An Investigation Into Partitioning Algorithms for Automatic Heterogeneous Compilers

Abstract

Automatic Heterogeneous Compilers allows blended hardware-software solutions to be explored without the cost of a full-fledged design team, but limited research exists on current partitioning algorithms responsible for separating hardware and software. The purpose of this thesis is to implement various partitioning algorithms onto the same automatic heterogeneous compiler platform to create an apples to apples comparison for AHC partitioning algorithms. Both estimated outcomes and actual outcomes for the solutions generated are studied and scored. The platform used to implement the algorithms is Cal Poly's own Twill compiler, created by Doug Gallatin last year. Twill's original partitioning algorithm is chosen along with two other partitioning algorithms: Tabu Search + Simulated Annealing (TSSA) and Genetic Search (GS). These algorithms are implemented inside Twill and test bench input code from the CHStone HLS Benchmark tests is used as stimulus. Along with the algorithms cost models, one key attribute of interest is queue counts generated, as the more cuts between hardware and software requires queues to pass the data between partition crossings. These high communication costs can end up damaging the heterogeneous solution's performance. The Genetic, TSSA, and Twill's original partitioning algorithm are all scored against each other's cost models as well, combining the fitness and performance cost models with queue counts to evaluate each partitioning algorithm. The solutions generated by TSSA are rated as better by both the cost model for the TSSA algorithm and the cost model for the Genetic algorithm while producing low queue counts. v ACKNOWLEDGMENTS Through the five years I've spent at Cal Poly I have countless people to thank. Firstly, Iwant to thank Dr. Danowitz for taking me under as one of his first graduate students and Dr. Oliver for pointing me towards the master's degree. Also I would like to thank Toby