The Research On NoC Architecture And Mapping Technology For Wireless Communication

With the development of semiconductor integrated circuits,the number of processing cores that can be integrated on a single chip has gradually increased,and the traditional bus structure has become a bottleneck restricting the development of chip size,speed,power consumption,and communication requirements.As a brand new solution,Network on Chip(NoC),transplants Internet thoughts into System on Chip(So C).By separating the communication and calculation parts,it solves the problems of the bus architecture very well.At the same time,with the rapid development of wireless communication technology,the amount and complexity of the internal calculation of the system are getting higher and higher,and large-scale parallel computing has become a development trend.Because of some advantages of low power consumption,good scalability,and reliable transmission of the NoC,the multi-core parallel processing platform based on the NoC can provide strong support for wireless communication computing while minimizing power consumption and area.Starting from the basic theory of NoC,this thesis focuses on the mapping problem,which is one of the key technologies in NoC design.The mapping problem of NoC is a non-deterministic polynomial(NP)difficult problem.The search space of solution of NP problem is too large,and an approximate solution can only be obtained through some heuristic search algorithms.At the same time,this thesis focuses on the two goals of power consumption and delay.The solution of the goals is a multi-objective problem.Usually,the goals of multi-objective problem restrict each other,so it is difficult to obtain the optimal value on each design goal.Therefore,this thesis focuses on the NoC mapping technique for wireless communication,and carries out the following research work for the task division and mapping solution optimization:(1)Analyzed the typical features and related calculation examples in the wireless communication algorithm,namely the Singular Value Decomposition(SVD)algorithm and the Minimum Mean Squared Error-Interference Rejection Combining(MMSE-IRC)algorithm,designed NoC based on the 2D-Mesh topology,and visualized the calculation process of the wireless communication example into a directed acyclic task flow graph.Among them,the nodes and edges of the task flow graph respectively represent the calculation instructions and data flow direction of the calculation example.(2)According to the task flow diagram of the wireless communication algorithm,the NoC mapping energy consumption and delay calculation model is established,and the corresponding evaluation function and constraint conditions are proposed.Because the number of task flow graph nodes is too large,the search space of solution of mapping is too large,the calculation complexity is high.Therefore,this thesis proposes a task flow graph partition technique based on minimizing the sum of weight of all global edges,which merges some nodes in the task flow graph,reduces the scale of the task flow graph.The simulation results prove that the strategy can reduce the complexity of subsequent mapping obviously.In the selected SVD calculation example,the power consumption of the divided graph mapping result is reduced by 33.23%,and the delay is reduced by20.18%;in the MMSE-IRC calculation example,the divided graph mapping result is reduced in power consumption Increased by 9.33%,and reduced the delay by 18.76%.(3)In order to enhance the search ability of the solution space based on the traditional heuristic algorithm,an improved algorithm based on Generative Adversarial Networks(GAN)is proposed.As an unsupervised learning technology that has been widely used in the field of data enhancement and generation in recent years,GAN is composed of a generator network and a discriminator network.It borrows the ideas of game theory,using real samples and random samples to train both networks,and finally achieved a balance between generator network and discriminator network.In this thesis,GAN is applied to the multi-objective mapping optimization problem.The simulation results prove that the solution generated after GAN training performs better in terms of distribution and convergence,and the optimal solution generated after training is better than the sample solution,which is closer to the real optimal solution.The data shows that in the MMSEIRC example,the optimal solution obtained by GAN reduces power consumption by24.49%,and the delay is reduced by 10.89%;in the SVD example,the optimal solution obtained by GAN is in power consumption.It is reduced by 17.54%,and the delay is reduced by 8.77%.The algorithm proposed in this thesis has achieved a breakthrough in its performance.Besides,the complete algorithm flow and general mapping optimization model from certain valuable reference for the specific implementation and optimization of related technologies in the field of wireless communication and NoC mapping in the future.