benchmark.py

import subprocess, time, os, sys
import numpy as np
sys.path.insert(0, os.path.expanduser('~/projects/machineai'))
os.chdir(os.path.expanduser('~/projects/machineai'))
from stable_baselines3 import PPO
from ir_feature_extractor import extract_features
import glob

SOURCE_FILES = glob.glob("benchmarks/*.c")
model = PPO.load("models/x86v2_base")

passes = {0:"",1:"loop-vectorize",2:"inline,loop-vectorize",
          3:"loop-unroll,loop-vectorize",4:"inline,loop-unroll,loop-vectorize",5:"loop-unroll"}

def measure(exe, n=9):
    times = []
    for _ in range(n):
        t0 = time.perf_counter()
        r = subprocess.run([exe], capture_output=True)
        t1 = time.perf_counter()
        if r.returncode == 0:
            times.append(t1 - t0)
    return float(np.median(times)) if times else 999.0

print("=" * 60)
print("  BENCHMARK RESULTS — AI vs -O1 baseline")
print("=" * 60)

total_pct = []
for src in SOURCE_FILES:
    bc = src.replace(".c", ".bc")
    subprocess.run(["clang", "-O1", "-emit-llvm", "-c", src, "-o", bc], capture_output=True)
    obs = extract_features(bc)
    action, _ = model.predict(obs, deterministic=True)

    out_bc = bc.replace(".bc", "_ai.bc")
    p = passes[int(action)]
    if p:
        subprocess.run(["opt", f"--passes={p}", bc, "-o", out_bc], capture_output=True)
    else:
        out_bc = bc

    base_exe = src.replace(".c", "_o1_exe")
    ai_exe   = src.replace(".c", "_ai_exe")
    subprocess.run(["clang", "-O1", bc,     "-o", base_exe, "-lm"], capture_output=True)
    subprocess.run(["clang", "-O1", out_bc, "-o", ai_exe,   "-lm"], capture_output=True)

    t_base = measure(base_exe)
    t_ai   = measure(ai_exe)
    pct = (t_base - t_ai) / (t_base + 1e-9) * 100
    total_pct.append(pct)

    name = os.path.basename(src)
    print(f"  {name:<20} action={action}  base={t_base*1000:.1f}ms  ai={t_ai*1000:.1f}ms  {pct:+.1f}%")

print("=" * 60)
print(f"  평균 speedup: {np.mean(total_pct):+.1f}%")
print("=" * 60)