Hardware/Software Partitioning And Task Scheduling Based On Graph Neural Network

Based on the rapid development of embedded technology,there are a series of embedded electronic systems appeared,such as face recognition,audio decoding,and intelligent diagnosis.The embedded systems generally include software(SW)design and hardware(HW)design.Since the properties of software and hardware should be taken into account during the design process,traditional system design cannot meet this requirement,and HW/SW co-design emerges.Hardware/software partitioning(HSP)is a critical part in HW/SW co-design,and the result of the HSP affects the quality and efficiency of system design.To this end,we focus on the HSP problem and propose a new HSP algorithm to improve the performance and the speed of HSP.First,to address the defeats of previous partitioning methods,we propose a partitioning model that combines HSP and task scheduling(Graph Attention Network-based Partitioning and Task Priority-based Static Scheduling,GATP-TPSS).GATP is the partitioning part that uses the graph attention network as the main framework,the core operation is matrix operation,and the optimization method for GATP is the classic gradient descent method.The input of GATP is a directed acyclic graph(DAG),and graph convolution operations are used to extract the features of nodes in the graph.Then,a linear classifier is utilized to achieve the HW/SW probabilities of the nodes.The output of GATP is the labels of the nodes,which is generated by the greedy selection approach.TPSS is a static scheduling algorithm based on task priorities.Scheduling is carried out after partitioning,and the optimal scheduling result of the task graph is achieved by minimizing task execution time under the hardware area constraints.Besides,the result will be fed back to the GATP model to further improve the performance of GATP.Second,in order to evaluate the effectiveness of the GATP-TPSS algorithm,we set up four comparative experiments,and we compare the proposed GATP-TPSS algorithm with three heuristic methods.The experimental results show that:(1)GATP-TPSS outperforms the state-of-the-art methods for large-scale graphs(the number of nodes is more than 200);(2)The time complexity of GATP-TPSS is lower than other algorithms,which can make our algorithm converge faster.Finally,based on the proposed GATP-TPSS algorithm,we use the Py Qt5 framework to develop a simple HSP software tool.This tool can call the GATP-TPSS model for processing the input task graph,and the partitioning result will be displayed,which can be referenced for the next procedures of system design.