GA driven integrated exploration of loop unrolling factor and datapath for optimal scheduling of CDFGs during high level synthesis

Abstract

A novel solution to the problem of integrated exploration of datapath resource configuration and loop unrolling factor (UF) for control data flow graphs (CDFGs) during high level synthesis (HLS) for optimal scheduling is presented in this paper. The proposed approach is fast as it is driven through an adaptive genetic algorithm (GA) process, capable of escaping local minima and an estimation model for single loops that determines the total delay without tediously unrolling CDFG. The contributions of the presented work are as follows: (a) novel encoding scheme for chromosomes comprising of ‘datapath string’ (b) novel hybrid encoding scheme for ‘auxiliary string’ that acts a priority resolver during scheduling (c) consideration of operation chaining in scheduling during delay evaluation as well as impact of loop unrolling on the configuration of multiplexer size during power evaluation (d) sensitivity analysis on the effect of ‘mutation probability (P M )’ and ‘fraction of population mutated (F M )’ on final cost and exploration time. Results of the proposed approach yielded reduced final cost (real optimal solution) and exploration speed for all tested benchmarks at P M = 0.5 and F M = lower 50 % chromosome, when compared to previous GA approaches.