cpython/Tools/peg_generator/scripts/benchmark.py

103 lines
2.5 KiB
Python

#!/usr/bin/env python3
import argparse
import ast
import sys
import os
from time import time
try:
import memory_profiler
except ModuleNotFoundError:
print(
"Please run `make venv` to create a virtual environment and install"
" all the dependencies, before running this script."
)
sys.exit(1)
sys.path.insert(0, os.getcwd())
from scripts.test_parse_directory import parse_directory
argparser = argparse.ArgumentParser(
prog="benchmark", description="Reproduce the various pegen benchmarks"
)
argparser.add_argument(
"--target",
action="store",
choices=["xxl", "stdlib"],
default="xxl",
help="Which target to use for the benchmark (default is xxl.py)",
)
subcommands = argparser.add_subparsers(title="Benchmarks", dest="subcommand")
command_compile = subcommands.add_parser(
"compile", help="Benchmark parsing and compiling to bytecode"
)
command_parse = subcommands.add_parser("parse", help="Benchmark parsing and generating an ast.AST")
def benchmark(func):
def wrapper(*args):
times = list()
for _ in range(3):
start = time()
result = func(*args)
end = time()
times.append(end - start)
memory = memory_profiler.memory_usage((func, args))
print(f"{func.__name__}")
print(f"\tTime: {sum(times)/3:.3f} seconds on an average of 3 runs")
print(f"\tMemory: {max(memory)} MiB on an average of 3 runs")
return result
return wrapper
@benchmark
def time_compile(source):
return compile(source, "<string>", "exec")
@benchmark
def time_parse(source):
return ast.parse(source)
def run_benchmark_xxl(subcommand, source):
if subcommand == "compile":
time_compile(source)
elif subcommand == "parse":
time_parse(source)
def run_benchmark_stdlib(subcommand):
modes = {"compile": 2, "parse": 1}
for _ in range(3):
parse_directory(
"../../Lib",
verbose=False,
excluded_files=["*/bad*", "*/lib2to3/tests/data/*",],
short=True,
mode=modes[subcommand],
)
def main():
args = argparser.parse_args()
subcommand = args.subcommand
target = args.target
if subcommand is None:
argparser.error("A benchmark to run is required")
if target == "xxl":
with open(os.path.join("data", "xxl.py"), "r") as f:
source = f.read()
run_benchmark_xxl(subcommand, source)
elif target == "stdlib":
run_benchmark_stdlib(subcommand)
if __name__ == "__main__":
main()