import os
import pandas as pd
import numpy as np
WARM_UP_TIMES = 10
PERF_TEST_CYCLE_TIMES = 40
def open_input_file(input_file):
df = pd.read_csv(input_file)
return df
def get_time_data(df, test_line_num: int):
df = df[df['kernel_type'].astype(str).str.contains("CORE|AIV", na=False)]
df = df.reset_index(drop=True)
time_data = []
total_rows = len(df)
data_rows = total_rows // test_line_num
coc_tiling_num = (data_rows - WARM_UP_TIMES) // PERF_TEST_CYCLE_TIMES
for i in range(test_line_num):
start_row = i * data_rows + WARM_UP_TIMES
for j in range(coc_tiling_num):
current_row = start_row + j * PERF_TEST_CYCLE_TIMES
group = df.iloc[current_row: current_row + PERF_TEST_CYCLE_TIMES]["task_time(us)"]
avg_value = group.mean()
time_data.append(avg_value)
return time_data
def get_time_file(path):
for data_path in os.listdir(path):
if os.path.basename(data_path) != "mindstudio_profiler_output":
continue
profiler_path = os.path.join(path, data_path)
for f in os.listdir(profiler_path):
if os.path.basename(f)[:9] == "task_time":
res = os.path.join(profiler_path, f)
return res
return ""
def get_pref_path(path):
perf_list = list(filter(lambda item: item.startswith("PROF"), os.listdir(path)))
task_time_list = []
for perf_dir in perf_list:
perf_path = os.path.join(path, perf_dir)
time_file = get_time_file(perf_path)
task_time_list.append(time_file)
return task_time_list
def process_kernel_data():
script_dir = os.path.dirname(os.path.abspath(__file__))
parent_dir = os.path.dirname(script_dir)
tiling_df = open_input_file(os.path.join(parent_dir, "results.csv"))
print(tiling_df)
pref_file_list = get_pref_path(os.path.join(parent_dir, "output"))
print(pref_file_list)
case_num = len(tiling_df)
perf_output = np.zeros((case_num, )).astype(np.float32)
for pref_file in pref_file_list:
pref_df = open_input_file(pref_file)
pref_data = get_time_data(pref_df, case_num)
perf_output += np.array(pref_data).astype(np.float32)
perf_output = perf_output / len(pref_file_list)
tiling_df['Time(us)'] = perf_output
tiling_df.to_csv(os.path.join(parent_dir, "result.csv"), index=False)
if __name__ == '__main__':
process_kernel_data()