Swarm‐inspired exploration of architecture and unrolling factors for nested‐loop‐based application in architectural synthesis

Abstract

A novel framework for automated design space exploration (DSE) of architecture and unrolling factor (UF) for perfectly nested loops in architectural synthesis (AS) is presented. The sub-contributions are that it: (a) proposes a novel model for determining execution time based on architecture and UFs for nested loop without tiresomely unrolling completely; (b) proposes a novel methodology for exploration of architecture and loop UFs for a perfectly nested loop, based on swarm intelligence (SI); and (c) maintains a trade-off between power and performance during exploration and also resolves orthogonal issues such as exploration speed and quality of result (QoR). Results of the proposed approach indicate an average improvement in QoR of >33% and reduction in runtime of >34% compared with recent approaches.Introduction: The design space exploration (DSE) for applicationspecific processors is intricate involving identification of best solution from a set of massive design alternatives. In [1], the DSE problem of only single-loop-based applications with no focus on nested-loop-based control data flow graph (CDFG) is solved. Moreover, the delay model proposed in [1] is not useful when the number of independent operations required to be performed in parallel due to unrolling exceeds the available resource units specified in a solution.Many researchers have addressed DSE in architectural synthesis. For example, in [2], the SPARK tool was introduced where the loop unrolling factor (UF) being user-directed is not able to explore the combination of UF and architecture together automatically. Moreover, in [3] Mishra and Sengupta proposed a particle swarm optimisation (PSO) DSE where the loop unrolling was not considered during exploration. In [4] evolutionary algorithm-based DSE required manual intervention to decide UF besides considering only UFs which are multiples of iteration count. Harish et al.[5] used a genetic algorithm and weighted sum PSO-based DSE which did not consider constraints in cost function and UF during DSE.

Development of Multiobjective High-Level Synthesis for FPGAs

“…An adaptive DSE framework called integrated Particle Swarm Optimization (i-PSO) for delay and power as objective functions in HLS is presented in [130], including a sensitivity analysis of the algorithm. e use of PSO for the DSE of data paths in HLS is also proposed in [131][132][133][134], and the MediaBench benchmark and another DSP benchmark (the paper does not provide details of the benchmark name) to measure the optimization quality of the simultaneous exploration of data path and loop unrolling factor are used. Other authors published a similar strategy, but delay and area were optimized in [135,136].…”

Intrusion Detection System Employing Multi-level Feed Forward Neural Network along with Firefly Optimization (FMLF2N2)

“…Multi-layer neural systems can be set up from multiple points of view. Normally, they have something like one info layer, which sends weighted contributions to a progression of concealed layers, and a yield layer toward the end [21]. These more advanced setups are additionally connected with nonlinear forms utilizing sigmoids and different capacities to coordinate the terminating or initiation of fake neurons.…”