A comparison of heuristic algorithms for custom instruction selection

Extensible processors with custom function units (CFU) that implement parts of the application code can make good trade-off between performance and flexibility. In general, deciding profitable parts of the application source code that run on CFU involves two crucial steps: subgraph enumeration and subgraph selection. In this paper, we focus on the subgraph selection problem, which has been widely recognized as a computationally difficult problem. We have formally proved that the upper bound of the number of feasible solutions for the subgraph selection problem is 3 n /3 , where n is the number of subgraph candidates. We have adapted and compared five popular heuristic algorithms: simulated annealing (SA), tabu search (TS), genetic algorithm (GA), particle swarm optimization (PSO) and ant colony optimization (ACO), for the subgraph selection problem with the objective of minimising execution time under non-overlapping constraint and acyclicity constraint. The results show that the standard SA algorithm can produce the best results while taking the least amount of time among the five standard heuristics. In addition, we have introduced an adaptive local optimum searching strategy in ACO and PSO to further improve the