#!/usr/bin/env python3 """numpy/nax 参考实现的微基准,与 `crates/nax-core/examples/bench.rs` 对齐。 同一个机器、同样规模、同样算子,唯一变量是"numpy 的 u64 路径" vs "Rust 的 u64 路径"。这个对比的意义在于:ML 框架里所有性能优化都建立在 浮点 BLAS 上,而 MPC 用的是 **u64 环算术** —— BLAS 完全帮不上忙, numpy 只能走通用迭代器循环。这正是 Rust 算子核最大的价值来源。 用法: .venv/bin/python tools/bench_numpy.py """ from __future__ import annotations import importlib.util import pathlib import time import numpy as np ROOT = pathlib.Path(__file__).resolve().parent.parent REF_RING = ROOT / "refs" / "nax" / "nax" / "crypto" / "ring.py" def load_ring(): spec = importlib.util.spec_from_file_location("ref_ring", REF_RING) mod = importlib.util.module_from_spec(spec) assert spec.loader is not None spec.loader.exec_module(mod) return mod def bench(label: str, iters: int, fn) -> float: fn() # warmup best = float("inf") for _ in range(iters): t = time.perf_counter() fn() dt = time.perf_counter() - t best = min(best, dt) print(f"{label:<42} {best * 1e3:>10.3f} ms") return best * 1e3 def main() -> int: ring = load_ring() rng = np.random.default_rng(0xC0FFEE) print("numpy 参考实现微基准") print(f"numpy {np.__version__}") print("=" * 30) print() print("[矩阵乘 u64 @ u64(mod 2^64)]") for n in (64, 128, 256): a = rng.integers(0, 1 << 63, size=(n, n), dtype=np.uint64) b = rng.integers(0, 1 << 63, size=(n, n), dtype=np.uint64) iters = 20 if n <= 64 else (5 if n <= 128 else 2) with np.errstate(over="ignore"): ms = bench(f" matmul {n}x{n}x{n}", iters, lambda: ring.matmul(a, b)) ops = n**3 print(f" -> {ops / (ms / 1e3) / 1e9:.2f} Gop/s") print() print("[逐元素算子,16M 元素]") length = 16 * 1024 * 1024 a = rng.integers(0, 1 << 63, size=length, dtype=np.uint64) b = rng.integers(0, 1 << 63, size=length, dtype=np.uint64) for name, fn in ( ("mul (16M)", lambda: ring.mul(a, b)), ("add (16M)", lambda: ring.add(a, b)), ("arshift 18 (16M)", lambda: ring.arshift(a, 18)), ): with np.errstate(over="ignore"): ms = bench(f" {name}", 5, fn) print(f" -> {length / (ms / 1e3) / 1e9:.2f} Gop/s") return 0 if __name__ == "__main__": raise SystemExit(main())