Optimizing Loop Partitioning in TVM

Group: Yingjing Lu, Wuwei Lin

Project Proposal

75% goal: some example loop partitioning microkernels + benckmark cost model for runtime and data reuse
100% goal: implement the loop partitioning optimization pass with a given set of optimized microkernels
125% goal: Extend this into a module in TVM tuning stack. Combine this pass with TVM IR scheduling part for cost and partitioning calculation

nearly achieved hthe 75% goal as stated in the prposal, some minor changes in the schedule and 125% goal which is updated

A demo of the result of our runtime cost model benchmarking three different simple nested loop program with the same loop structure.