# 两组 native codegen 诊断：完整数据表

**全部计时仅供诊断，不纳入性能达标、默认策略或模型校准。** 两组实验分开执行，不能横向拼接绝对时间。

FP32、单 CPU 线程；单位 µs。每个时间是六个轮次 p50 的中位数；每个轮次含七个样本。
配对比值是六个同轮次时间比的中位数，不是显示时间的商；范围为六轮最小–最大值，不是置信区间。
`On<Off` 只是观测轮次数，不是统计显著性或已接受的胜场。完整样本与输出校验在各阶段 JSON 中。

## packet_loop_inlining

| 算子 | 尺寸 | Off µs | On µs | Inductor µs | On/Off（范围） | On<Off | On/Inductor（范围） |
|---|---|---:|---:|---:|---:|---:|---:|
| rmsnorm | 17×65 | 0.252096 | 0.248443 | 0.732910 | 0.985583 (0.980055–0.987126) | 6/6 | 0.339001 (0.337988–0.339552) |
| rmsnorm | 129×768 | 24.608165 | 24.854502 | 60.456585 | 0.990936 (0.977165–1.021529) | 4/6 | 0.409000 (0.407132–0.414664) |
| rmsnorm | 257×1538 | 109.015828 | 107.855225 | 236.232586 | 0.993596 (0.980547–0.998340) | 6/6 | 0.455741 (0.448595–0.463715) |
| rmsnorm | 1024×4097 | 1168.385406 | 1149.473953 | 2521.460938 | 0.986647 (0.969676–0.988678) | 6/6 | 0.456163 (0.455373–0.457843) |
| layernorm | 17×65 | 0.616044 | 0.559009 | 0.606978 | 0.906805 (0.905157–0.911907) | 6/6 | 0.919972 (0.915148–0.927710) |
| layernorm | 129×768 | 36.496500 | 36.346538 | 67.291301 | 0.996036 (0.990743–0.997287) | 6/6 | 0.540052 (0.537282–0.540998) |
| layernorm | 257×1538 | 176.906330 | 190.001221 | 465.223141 | 1.074727 (1.070896–1.078369) | 0/6 | 0.407700 (0.404922–0.410237) |
| layernorm | 1024×4097 | 1848.743484 | 1964.393219 | 4906.002625 | 1.065236 (1.057859–1.074113) | 0/6 | 0.400041 (0.399623–0.402919) |
| masked_softmax | 17×65 | 3.477432 | 3.414625 | 2.169772 | 0.977847 (0.962730–0.995378) | 6/6 | 1.577932 (1.549827–1.604658) |
| masked_softmax | 129×768 | 233.138023 | 230.640381 | 171.206951 | 0.988850 (0.986216–0.990361) | 6/6 | 1.339391 (1.337276–1.354267) |
| masked_softmax | 257×1538 | 1081.166672 | 1081.192703 | 677.452805 | 1.000798 (0.980690–1.006726) | 2/6 | 1.594132 (1.589351–1.613954) |
| masked_softmax | 1024×4097 | 11511.244875 | 11531.354250 | 7381.171875 | 1.000411 (0.971940–1.018263) | 3/6 | 1.562731 (1.543653–1.583960) |
| swiglu | 17×65 | 1.535465 | 1.446132 | 1.310667 | 0.942355 (0.938598–0.947259) | 6/6 | 1.103640 (1.100186–1.104649) |
| swiglu | 129×768 | 121.221926 | 121.047361 | 115.648112 | 0.998141 (0.997234–1.000540) | 5/6 | 1.042412 (1.035688–1.052597) |
| swiglu | 257×1538 | 495.576172 | 494.751305 | 470.404297 | 1.006203 (0.995182–1.012215) | 2/6 | 1.051672 (1.050007–1.062292) |
| swiglu | 1024×4097 | 5131.416750 | 5113.164062 | 4912.502625 | 0.997330 (0.992180–0.998223) | 6/6 | 1.040804 (1.039888–1.041767) |
| gelu_residual | 17×65 | 2.932181 | 2.853606 | 4.742200 | 0.973307 (0.972784–0.973865) | 6/6 | 0.601538 (0.600437–0.602104) |
| gelu_residual | 129×768 | 230.646809 | 231.054932 | 424.293621 | 1.001048 (1.000627–1.003861) | 0/6 | 0.544334 (0.534910–0.545492) |
| gelu_residual | 257×1538 | 938.003258 | 928.708328 | 1697.315766 | 0.991842 (0.988340–0.999518) | 6/6 | 0.547535 (0.539707–0.552168) |
| gelu_residual | 1024×4097 | 9699.270875 | 9863.729250 | 17756.468750 | 1.016685 (1.002557–1.022789) | 0/6 | 0.555376 (0.555065–0.565387) |
| rope | 17×66 | 0.261478 | 0.117460 | 0.113291 | 0.450091 (0.437343–0.456442) | 6/6 | 1.036088 (1.022568–1.041695) |
| rope | 129×768 | 16.271189 | 16.140564 | 14.819875 | 0.986768 (0.984361–0.998834) | 6/6 | 1.091991 (1.073936–1.102053) |
| rope | 257×1538 | 67.110637 | 67.647419 | 71.112874 | 1.007821 (1.004370–1.010937) | 0/6 | 0.952448 (0.948517–0.954389) |
| rope | 1024×4098 | 837.811836 | 833.438156 | 934.690094 | 0.994131 (0.987933–1.017544) | 5/6 | 0.888549 (0.864399–0.913687) |

## integer_lane_projection

| 算子 | 尺寸 | Off µs | On µs | Inductor µs | On/Off（范围） | On<Off | On/Inductor（范围） |
|---|---|---:|---:|---:|---:|---:|---:|
| rmsnorm | 17×65 | 0.243421 | 0.245699 | 0.726011 | 0.996945 (0.984463–1.043136) | 4/6 | 0.337791 (0.337193–0.343564) |
| rmsnorm | 129×768 | 25.180511 | 25.266896 | 61.142985 | 1.006517 (0.993203–1.010631) | 2/6 | 0.412819 (0.411755–0.416315) |
| rmsnorm | 257×1538 | 109.746947 | 110.249755 | 236.409508 | 1.005124 (0.994559–1.010910) | 1/6 | 0.466321 (0.464416–0.469805) |
| rmsnorm | 1024×4097 | 1157.927078 | 1159.555984 | 2521.154937 | 0.988876 (0.970138–1.016032) | 4/6 | 0.457335 (0.453763–0.466563) |
| layernorm | 17×65 | 0.612148 | 0.539337 | 0.569097 | 0.888309 (0.872869–0.897861) | 6/6 | 0.949230 (0.937700–0.974102) |
| layernorm | 129×768 | 37.147706 | 37.150004 | 69.241496 | 1.001981 (0.995350–1.003328) | 2/6 | 0.537121 (0.535605–0.542815) |
| layernorm | 257×1538 | 176.745604 | 155.800211 | 463.724773 | 0.884583 (0.869551–0.892077) | 6/6 | 0.335504 (0.325239–0.347657) |
| layernorm | 1024×4097 | 1895.447906 | 1658.115234 | 5246.851625 | 0.873537 (0.844216–0.894469) | 6/6 | 0.315989 (0.304670–0.319525) |
| masked_softmax | 17×65 | 3.578004 | 3.557218 | 2.214266 | 0.994486 (0.991460–0.999877) | 6/6 | 1.608751 (1.557026–1.614490) |
| masked_softmax | 129×768 | 234.711506 | 234.269127 | 172.534668 | 0.999199 (0.996848–1.003136) | 3/6 | 1.357623 (1.351477–1.366697) |
| masked_softmax | 257×1538 | 1097.651047 | 1087.765625 | 687.593742 | 0.990297 (0.987523–0.996001) | 6/6 | 1.581084 (1.576303–1.591342) |
| masked_softmax | 1024×4097 | 11487.541500 | 11379.192625 | 7383.468750 | 0.989797 (0.972727–0.994078) | 6/6 | 1.541500 (1.519808–1.548446) |
| swiglu | 17×65 | 1.558191 | 1.539029 | 1.333104 | 0.988016 (0.972854–1.001788) | 5/6 | 1.154153 (1.137155–1.157335) |
| swiglu | 129×768 | 123.877848 | 123.861490 | 119.119303 | 0.999311 (0.997550–1.002460) | 4/6 | 1.040935 (1.031702–1.042658) |
| swiglu | 257×1538 | 508.394531 | 507.136719 | 476.237633 | 1.000232 (0.996029–1.017063) | 3/6 | 1.064579 (1.051801–1.087585) |
| swiglu | 1024×4097 | 5205.570312 | 5183.145875 | 4949.697875 | 0.996854 (0.985495–0.999281) | 6/6 | 1.042093 (1.040251–1.062585) |
| gelu_residual | 17×65 | 3.025205 | 3.016801 | 4.824158 | 0.996947 (0.992950–1.000244) | 5/6 | 0.624921 (0.623907–0.625914) |
| gelu_residual | 129×768 | 234.787602 | 235.356281 | 432.189941 | 0.999206 (0.975624–1.009386) | 3/6 | 0.543818 (0.524791–0.552922) |
| gelu_residual | 257×1538 | 924.999344 | 924.253258 | 1674.292969 | 0.999634 (0.986240–1.012996) | 3/6 | 0.552091 (0.544096–0.561521) |
| gelu_residual | 1024×4097 | 9710.588625 | 9705.854250 | 17759.906500 | 0.998092 (0.996457–1.012892) | 5/6 | 0.545516 (0.526908–0.553478) |
| rope | 17×66 | 0.258619 | 0.198356 | 0.120037 | 0.767042 (0.760053–0.769907) | 6/6 | 1.652520 (1.646481–1.658297) |
| rope | 129×768 | 16.412404 | 15.054057 | 14.880544 | 0.916398 (0.915361–0.918729) | 6/6 | 1.011462 (1.009173–1.013059) |
| rope | 257×1538 | 66.495728 | 64.531250 | 71.354615 | 0.970257 (0.959751–0.986584) | 6/6 | 0.903806 (0.900169–0.920699) |
| rope | 1024×4098 | 822.063148 | 810.626953 | 945.413398 | 0.987125 (0.985173–0.989195) | 6/6 | 0.858383 (0.838527–0.903596) |