Research On Multi-issue Technology Of A NoC Platform Faced To High-density Computing

With the development of science and technology, and informatization of production and living, there’s many areas put forward higher demands to the computing power of processors. For the algorithm, people put forward a variety of efficient algorithms to solve the actual problems, the development trend of part of them is that, the computing density increase sharply and support parallel computing. And in terms of computing platform, it also developed a variety of specially optimized dedicated platform to meet the stringent requirements. Among them, heterogeneous multi-core processor based on Network on Chip (NoC) architecture is one of the important branch. Researching improve the performance of the heterogeneous multi-core NoC processor is an important task to the semiconductor industry.According to the above problems, this paper researched and experimented the multi-issue technology on the NoC platform for high density calculation.In this paper, the main work done is as follows:Firstly, it analysis the improvement based on single issue NoC platform prototype system, and put forward the main controller designing based on scalable transmission channel, in order to manage the dynamic scheduling for multi-issue.Secondly, it considered the features of using computing cluster resources in multi-issue system, designed the free cluster queue and the corresponding sorting algorithm for resources dynamic allocation. It considered the characteristics of dynamical access to the DDR when calculating the sub tasks, designed address mapping look-up table, in order to dynamically mapping virtual address/real address. It considered the characteristics of dynamical distributing computing clusters and real address, designed configuration information filling module. It considered the controlling characteristics of multi-issue improvement, finished the design for DDR controller with no channel structure and the design for status network function enhance, in order to fit the multi-issue improvement. It considered the characteristics of bottleneck sub tasks, finished the improving design support local algorithm level out-of-order multi-issue.Thirdly, it build a software model for the multi-issue main controller, and added on time variables in order to gain the performance information during system running. And it finished analysis and mapping for a kind of high density algorithm during the preparation stage for experiment. Then, it load this high density algorithm to software model system to check the function of multi-issue and the performance improvement, and compared the impact to the performance of the system whether open or not open local algorithm level out-of-order multi-issue function. It put forward to a method which optimizing the system computing clusters structure and algorithm mapping, according to the result of performance analysis, and it has been proved effective by the software model. At last, it found a way to resolve the performance gap between odd number of times and even number of times calculating, and it also has been proved effective by the software model.In the end, it wrote the RTL code for the multi-issue main controller in this paper, and finished system integration. Then firstly verified the functional correctness using ModelSim simulator, under the shrinkage ratio of 4096:1 for the sub tasks scale. Soon the full scale experiment was done on the FPGA development board, it verified the feasibility of the design and the accuracy of the software model for performance information.In this paper, design improved multi-issue system has the following characteristics:1) The task scheduling method is suitable for high density type of calculation, and the multi-issue method is suitable for multi-channel parallel computing with high density algorithms.2) It supported different kind of computing in each issue channel at the same time.3) The object that multi-issue system faced to is computing sub tasks, using dynamic scheduling.4) Multi channels sharing all system cluster resources, and the number of channels can be extended according to need.5) It could dynamically distribute the most optimal computing cluster resources according to different issue channel.6) It has compacted configuration information, and could dynamically mapping real address/virtual address space.7) It has the function of local algorithm level out-of-order multi-issue which could be use or not.