#!/usr/bin/env python3 """the benchmark suite behind `make bench`, `make bench-save` and `make bench-check`. usage: tests/bench/suite.py run [--zy PATH] [--out FILE] tests/bench/suite.py check [--zy PATH] [--baseline FILE] [--out FILE] each program is a kernel of tests/programs/bench.zc (the kernels bench.sh times, plain and o3) or of tests/bench/sweep.zc (a 256-word pointer sweep, checked and unbounded), run on one core (build/bin/zy by default, the host's jit where there is one). a kernel is selected by a letter and a repetition count on stdin, after the source. the time per repetition is the slope between 1 and N repetitions, so starting the vm and compiling the program cancel out; N grows until the N run takes at least MIN_S seconds. each cell is the median of TRIALS slopes. columns: ms per repetition; checks, the number of bounds checks the compiler keeps (from --explain-checks, which does not exist yet: '-'); peak heap words and gc mark time, which the runtime does not expose: '-'. check compares with the baseline (tests/bench.baseline): it fails when a program is more than TOLERANCE (10%) slower, or when both sides have a checks count and it rose. a program that looks slower is measured again, up to RETRIES more times, and its best median is kept, so one noisy trial does not fail the check. """ import argparse import os import statistics import subprocess import sys import time from pathlib import Path ROOT = Path(__file__).resolve().parent.parent.parent TRIALS = int(os.environ.get("TRIALS", "5")) MIN_S = float(os.environ.get("MIN_S", "0.4")) TOLERANCE = float(os.environ.get("TOLERANCE", "0.10")) RETRIES = int(os.environ.get("RETRIES", "2")) # name, source, kernel letter PROGRAMS = [ ("collatz", "tests/programs/bench.zc", "a"), ("collatz.o3", "tests/programs/bench.zc", "b"), ("sieve", "tests/programs/bench.zc", "c"), ("sieve.o3", "tests/programs/bench.zc", "d"), ("digits", "tests/programs/bench.zc", "e"), ("digits.o3", "tests/programs/bench.zc", "f"), ("fib", "tests/programs/bench.zc", "g"), ("fib.o3", "tests/programs/bench.zc", "h"), ("sweep.checked", "tests/bench/sweep.zc", "c"), ("sweep.unbounded", "tests/bench/sweep.zc", "u"), ] COLUMNS = ["program", "ms", "checks", "peak_heap_words", "gc_mark_ms"] def run_once(zy, src, letter, reps): data = src + b"`" + letter.encode() + str(reps).encode() t = time.perf_counter() r = subprocess.run([zy], input=data, stdout=subprocess.PIPE, stderr=subprocess.PIPE) dt = time.perf_counter() - t if r.returncode != 0: sys.exit("bench: %s %s failed: %s" % (zy, letter, r.stderr.decode(errors="replace")[:300])) return dt def slope_ms(zy, path, letter): src = (ROOT / path).read_bytes() n = 4 while True: one = min(run_once(zy, src, letter, 1) for _ in range(2)) many = run_once(zy, src, letter, n) if many - one >= MIN_S or n >= 1 << 16: break n *= 4 slopes = [] for _ in range(TRIALS): one = run_once(zy, src, letter, 1) many = run_once(zy, src, letter, n) slopes.append((many - one) / (n - 1) * 1000) return statistics.median(slopes) def measure(zy, name): for p, path, letter in PROGRAMS: if p == name: return slope_ms(zy, path, letter) raise KeyError(name) def fmt_row(cells): return "%-16s %10s %8s %16s %11s" % tuple(cells) def write_table(path, rows): lines = ["# tinyvm bench: ms per repetition (median of %d slopes); '-' where a figure is not available" % TRIALS, fmt_row(COLUMNS)] for r in rows: lines.append(fmt_row([r["program"], "%.3f" % r["ms"], r["checks"], r["peak"], r["gc"]])) path.parent.mkdir(parents=True, exist_ok=True) path.write_text("\n".join(lines) + "\n") def read_table(path): rows = {} for line in path.read_text().splitlines(): if not line.strip() or line.startswith("#") or line.split()[0] == "program": continue w = line.split() rows[w[0]] = {"program": w[0], "ms": float(w[1]), "checks": w[2], "peak": w[3], "gc": w[4]} return rows def run_all(zy): rows = [] for name, _, _ in PROGRAMS: ms = measure(zy, name) print("bench: %-16s %10.3f ms" % (name, ms), file=sys.stderr, flush=True) rows.append({"program": name, "ms": ms, "checks": "-", "peak": "-", "gc": "-"}) return rows def main(): ap = argparse.ArgumentParser(description="tinyvm benchmark suite") ap.add_argument("mode", choices=["run", "check"]) ap.add_argument("--zy", default=str(ROOT / "build/bin/zy")) ap.add_argument("--out", default=str(ROOT / "build/bench.txt")) ap.add_argument("--baseline", default=str(ROOT / "tests/bench.baseline")) a = ap.parse_args() rows = run_all(a.zy) if a.mode == "run": write_table(Path(a.out), rows) print(Path(a.out).read_text(), end="") return 0 base_path = Path(a.baseline) if not base_path.exists(): sys.exit("bench: no baseline at %s (run make bench-save)" % base_path) base = read_table(base_path) bad = 0 print("%-16s %10s %10s %8s %9s %9s %s" % ("program", "base ms", "now ms", "change", "base chk", "now chk", "verdict")) for r in rows: b = base.get(r["program"]) if not b: print("%-16s %10s %10.3f %8s %9s %9s new" % (r["program"], "-", r["ms"], "-", "-", r["checks"])) continue tries = 0 while r["ms"] > b["ms"] * (1 + TOLERANCE) and tries < RETRIES: r["ms"] = min(r["ms"], measure(a.zy, r["program"])) tries += 1 change = (r["ms"] - b["ms"]) / b["ms"] * 100 if b["ms"] > 0 else 0.0 verdict = "ok" if r["ms"] > b["ms"] * (1 + TOLERANCE): verdict = "SLOWER" if b["checks"] != "-" and r["checks"] != "-" and int(r["checks"]) > int(b["checks"]): verdict = "MORE CHECKS" if verdict == "ok" else verdict + ", MORE CHECKS" if verdict != "ok": bad += 1 print("%-16s %10.3f %10.3f %+7.1f%% %9s %9s %s" % (r["program"], b["ms"], r["ms"], change, b["checks"], r["checks"], verdict)) for name in base: if name not in {r["program"] for r in rows}: print("%-16s %10.3f %10s %8s %9s %9s gone" % (name, base[name]["ms"], "-", "-", base[name]["checks"], "-")) write_table(Path(a.out), rows) print("bench-check: %s (%d of %d programs over the limit; tolerance %d%%)" % ("FAIL" if bad else "ok", bad, len(rows), round(TOLERANCE * 100))) return 1 if bad else 0 if __name__ == "__main__": sys.exit(main())