Hi everyone,
We recently introduced a kernel optimization algorithm for AutoTVM ([AutoTVM] Added Droplet algorithm in TVM by canesche · Pull Request #14683 · apache/tvm · GitHub). The paper detailing the algorithm has been accepted in ACM TACO: The Droplet Search Algorithm for Kernel Scheduling | ACM Transactions on Architecture and Code Optimization. Droplet Search operates by treating the annotations of optimizations (e.g., length of the tiling window, unrolling factor, threading factor) as coordinates in the optimization space. It assumes that the unoptimized kernel and the optimal kernel exist within the same convex region of this space. Based on this observation, Droplet Search tries to find the optimal optimization parameters through the coordinate descent search technique (Coordinate descent - Wikipedia).
We are currently in the process of developing a patch to integrate Droplet Search into Ansor ([ANSOR][AUTOTVM] Combine Ansor and AutoTVM to Improve Scheduling by canesche · Pull Request #16499 · apache/tvm · GitHub).
Regards,
Michael