UvmPinAsync / workloads /micro /run_micro_perf.py
lrh12580
first commit
5cb6c4b
import os
import argparse
import sys
import subprocess
import psutil
import os
import collections
import csv
import numpy as np
import pandas as pd
import matplotlib
from matplotlib import pyplot as plt
import matplotlib.ticker as mticker
import xlsxwriter
import seaborn as sns
from matplotlib.ticker import FormatStrFormatter
from matplotlib.legend_handler import HandlerTuple
from subprocess import Popen, PIPE
from scipy.stats import gmean
prefix = 'run_'
parameter_super_list = ['super']
config_super_list = ['standard', 'async', 'uvm', 'uvm_prefetch', 'uvm_prefetch_async']
workload_super_list = ['gemm']
def dict_to_list(input_dict):
return_list = []
for elemement in input_dict:
return_list.append(elemement)
return return_list
def addOptions(parser):
parser.add_argument("-i", "--iterations", type=int, default=1,
help="Number of iterations")
parser.add_argument("-c", "--csv", type=str, default='output.xlsx',
help="output trace log file")
parser.add_argument("-f", "--figure", type=str, default='micro',
help="output pdf file")
parser.add_argument("-p", "--profiling", action='store_true',
help="whether to run profiling or just parse results")
def get_config_list(root_directory):
config_list = []
for dict in os.listdir(root_directory):
if os.path.isdir(dict) and dict in config_super_list:
config_list.append(dict)
return config_list
def get_workload_dict(root_directory, config_list):
workload_list = []
workload_dict = dict()
for config in config_list:
config_dir = root_directory + '/' + config
for root, directories, files in os.walk(config_dir, topdown=False):
for dir in directories:
if dir in workload_super_list:
if dir not in workload_dict:
workload_dict[dir] = dict()
workload_dict[dir][config] = os.path.join(root, dir + '_perf')
if dir not in workload_list:
workload_list.append(dir)
return workload_list, workload_dict
def get_run_cmd(bash_file):
return_txt = ''
text = open(bash_file, "r")
for line in text:
return_txt += line.rstrip()
return return_txt
def execute_bashes(workload_dict, iterations, perf_list):
for workload in workload_dict:
if workload in workload_super_list:
for config in workload_dict[workload]:
if config in config_super_list:
cur_dir = workload_dict[workload][config]
pwd = os.getcwd()
os.chdir(cur_dir)
os.system('make')
for para in parameter_super_list:
for iter in range(0, iterations):
sh_file = './' + prefix + para + '.sh'
perf_cmd = ''
for i in range(0, len(perf_list)):
perf_cmd += perf_list[i]
if i != len(perf_list) - 1:
perf_cmd += ','
profile_cmd = 'ncu --metrics '
profile_cmd += perf_cmd
profile_cmd += ' --csv --log-file ' + para + '_' + str(iter) + '.profile.csv '
profile_cmd += get_run_cmd(sh_file)
os.system(profile_cmd)
os.chdir(pwd)
def main():
parser = argparse.ArgumentParser()
addOptions(parser)
options = parser.parse_args()
iterations = options.iterations
output_csv_file = options.csv
output_figure_file = options.figure
profiling = options.profiling
perf_list = []
perf_list.append('smsp__inst_executed.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_memory_pred_on.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_control_pred_on.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_integer_pred_on.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_fp16_pred_on.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_fp32_pred_on.sum')
perf_list.append('smsp__sass_thread_inst_executed_op_fp64_pred_on.sum')
perf_list.append('l1tex__t_sectors_pipe_lsu_mem_global_op_ld.sum')
perf_list.append('l1tex__t_sectors_pipe_lsu_mem_global_op_ld_lookup_hit.sum')
perf_list.append('l1tex__t_sectors_pipe_lsu_mem_global_op_st.sum')
perf_list.append('l1tex__t_sectors_pipe_lsu_mem_global_op_st_lookup_hit.sum')
root_directory = './'
config_list = get_config_list(root_directory)
print(config_list)
workload_list, workload_dict = get_workload_dict(root_directory, config_list)
print(workload_dict)
if profiling:
execute_bashes(workload_dict, iterations, perf_list)
if __name__ == '__main__':
main()