# 完整 native 对照

M1 Max / FP32 / 单线程；单位 µs。Whole/local 分别固定 local_lanes=1/8，其余编译开关相同。
时间是六轮 p50 的中位数；比值是同轮比值的中位数，不是表内时间相除。
范围是六轮观测范围，不是置信区间。所有 24 个尺寸均保留；更小的比值更好。

| 算子 | 尺寸 | Whole | Local | Inductor | Local/Inductor | 六轮范围 | 胜出轮数 | Local/Whole |
|---|---|---:|---:|---:|---:|---|---:|---:|
| rmsnorm | 17×65 | 1.147 | 0.246 | 0.730 | 0.337 | 0.337–0.340 | 6/6 | 0.215 |
| rmsnorm | 129×768 | 80.512 | 24.728 | 60.388 | 0.409 | 0.408–0.411 | 6/6 | 0.307 |
| rmsnorm | 257×1538 | 312.032 | 112.174 | 238.280 | 0.469 | 0.452–0.499 | 6/6 | 0.362 |
| rmsnorm | 1024×4097 | 3444.264 | 1158.359 | 2564.409 | 0.453 | 0.440–0.456 | 6/6 | 0.336 |
| layernorm | 17×65 | 1.392 | 0.421 | 0.574 | 0.735 | 0.730–0.741 | 6/6 | 0.303 |
| layernorm | 129×768 | 101.102 | 42.320 | 68.713 | 0.615 | 0.614–0.618 | 6/6 | 0.420 |
| layernorm | 257×1538 | 388.809 | 176.939 | 474.896 | 0.374 | 0.371–0.381 | 6/6 | 0.455 |
| layernorm | 1024×4097 | 4207.135 | 1922.582 | 5066.773 | 0.384 | 0.373–0.390 | 6/6 | 0.457 |
| masked_softmax | 17×65 | 4.291 | 3.535 | 2.199 | 1.607 | 1.604–1.618 | 0/6 | 0.823 |
| masked_softmax | 129×768 | 305.579 | 231.400 | 171.719 | 1.350 | 1.339–1.357 | 0/6 | 0.757 |
| masked_softmax | 257×1538 | 1155.162 | 1082.837 | 678.864 | 1.595 | 1.550–1.601 | 0/6 | 0.939 |
| masked_softmax | 1024×4097 | 12431.953 | 11657.000 | 7459.193 | 1.564 | 1.518–1.582 | 0/6 | 0.938 |
| swiglu | 17×65 | 3.404 | 1.624 | 1.303 | 1.244 | 1.239–1.256 | 0/6 | 0.477 |
| swiglu | 129×768 | 234.199 | 134.854 | 115.806 | 1.166 | 1.161–1.170 | 0/6 | 0.577 |
| swiglu | 257×1538 | 899.092 | 552.209 | 466.095 | 1.185 | 1.168–1.207 | 0/6 | 0.613 |
| swiglu | 1024×4097 | 9421.838 | 5769.779 | 4954.175 | 1.167 | 1.066–1.175 | 0/6 | 0.612 |
| gelu_residual | 17×65 | 5.695 | 3.032 | 4.726 | 0.645 | 0.636–0.655 | 6/6 | 0.533 |
| gelu_residual | 129×768 | 366.457 | 245.764 | 424.596 | 0.580 | 0.571–0.588 | 6/6 | 0.672 |
| gelu_residual | 257×1538 | 1417.951 | 1017.593 | 1693.660 | 0.602 | 0.583–0.607 | 6/6 | 0.717 |
| gelu_residual | 1024×4097 | 14449.375 | 10370.698 | 18308.188 | 0.568 | 0.548–0.601 | 6/6 | 0.718 |
| rope | 17×66 | 1.728 | 0.146 | 0.105 | 1.389 | 1.372–1.398 | 0/6 | 0.084 |
| rope | 129×768 | 104.219 | 28.736 | 14.794 | 1.944 | 1.904–1.968 | 0/6 | 0.275 |
| rope | 257×1538 | 404.327 | 126.292 | 70.319 | 1.795 | 1.785–1.806 | 0/6 | 0.313 |
| rope | 1024×4098 | 4743.495 | 1256.193 | 1016.388 | 1.247 | 1.144–1.350 | 0/6 | 0.268 |

源记录：`replay.json.gz`；逐次输出/guard 校验和独立复核：`audit.json`。
计时包含实际入口、共同 C++ callback、必要的 block 遍历/launch reset，以及生成代码内部的 libc/分配调用；
不包含 Runtime/Python/JIT 或 caller 的分配。不是硬件周期计数，也不是 E2E 延迟。
