Research Of Vision Processing API Library Based On General Purpose Reconfigurable CPU

With the development of video processing technology,there are higher requirements for the data volume,circuit performance,power and flexibility.The parallel processing technology of reconfigurable processor becomes a good solution to improve the computational efficiency and reduce the power consumption.At present most of visual information processing algorithms have the features of regular data dependencies,centralized computing complexity and concurrent execution based on block.Reconfigurable processors have obvious advantages on the visual information processing algorithm.Based on the architecture of GReP and its programming mode,this paper designs 4 kinds of fixed point division,analyzes their performance and resource cost,and gives the scope of them.Then this paper presents a method of floating-point multiplication and addition implemented on GReP.The performance improvement of this configuration is analyzed,and the specific usage examples are given under different algorithms.Finally,this paper presents a visual information processing API function library interface specification,and carries out the general mapping flow of visual information processing API functions.According to the ALU function and designed division and floating point arithmetic kernel,this paper maps three kinds of visual information processing algorithm and establishes the API function libraries.The serial computing performance of algorithm has been tested on Atom 230 test platform,and the parallel computing performance has been tested on reconfigurable processor simulation test platform based on SoC Designer.The experimental results show that the Harris corner detection,the normalized cross-correlation,random sampling consensus,two-dimensional discrete cosine transform,bicubic interpolation and median filtering algorithm of parallel efficiency are the traditional serial implementation of 3.34 times,19.24 times,7.29 times,4.73 times,7.23 times,and 8.98 times.The performance of floating-point multiplication and addition improve in 2.09 and 1.67 times.The performance of 4*4 matrix multiplication increases 1.88 times,and 7*7 gauss filtering increase 1.87 times.Therefore,the designed visual information processing API function library not only reduces the difficulty of application development,but also greatly improve the performance of the visual information processing application.