Research On Hybrid-Grained Reconfigurable Array Architecture And Simulator Based On Data Stream

In recent years,computing demands are growing because of the emergence of data centers,which needs higher calculate ability.The reconfigurable computing platform makes full use of the characteristics of spatial computing,which is suitable for a variety of data-intensive computing scenarios,providing acceleration for most types of applications.However,the existing reconfigurable computing arrays and their research are mostly based on coarse-grained arrays,which leads to insufficient controllability,and it is difficult to support control-type applications.At the same time,the execution mechanism driven by configuration makes array switch configuration frequently,which reduce effectiveness.This thesis studies a hybrid-grained reconfigurable array structure based on data stream,which integrates fine-grained processing elements and coarse-grained processing elements,fine-grained elements implement flexible control while coarse-grained elements implement efficient calculation based on data flow.In order to use this hybrid-grained array structure,the task is divided into data flow graph(DFG,data flow graph)and control flow graph(CFG)before execution.The data flow graph is mapped to the coarse-grained array,and the control flow graph is mapped to the fine-grained array,The combination of the two kinds of node increases autonomy of array.During the execution,the data flows between the array elements according to the configured connection,completing the integration of operations and data,control and computation and support computeintensive and control-intensive algorithms.At the same time,the model of hybrid-grained reconfigurable array is built,we achieve a periodic-level software simulator to evaluate the performance of different algorithms.The thesis also builds a system verification and test platform,which is used for evaluating some different algorithms.The experimental results show that the hybrid-grained reconfigurable array structure in this thesis can support the execution of many complex algorithms and has high execution efficiency,which will provide useful reference for other researchers in the future.