forked from mindspore-Ecosystem/mindspore
Fix offline dbg shows null for fusion nodes using async dump data
This commit is contained in:
parent
6c20f55a51
commit
6a65e02740
|
|
@ -17,15 +17,11 @@ Module to provide conversion capabalities from .timestamp async dump files to .n
|
|||
It's an internal module for debugger backend but not exposed to users.
|
||||
"""
|
||||
import os
|
||||
import glob
|
||||
import stat
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from importlib import import_module
|
||||
from collections import namedtuple
|
||||
|
||||
import numpy as np
|
||||
|
||||
|
||||
class ConvertToolLoader:
|
||||
"""
|
||||
|
|
@ -119,7 +115,7 @@ def parse_args(file_list, output_path):
|
|||
args_dict = dict()
|
||||
args_dict['dump_version'] = '2.0'
|
||||
args_dict['format'] = 'NCHW'
|
||||
args_dict['output_file_type'] = 'npy'
|
||||
args_dict['output_file_type'] = 'msnpy'
|
||||
args_dict['dump_path'] = output_path
|
||||
args_dict['output_path'] = output_path
|
||||
args_dict['file_list'] = file_list
|
||||
|
|
@ -170,50 +166,13 @@ class AsyncDumpConverter:
|
|||
convert.check_arguments_valid()
|
||||
# 2. convert format for dump data
|
||||
ret_code = self.handle_multi_process(convert, self.files_to_convert)
|
||||
self._rename_generated_npy_files()
|
||||
if ret_code != self.convert_tool.compare_none_error:
|
||||
if os.path.exists(self.failed_file_path):
|
||||
self.convert_failed_tensors()
|
||||
self.convert_tool.log.print_info_log('An error has occurred while converting format.')
|
||||
finally:
|
||||
# clean up sys.path no matter conversion is successful or not to avoid pollution
|
||||
self.convert_tool.reset_system_path()
|
||||
self.convert_tool.log.print_info_log('Finish to convert async dump files.')
|
||||
|
||||
def convert_failed_tensors(self):
|
||||
"""
|
||||
Convert the failed tensor recorded in the failed txt file.
|
||||
"""
|
||||
self.convert_tool.log.print_info_log(
|
||||
'Start to convert failed tensors recorded in ' + self.failed_file_path + '.')
|
||||
with open(self.failed_file_path) as failed_lines:
|
||||
for failed_line in failed_lines:
|
||||
try:
|
||||
failed_line_list = failed_line.rstrip().split(',')
|
||||
self.convert_one_failed_tensor(failed_line_list)
|
||||
except (ValueError, OSError, AttributeError, self.convert_tool.compare_exception) as err:
|
||||
self.convert_tool.log.print_error_log(
|
||||
'Failed to convert ' + failed_line + ' to Host format: ' + str(err))
|
||||
|
||||
def convert_one_failed_tensor(self, failed_tensor):
|
||||
"""
|
||||
Convert failed operator one by one.
|
||||
"""
|
||||
if len(failed_tensor) <= 1:
|
||||
raise ValueError(
|
||||
"Invalid tensor info in convert_failed_file_list.txt")
|
||||
file_path = failed_tensor[0]
|
||||
type_index = failed_tensor[1:]
|
||||
op_data = self.convert_tool.utils.parse_dump_file(
|
||||
file_path, self.args.dump_version)
|
||||
for type_index_item in type_index:
|
||||
tensor_type, index = type_index_item.split(':')
|
||||
index = int(index)
|
||||
tensor = getattr(op_data, tensor_type)[index]
|
||||
dump_data_array = self.convert_tool.utils.deserialize_dump_data_to_array(tensor)
|
||||
array = dump_data_array.reshape(tensor.shape.dim)
|
||||
out_path = self._generate_path(file_path, tensor_type, index, tensor.format)
|
||||
self._save_tensor_to_npy_file(out_path, array)
|
||||
|
||||
def handle_multi_process(self, convert_obj, files):
|
||||
"""
|
||||
Convert async format files to npy in a multithreaded manner.
|
||||
|
|
@ -281,45 +240,3 @@ class AsyncDumpConverter:
|
|||
if ret_bf != self.convert_tool.compare_none_error:
|
||||
return_code = ret_bf
|
||||
return return_code
|
||||
|
||||
@staticmethod
|
||||
def _save_tensor_to_npy_file(out_path, dump_data_array):
|
||||
"""
|
||||
Save tensor file into npy format.
|
||||
"""
|
||||
np.save(out_path, dump_data_array)
|
||||
os.chmod(out_path, stat.S_IRUSR)
|
||||
|
||||
def _generate_path(self, file_path, tensor_type, idx, tensor_format):
|
||||
"""
|
||||
Generate path and filename to the target npy files
|
||||
"""
|
||||
file_name = os.path.basename(file_path)
|
||||
name_splits = file_name.split('.')
|
||||
name_splits[1] = name_splits[1].split('_')[-1]
|
||||
file_name_no_scope = '.'.join(name_splits)
|
||||
out_file_name = "%s.%s.%d.%s.npy" % (
|
||||
file_name_no_scope,
|
||||
tensor_type,
|
||||
idx,
|
||||
self.convert_tool.common.get_format_string(tensor_format)
|
||||
)
|
||||
return os.path.join(self.output_path, out_file_name)
|
||||
|
||||
def _rename_generated_npy_files(self):
|
||||
"""
|
||||
In order to follow dump naming convention, rename npy files generated by CANN conversion tool.
|
||||
"""
|
||||
target_file_list = []
|
||||
for in_file in self.files_to_convert:
|
||||
target_file_list.extend(glob.glob(in_file + "*.npy"))
|
||||
for target_file in target_file_list:
|
||||
old_filename = os.path.basename(target_file)
|
||||
name_splits = old_filename.split('.')
|
||||
name_splits[1] = name_splits[1].split('_')[-1]
|
||||
name_splits[-2] = self.args.format
|
||||
new_file_name = '.'.join(name_splits)
|
||||
out_path = os.path.join(self.output_path, new_file_name)
|
||||
os.rename(target_file, out_path)
|
||||
os.chmod(out_path, stat.S_IRUSR)
|
||||
self.convert_tool.log.print_info_log("Rename file " + target_file + " to " + out_path)
|
||||
|
|
|
|||
Loading…
Reference in New Issue