Performance question about Auto tuning on x86 CPU

Development
1

Hello, I am trying to replicate the performance that is shown in this tutorial: https://tvm.apache.org/docs/tutorials/autotvm/tune_relay_x86.html

The problem is that I am not able to get quite as good performance. I did some digging and set the target that some others did for the similar CPU (i am using i7-8665U) but I am not getting as good of optimization can someone help me?

I am using target as “llvm -mcpu=core-avx2”

This is the autotuned performance output, as you can see I get 29ms but in the tutorial they get 3ms. Everything else is the same as the tutorial.

29.66ms
29.66ms1771×468 84.8 KB

2

The final performance may vary depending on your specific hardware. 3ms is just for reference.

3

Thank you for your replay, The only thing is that it’s not doing better than non tvm optimized model. I am getting 8 to 10ms without using tvm at all. is that okay for something like resnet-18 ?

4

Are you using single or multiple cores?

5

I am using all 4 cores.

6

I think there may be some issues with the auto-tune script. It will always use the first config and not reset even you provide the generated log file.

You can try to run the resnet model using the autotune log separately after you finished the autotune. The result will look better.