mberenjk
519843d2cf
* adding rocprof parser script * adding the support for multiple json files * adding pytorch profiler script * remove filtering from pytorch log * adding the addressing the comments and add the feature to parse all kernels * completing the report for torch profiler --------- Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>
119 Zeilen
4.4 KiB
Python
119 Zeilen
4.4 KiB
Python
import json
|
|
import argparse
|
|
import os
|
|
import csv
|
|
import sys
|
|
|
|
@staticmethod
|
|
def get_num_gpu():
|
|
return 8
|
|
def load_json_files(directory):
|
|
json_data = {}
|
|
for root, _, files in os.walk(directory):
|
|
for file in files:
|
|
if file.endswith('json'):
|
|
with open(directory+file, 'r') as f:
|
|
data = json.load(f)
|
|
json_data.setdefault('traceEvents',[]).append(data['traceEvents'])
|
|
return json_data
|
|
|
|
|
|
def parse(json_file_path, output_file_name, function_name):
|
|
|
|
data_all = load_json_files(json_file_path)
|
|
data = data_all['traceEvents']
|
|
|
|
kernels = []
|
|
found = False
|
|
for entries in data:
|
|
for entry in entries:
|
|
if 'name'in entry and 'cat' in entry and (entry['cat'] == 'kernel' ):
|
|
if function_name == 'all':
|
|
kernels.append(entry)
|
|
found = True
|
|
elif function_name in entry['name']:
|
|
kernels.append(entry)
|
|
found = True
|
|
if not found:
|
|
print('There is no ' + function_name +' in this log')
|
|
return
|
|
|
|
|
|
sorted_kernels = sorted(kernels, key=lambda x: ( x['ts'], x['pid']))
|
|
|
|
csv_file_name = output_file_name + '.csv'
|
|
json_file_out = output_file_name + '.json'
|
|
|
|
json_data_out = {}
|
|
json_data_out.setdefault('traceEvents',[]).append({})
|
|
|
|
with open(csv_file_name, 'w', newline='') as csvfile:
|
|
fieldnames = ['pid', 'dur', 'ts', 'min_dur', 'max_dur', 'min_start', 'max_start', 'latency_before_first_gpu', 'max_dur - min_dur', 'duration_from_last_arrival', 'first_gpu', 'last_gpu', 'shortest_gpu', 'longest_gpu']
|
|
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
|
|
writer.writeheader()
|
|
i = 1
|
|
min_dur = sys.float_info.max
|
|
max_dur = sys.float_info.min
|
|
min_start = sys.float_info.max
|
|
max_start = sys.float_info.min
|
|
first_gpu = 0
|
|
last_gpu = 0
|
|
longest_gpu = 0
|
|
shortest_gpu = 0
|
|
for entry in sorted_kernels:
|
|
record = {'pid': entry['pid'], 'dur': entry['dur'],
|
|
'ts': entry['ts']}
|
|
json_data_out.setdefault('traceEvents',[]).append(entry)
|
|
if entry['dur'] < min_dur :
|
|
min_dur = min(min_dur, entry['dur'])
|
|
shortest_gpu = entry['pid']
|
|
|
|
if entry['dur'] > max_dur :
|
|
max_dur = max(max_dur, entry['dur'])
|
|
longest_gpu = entry['pid']
|
|
|
|
if entry['ts'] < min_start :
|
|
min_start = min(min_start, entry['ts'])
|
|
first_gpu = entry['pid']
|
|
|
|
if entry['ts'] > max_start:
|
|
max_start = max(max_start, entry['ts'])
|
|
duration_from_last_arrival = entry['dur']
|
|
last_gpu = entry['pid']
|
|
|
|
writer.writerow(record)
|
|
if (i) % get_num_gpu() == 0:
|
|
record = {'min_dur': min_dur, 'max_dur':max_dur, 'min_start':min_start, 'max_start':max_start,'latency_before_first_gpu':max_start-min_start, 'max_dur - min_dur':max_dur-min_dur , 'duration_from_last_arrival':duration_from_last_arrival , 'first_gpu': first_gpu, 'last_gpu':last_gpu, 'shortest_gpu':shortest_gpu, 'longest_gpu':longest_gpu}
|
|
writer.writerow(record)
|
|
csvfile.write('\n')
|
|
min_dur = sys.float_info.max
|
|
max_dur = sys.float_info.min
|
|
min_start = sys.float_info.max
|
|
max_start = sys.float_info.min
|
|
first_gpu = 0
|
|
last_gpu = 0
|
|
longest_gpu = 0
|
|
shortest_gpu = 0
|
|
|
|
i = 0
|
|
i = i + 1
|
|
|
|
with open(json_file_out, 'w') as jsonfileout:
|
|
json.dump(json_data_out, jsonfileout, indent=4)
|
|
|
|
print(f"Data successfully written to {csv_file_name} and {json_file_out}.")
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description='Json file and the function to parse.')
|
|
|
|
parser.add_argument('json_file_path', metavar='file_path', type=str, help='Path to the JSON file to process')
|
|
parser.add_argument('output_file_name', type=str, help='Output File Name')
|
|
parser.add_argument('function_name', type=str, help='Kernel Function Name, e.g., oneShotAllReduce, ncclDevKernel_Generic, mscclKernel')
|
|
|
|
args = parser.parse_args()
|
|
parse(args.json_file_path, args.output_file_name, args.function_name)
|
|
|
|
if __name__ == '__main__':
|
|
main()
|
|
|