Thanks for the detailed (and interesting) answer.
So, right: the sketch defines the dag, and then different states share the same (initial) dag. But different sketches have different dags. Is this correct?
A very interesting thing you said: “The autoscheduler infers a good layout from the schedule and does the layout rewrite automatically”. Let’s say I have a GEMM computation, and then I want to pack my matrix A to have a more friendly data-layout (see this RFC, for instance). Can I inject in the autoscheduler my data-packing?
As for debugging further, once I have autoscheduler in tvmc I will send you the tvmc command so that we can start from a common ground. Meanwhile, I will try to put some printfs in the code to see where the error is showing.
Thanks,