The Research And Design Of Multi-objective Mapping Algorithm For Network On Chip

Under the guidance of Moore's Law,the semiconductor process node are continued to advance and the IC design industry are developed rapidly.The number of IP cores integrated on a single chip is increasing.The traditional bus interconnection methods can no longer meet the interconnection requirements of IP cores in chips in terms of scalability,compatibility,communication speed,and power consumption.By combining the ideas of distributed design and OSI communication model,Network on Chip is applied in the So C interconnection technology,effectively improving the communication efficiency,scalability and compatibility between IP cores.The multi-objective mapping of Noc,many-objective mapping of No C,topology structure,routing algorithm and router micro-architecture of the No C are researched in this thesis.The design and implementation of multi-objective mapping algorithm and many-objective mapping algorithm for No C are respectively carried out.RNSGA2 are proposed(Reprocessed Nondominated Sorting Genetic Algorithm)in this thesis which is based on the characteristics of NSGA2 and No C.RNSGA2 adopts a positionbased coding method.Each solution vector is represented by an array P[i].The value of the array element P[i] represents the position of the No C node.The element number i of the array element represents the IP core number.The global search of RNSGA2 is performed through a single-point crossover and local search is performed through three mutation methods.It retains the fast non-dominated sorting in NSGA2,the crowded distance,the crowd comparison operator,and adds a reprocess operator to the main loop of the algorithm.The redundant solutions in the first non-dominated solution set are found and crossed at multiple points by the reprocess operator.Then they are put into the next generation to participate in inheritance.There are two characteristics of RNSGA2: First,it can solve the mapping problem of two objective functions at the same time.Second,it effectively removes the redundancy solution in the solution set of No C mapping problem and guarantees the invariable population number and the retention of excellent genes.Designs,implements and simulates each module of RNSGA2 are carried out.The RNSGA2 modules were designed,implemented,and simulated using MATLAB.RNSGA3 are proposed in this thesis which is based on the characteristics of NSGA3 and No C many-objective mapping problems.It adds reference points and reference vectors based on RNSGA2,normalizes the fitness values of all individuals and selects the number of individuals associated with the reference vector.The reprocess operator is used to remove redundant solutions in the first non-dominated solution set.There are two main features of RNSGA3: First,the search space can be divided by reference vectors,the search range can be reduced,three or more many-objective mapping problems can be solved,and the No C mapping problem solution can be effectively eliminated.The centralized redundancy solution guarantees the population constant and the retention of excellent genes.The advantages and disadvantages of RNSGA2 and RNSGA3 are analyzed in different application scenarios in this thesis.The RNSGA3 modules were designed,implemented,and simulated using MATLAB.A 2DMesh-based topology and a clock-accurate No C evaluation platform are implemented using synthesizable Verilog HDL in this thesis.2DMesh topology,XY routing algorithm,wormhole switching and ACK-NACK flow control strategy are adopted in this Network on Chip.The evaluation platform is designed with a software interface,the solution to the No C mapping problem is input to the hardware platform for simulation,and a network performance report can be given.The results show that the RNSGA2 solution has a higher saturated throughput rate than the NSGA2 solution of 11.76%,23.26%,50%,48.65%,41.67% and 33.33%,respectively,under the six traffic patterns.And the average network performance increased by 35.1%.And the RNSGA3 solution has a higher saturation throughput than the RNSGA2 solution of 57.7%,46.67%,20.83%,52.78%,59.38% and 39.53% under the six traffic patterns.And the average network performance increased by 43.72%.