Is it possibile that by using AUTOTvm i get a module that performe worse with respect the not tuned model?
the AutoTVM builds configurations for kernels from hand-made schedules from python/tvm/topi. At the same time, the default configuration is from the same search space as AutoTVM. I.e. AutoTVM has an ability to find the configuration used by default compilation.
On the other hand, It is not pointed in the AutoTVM algorithm to use this default point anyhow. The first configurations are selected randomly and then are tried to be improved using xgboost mechanism or genetic algorithms.
The default configuration is defined by the developer. We can hope that this configuration is selected by developer based on analysis how kernel works with different parameters on different hardware but by fact we cannot guarantee that default point is selected the best way because it is not standardized, depends on the developer and effort which he spends to find such point.
Return to your use case - it seems the developer tried to find good default configuration. If we remeasure all configurations for layer by AutoTVM from search space (that can be quite huge, sometimes it has millions configurations), the default point will be definitely in measured. But it is very unlikely that it will be selected as starting point for tuning.
This issue should be addressed in the next generation of the tuner - Meta Scheduler, but it is mostly under development. MetaScheduler should include default point of the handmade schedule in the initial measurement