def _analyser_op_info(self):
"""Analyse the operator information."""
integrator = Integrator(self._output_path, self._dev_id)
integrator.integrate()
aicore_type_result = self._query_op_type_info()
detail_file_path = os.path.join(
self._output_path,
'output_op_compute_time_detail_{}.txt'.format(self._dev_id))
fwrite_format(detail_file_path, data_source='title:op compute time')
display_names = [
'optype_name', 'compute_time(ms, per-step)',
'called_times(per-step)', 'percent'
]
fwrite_format(detail_file_path,
data_source=" ".join(display_names),
is_print=True)
fwrite_format(detail_file_path,
data_source=aicore_type_result,
is_print=True)
op_type_order = [item[0] for item in aicore_type_result]
aicore_detail_result = self._query_op_detail_info(op_type_order)
fwrite_format(detail_file_path, data_source='', is_print=True)
fwrite_format(detail_file_path, data_source='Detail:', is_print=True)
fwrite_format(detail_file_path,
data_source=" ".join(
aicore_detail_result.get('col_name_detail')),
is_print=True)
fwrite_format(detail_file_path,
data_source=aicore_detail_result.get('object'),
is_print=True)
def _analyse_memory_usage(self, points):
"""Analyse memory usage data."""
integrator = Integrator(self._output_path, self._dev_id)
aicore_detail_data = integrator.get_aicore_detail_data()
memory_parser = MemoryUsageParser(self._output_path, self._dev_id)
memory_parser.init_memory_usage_info(aicore_detail_data, points)
memory_parser.write_memory_files()
def _query_op_detail_info(self, op_type_order):
"""
Query AICORE operator detail information.
Args:
op_type_order(list): The name of the op type in order.
Returns:
dict, the AICORE operator detail information.
"""
op_type_condition = {}
if self._valid_optype_name:
op_type_condition['in'] = self._valid_optype_name
if self._filt_optype_names:
op_type_condition['not_in'] = self._filt_optype_names
subgraph_condition = {}
if self._subgraph != 'all':
subgraph_condition['in'] = [self._subgraph]
filter_condition = {
'op_type': op_type_condition,
'subgraph': subgraph_condition,
'is_display_detail': False,
'is_display_full_op_name': self._withfullpath
}
integrator = Integrator(self._output_path, self._dev_id)
return integrator.query_and_sort_by_op_type(filter_condition,
op_type_order)
def _query_op_type_info(self):
"""
Query AICORE operator type information.
Returns:
list[list], the AICORE operator type and execution time information.
"""
integrator = Integrator(self._output_path, self._dev_id)
return integrator.get_aicore_data()
def _analyse_timeline(self, aicpu_parser, optime_parser, source_path):
"""
Analyse and parse timeline info.
Args:
aicpu_parser (DataPreProcessParser): The parser instance for AI CPU operator
execution time calculation.
optime_parser (OPComputeTimeParserParser): The parser instance for AI Core
operator execution time calculation.
"""
timeline_analyser = AscendTimelineGenerator(self._output_path,
self._dev_id)
# Get framework info
integrator = Integrator(self._output_path, self._dev_id)
aicore_detail_data = integrator.get_aicore_detail_data()
aicore_detail_data_size = len(aicore_detail_data)
col_names = [
'op_name', 'op_type', 'avg_execution_time', 'subgraph',
'full_op_name', 'op_info'
]
framework_info = {
'col_name': col_names,
'object': aicore_detail_data,
'size': aicore_detail_data_size
}
all_reduce_info = integrator.query_for_all_reduce()
# Get timeline info
logger.info('Start writing timeline info...')
logger.info(
'Warm Prompt: It could take a few minutes if you are training '
'with a complex network or more than 10 steps.')
# Add info into timeline, such as AI CPU, AllReduce, framework info.
aicpu_info = aicpu_parser.query_aicpu_data()
min_cycle_counter = min(aicpu_parser.min_cycle_counter,
optime_parser.min_cycle_counter)
timeline_analyser.init_timeline(all_reduce_info, framework_info,
aicpu_info, min_cycle_counter,
source_path)
size_limit = 20 * 1024 * 1024 # 20MB
timeline_analyser.write_timeline(size_limit)
timeline_analyser.write_timeline_summary()
def _query_op_detail_info(self, op_type_order):
"""
Query AICORE operator detail information.
Args:
op_type_order(list): The name of the op type in order.
Returns:
dict, the AICORE operator detail information.
"""
op_type_condition = {}
if self._filt_optype_names:
op_type_condition['not_in'] = self._filt_optype_names
filter_condition = {
'op_type': op_type_condition,
'is_display_detail': False,
}
integrator = Integrator(self._output_path, self._dev_id)
return integrator.query_and_sort_by_op_type(filter_condition, op_type_order)
def _query_op_detail_info(self, op_type_order):
"""
Query AICORE operator detail information.
Args:
op_type_order(list): The name of the op type in order.
Returns:
dict, the AICORE operator detail information.
"""
op_type_condition = {}
if self._valid_optype_name:
op_type_condition['in'] = self._valid_optype_name
if self._filt_optype_names:
op_type_condition['not_in'] = self._filt_optype_names
subgraph_condition = {}
if self._subgraph != 'all':
subgraph_condition['in'] = [self._subgraph]
integrator = Integrator(self._output_path, self._dev_id)
return integrator.get_aicore_detail_data()
