[Question] After auto tune, the speed of resnest101 and resnet101 have big difference

I use tvm to speed resnet101 and resnest101 which are provided by mxnet . But after auto tuning the model, the resnet101 is more faster, but resnest101 is slower. To find the problem, I print the debug log to see the time of each layer , I found that although the same layer , time has a big difference. The first layer ,resnet cost 0.75ms, but resnest cost 37ms . The struct of the first layer just [conv + bn + relu] , I don’t know why resnest cost 37ms. who knows?

What is your target device?

I auto tune on a 1080ti