误差率、千分之五误差率并生成错误信息
+        """
         npu_bench_name_list = op_name_mapping_dict[npu_op_name]
-        data_name = npu_bench_name_list[1]
+        data_name = safe_get_value(npu_bench_name_list, 1, "npu_bench_name_list")
         error_file, relative_err, error_flag = None, None, False
+        bench_data_name = get_bench_data_name(bench_op_name, bench_data)
         if data_name == '-1' or data_name == -1:  # 没有真实数据路径
             n_value, b_value = CompareConst.READ_NONE, CompareConst.READ_NONE
             error_flag = True
+        elif not bench_data_name:
+            n_value, b_value, error_flag = CompareConst.READ_NONE, CompareConst.READ_NONE, True
+            error_file = 'no_bench_data'
         else:
             try:
                 read_npy_data = getattr(self, "read_npy_data")
@@ -335,19 +365,18 @@ class Comparator:
                 if frame_name == "MSComparator":
                     n_value = read_npy_data(input_param.get("npu_dump_data_dir"), npu_op_name + Const.NUMPY_SUFFIX)
                     if self.cross_frame:
-                        b_value = read_npy_data(input_param.get("bench_dump_data_dir"),
-                                                bench_op_name + Const.PT_SUFFIX, load_pt_file=True)
+                        b_value = read_npy_data(input_param.get("bench_dump_data_dir"), bench_data_name,
+                                                load_pt_file=True)
                     else:
-                        b_value = read_npy_data(input_param.get("bench_dump_data_dir"),
-                                                bench_op_name + Const.NUMPY_SUFFIX)
+                        b_value = read_npy_data(input_param.get("bench_dump_data_dir"), bench_data_name)
                 else:
                     n_value = read_npy_data(input_param.get("npu_dump_data_dir"), npu_op_name + Const.PT_SUFFIX)
-                    b_value = read_npy_data(input_param.get("bench_dump_data_dir"), bench_op_name + Const.PT_SUFFIX)
+                    b_value = read_npy_data(input_param.get("bench_dump_data_dir"), bench_data_name)
             except IOError as error:
                 error_file = error.filename
                 n_value, b_value = CompareConst.READ_NONE, CompareConst.READ_NONE
                 error_flag = True
-            except FileCheckException:
+            except (FileCheckException, CompareException):
                 error_file = data_name
                 n_value, b_value = CompareConst.READ_NONE, CompareConst.READ_NONE
                 error_flag = True
@@ -364,7 +393,7 @@ class Comparator:
             err_msg += " Fuzzy matching data, the comparison accuracy may be affected."
         result_list.append(err_msg)
         return result_list
-    
+
     def compare_core(self, input_parma, output_path, **kwargs):
         """
         Compares data from multiple JSON files and generates a comparison report.
@@ -378,8 +407,7 @@ class Comparator:
             - auto_analyze (bool, optional): If True, triggers automatic analysis after comparison. Defaults to True.
             - suffix (str, optional): Suffix to append to the output file name. Defaults to ''.
             - fuzzy_match (bool, optional): Enables fuzzy matching during comparison. Defaults to False.
-            - summary_compare (bool, optional): Enables summary comparison mode. Defaults to False.
-            - md5_compare (bool, optional): Enables MD5 comparison. Defaults to False.
+            - dump_mode (str): ALL, SUMMARY, MD5.
 
         Returns:
         """
@@ -388,41 +416,43 @@ class Comparator:
         auto_analyze = kwargs.get('auto_analyze', True)
         suffix = kwargs.get('suffix', '')
         fuzzy_match = kwargs.get('fuzzy_match', False)
-        summary_compare = kwargs.get('summary_compare', False)
-        md5_compare = kwargs.get('md5_compare', False)
+        dump_mode = kwargs.get('dump_mode', None)
 
         logger.info("Please check whether the input data belongs to you. If not, there may be security risks.")
         file_name = add_time_with_xlsx("compare_result" + suffix)
         file_path = os.path.join(os.path.realpath(output_path), file_name)
         remove_path(file_path)
-        highlight_dict = {'red_rows': [], 'yellow_rows': []}
+        highlight_dict = {"red_rows": set(), "yellow_rows": set(), "red_lines": [], "yellow_lines": []}
 
         npu_json = input_parma.get("npu_json_path")
         bench_json = input_parma.get("bench_json_path")
         stack_json = input_parma.get("stack_json_path")
         if self.data_mapping:
-            result_df = self.compare_process_custom([npu_json, bench_json, stack_json], stack_mode,
-                                                    summary_compare, md5_compare)
+            result_df = self.compare_process_custom([npu_json, bench_json, stack_json], stack_mode, dump_mode)
         else:
-            result_df = self.compare_process([npu_json, bench_json, stack_json], stack_mode, fuzzy_match,
-                                             summary_compare, md5_compare)
+            result_df = self.compare_process(
+                [npu_json, bench_json, stack_json],
+                stack_mode,
+                fuzzy_match,
+                dump_mode
+            )
 
         if not result_df.values.tolist():
             logger.warning("Can`t match any op.")
             return
 
-        if not md5_compare and not summary_compare:
-            result_df = self._do_multi_process(input_parma, result_df)
+        if dump_mode == Const.ALL:
+            result_df = self.do_multi_process(input_parma, result_df)
 
-        logger.info("Highlight suspicious API/Module start.")
-        find_compare_result_error_rows(result_df, highlight_dict, summary_compare, md5_compare)
+        find_compare_result_error_rows(result_df, highlight_dict, dump_mode)
         highlight_rows_xlsx(result_df, highlight_dict, file_path)
-        logger.info("Highlight suspicious API/Module finish.")
 
         if auto_analyze:
             advisor = Advisor(result_df, output_path, suffix)
             advisor.analysis()
-    
+
+        print_compare_ends_info()
+
     def compare_ops(self, idx, dump_path_dict, result_df, lock, input_param):
         cos_result = []
         max_err_result = []
@@ -431,13 +461,16 @@ class Comparator:
         one_thousand_err_ratio_result = []
         five_thousand_err_ratio_result = []
         is_print_compare_log = input_param.get("is_print_compare_log")
+        bench_data = load_json(input_param.get("bench_json_path")).get('data')
         for i in range(len(result_df)):
             npu_op_name = result_df.iloc[i, 0]
             bench_op_name = result_df.iloc[i, 1]
             if is_print_compare_log:
                 logger.info("start compare: {}".format(npu_op_name))
+
             cos_sim, max_abs_err, max_relative_err, one_thousand_err_ratio, five_thousand_err_ratio, err_msg = \
-                self.compare_by_op(npu_op_name, bench_op_name, dump_path_dict, input_param)
+                self.compare_by_op(npu_op_name, bench_op_name, dump_path_dict, input_param, bench_data)
+
             if is_print_compare_log:
                 logger.info(
                     "[{}] Compare result: cosine {}, max_abs_err {}, max_relative_err {}, {}, \
@@ -460,9 +493,9 @@ class Comparator:
             five_thousand_err_ratio_result=five_thousand_err_ratio_result
         )
 
-        return _save_cmp_result(idx, cr, result_df, lock)   
-    
-    def _do_multi_process(self, input_parma, result_df):
+        return _save_cmp_result(idx, cr, result_df, lock)
+
+    def do_multi_process(self, input_parma, result_df):
         try:
             result_df = _handle_multi_process(self.compare_ops, input_parma, result_df,
                                               multiprocessing.Manager().RLock())
@@ -470,4 +503,36 @@ class Comparator:
         except ValueError as e:
             logger.error('result dataframe is not found.')
             raise CompareException(CompareException.INVALID_DATA_ERROR) from e
-    
+
+def get_bench_data_name(bench_op_name, bench_data):
+    bench_name_list = re.split(r'\.(input|output|kwargs)\.', bench_op_name)
+    bench_data_bundle = bench_data.get(bench_name_list[0], {})
+    if not bench_data_bundle or len(bench_name_list) < 3:
+        return None
+    layers = bench_name_list[2].split(Const.SEP)
+
+    def get(key, container):
+        if isinstance(container, dict):
+            return container.get(key)
+        if isinstance(container, list):
+            try:
+                return container[int(key)]
+            except (ValueError, IndexError):
+                return None
+        return None
+
+    def get_by_layer(container):
+        data = container
+        for layer in layers:
+            data = get(layer, data)
+        return get(CompareConst.DATA_NAME.lower(), data)
+
+    if Const.INPUT == bench_name_list[1]:
+        return get_by_layer(bench_data_bundle.get(Const.INPUT, bench_data_bundle.get(Const.INPUT_ARGS)))
+    elif Const.KWARGS == bench_name_list[1]:
+        return get_by_layer(bench_data_bundle.get(Const.INPUT_KWARGS))
+    elif Const.OUTPUT == bench_name_list[1]:
+        return get_by_layer(bench_data_bundle.get(Const.OUTPUT))
+    else:
+        return None
+

msprobe/core/compare/check.py

@@ -35,18 +35,15 @@ dtype_mapping = {
     "BFloat16": "torch.bfloat16",
     "Complex64": "torch.complex64",
     "Complex128": "torch.complex128"
-    }
+}
 
 
-def check_struct_match(npu_dict, bench_dict, cross_frame=False):
+def check_struct_match(npu_dict, bench_dict):
     npu_struct_in = npu_dict.get("input_struct")
     bench_struct_in = bench_dict.get("input_struct")
     npu_struct_out = npu_dict.get("output_struct")
     bench_struct_out = bench_dict.get("output_struct")
 
-    if cross_frame:
-        npu_struct_in = [(dtype_mapping.get(item[0], item[0]), item[1]) for item in npu_struct_in]
-        npu_struct_out = [(dtype_mapping.get(item[0], item[0]), item[1]) for item in npu_struct_out]
     is_match = npu_struct_in == bench_struct_in and npu_struct_out == bench_struct_out
     if not is_match:
         if len(npu_struct_in) == 0 or len(bench_struct_in) == 0 or len(npu_struct_in) != len(bench_struct_in):

msprobe/core/compare/compare_cli.py

@@ -14,17 +14,22 @@
 # limitations under the License.
 
 import json
-from msprobe.core.common.file_utils import FileOpen, check_file_type
+from msprobe.core.common.file_utils import check_file_type, load_json
 from msprobe.core.common.const import FileCheckConst, Const
 from msprobe.core.common.utils import CompareException
 from msprobe.core.common.log import logger
 
 
 def compare_cli(args):
-    with FileOpen(args.input_path, "r") as file:
-        input_param = json.load(file)
+    input_param = load_json(args.input_path)
     npu_path = input_param.get("npu_path", None)
     bench_path = input_param.get("bench_path", None)
+    if not npu_path:
+        logger.error(f"Missing npu_path in configuration file {args.input_path}, please check!")
+        raise CompareException(CompareException.INVALID_PATH_ERROR)
+    if not bench_path:
+        logger.error(f"Missing bench_path in configuration file {args.input_path}, please check!")
+        raise CompareException(CompareException.INVALID_PATH_ERROR)
     frame_name = args.framework
     auto_analyze = not args.compare_only
     if frame_name == Const.PT_FRAMEWORK:
@@ -34,6 +39,9 @@ def compare_cli(args):
         from msprobe.mindspore.compare.ms_compare import ms_compare
         from msprobe.mindspore.compare.distributed_compare import ms_compare_distributed, ms_graph_compare
     if check_file_type(npu_path) == FileCheckConst.FILE and check_file_type(bench_path) == FileCheckConst.FILE:
+        if "stack_path" not in input_param:
+            logger.error(f"Missing stack_path in configuration file {args.input_path}, please check!")
+            raise CompareException(CompareException.INVALID_PATH_ERROR)
         input_param["npu_json_path"] = input_param.pop("npu_path")
         input_param["bench_json_path"] = input_param.pop("bench_path")
         input_param["stack_json_path"] = input_param.pop("stack_path")
@@ -56,7 +64,16 @@ def compare_cli(args):
 
             ms_compare(input_param, args.output_path, **kwargs)
     elif check_file_type(npu_path) == FileCheckConst.DIR and check_file_type(bench_path) == FileCheckConst.DIR:
-        kwargs = {"stack_mode": args.stack_mode, "auto_analyze": auto_analyze, "fuzzy_match": args.fuzzy_match}
+        kwargs = {
+            "stack_mode": args.stack_mode,
+            "auto_analyze": auto_analyze,
+            "fuzzy_match": args.fuzzy_match,
+            "is_print_compare_log": input_param.get("is_print_compare_log", True),
+            "cell_mapping": args.cell_mapping,
+            "api_mapping": args.api_mapping,
+            "data_mapping": args.data_mapping,
+            "layer_mapping": args.layer_mapping
+        }
         if input_param.get("rank_id") is not None:
             ms_graph_compare(input_param, args.output_path)
             return