Auto tuning of Linux RK3399 is useless


#1

I used the auto tuning to optimize the performance of mobilenetv2-1.0-224x224 in Linux RK3399, and the time is 62ms without auto tuning and 64ms with auto tuning.

In my case, the target = tvm.target.arm_cpu(“rk3399”) and other parts are same as the tutarials.

How do I debug this problem?


#2

This is expected behavior and not a bug, as tophub includes pretuned schedules for that particular workload and the RK3399. Your result is very close to the published blog post number as well: https://tvm.ai/2018/10/03/auto-opt-all.html.