Time domain finite difference method hardware accelerator architecture and implementation method thereof

Abstract

The invention discloses a time domain finite difference method hardware accelerator architecture. The architecture comprises a chain type calculation processing unit arrangement structure, multiple groups of storage blocks and an excitation source assignment device, the arrangement structure of the chained calculation processing unit is a combination of a plurality of PE arranged in the one-dimensional direction and is used for calculating the electric field value and/or the magnetic field value of the three-dimensional grid points simulated by the electromagnetic field; the storage block is composed of a plurality of block random access memories and is used for storing the electric field values and/or the magnetic field values of the grid points calculated by the PE; and the excitation source valuator is used for valuating the electric field value and/or the magnetic field value stored in the storage block. According to the invention, the utilization rate of the PE can be improved, the wiring difficulty and time delay comprehensively realized on the FPGA are reduced, the off-chip storage bandwidth capability is increased, and the expansibility and grid adaptability of the PE arrayare improved. The method can be applied to the field of hardware accelerators based on FPGA platforms.

Description

technical field [0001] The invention relates to the field of a hardware accelerator based on an FPGA platform, in particular to a time domain finite difference method hardware accelerator architecture and an implementation method thereof. Background technique [0002] When the existing FDT hardware accelerator based on the FPGA (Field Programmable Gate Array) platform calculates the electromagnetic field in three-dimensional space, the large grid that needs to be simulated will be calculated according to the logic resources and storage resources on the FPGA. Evenly divided into several small grids, and each small grid is allocated a computing processing unit PE (Processing Element) and a set of block random access memory BRAM (Block RAM) for storing data, forming a three-dimensional array of small grids . Each calculation processing unit PE is responsible for calculating the electric field value and magnetic field value of all grid points in the small grid, and updating the...

AI Technical Summary

Problems solved by technology

This defect will cause different projects to be simulated each time to re-program the FPGA bit stream to configure hardware accelerators with different architectures. The amount of engineering is very large, and it is difficult to be practical in actual projects.

That is, the scalability and grid adaptability of the PE array are very weak

Images