Have you found the cause of the problem？
What else information do you need to locate and solve the problem?
What is the cost of the best cfg for every depthwise conv2d? If those are similar to the actual execution time shown in debugger, it means auto-tuning depthwise conv2d didn’t achieve descent performance.
The problem is the graph tuner don’t work for this model, and when i set opt_level=3 the time is 2x slower than opt_level=1 or 2. And the modified mobilefacenet is 80ms using apply_history_best on opt_level=2, the ori mobilefacenet is 12ms, it is unreasonable due to the model complexity.
Opt level has no direct relation with graph tuner. Now the issue is depthwise conv2d is slow. Then does this come from autotvm, or other stuff such as layout transform? A simple way to verify is to check the best config cost of each workload in the auto tuned log file.