| With the rapid development of digital processing technology such as video structure and computing camera,the huge demand for computing power promotes the rapid development of high performance digital signal processor.Feiteng M7002 emerged as a domestic high performance DSP.In order to give full play to the hardware advantages of the high performance platform,It is a pressing problem for us to develop and optimize the corresponding high performance image processing algorithm for the specific architecture of domestic DSP,broaden the application direction of Chinese DSP chip,and then realize the efficient combination of hardware and software.In this thesis,three widely used image enhancement algorithms are implemented on FT DSP platform,and an optimization method matching the hardware characteristics is designing,which provides support for image enhancement function library on FT DSP platform.The work of this thesis is as follows:(1)Designing the overall implementation process of Laplacian image enhancement algorithm,Boxfilter image enhancement algorithm and Customfilter image enhancement algorithm on FT platform.By studying the principle of the algorithm and the characteristics of the software and hardware platform of FT-M7002,including its architecture,FT-MT2 kernel and memory hierarchy,three image enhancement algorithms for FT-M7002 are designed,and the correctness of the algorithm is verified by testing.(2)Combined with the program characteristics of the three image enhancement algorithms and the characteristics of FT DSP platform,the parallel optimization scheme supporting multiple data types and different convolution kernels is designed and implemented.Based on the program hotspot analysis,the parallel algorithms of Laplacian and Customfilter are designed for different data types,and the parallel implementation of 3*3,5*5 and 7*7 convolution kernels is carried out respectively.The parallel algorithm of column and column separable filter of Boxfilter is designed,and three convolution kernels of mean,sharpening and emrelief are implemented in parallel for different data types.At the same time,the data type conversion interface between character type and integer type,short integer type and integer type is proposed,which solves the compatibility and scalability problems of FT DSP platform for some data types.(3)Make full use of the ultra-long instruction word structure provided by FT DSP platform,optimize compilation options,Cache area,AM vector space and DMA high-speed transmission channel,etc.Make full use of to fetch optimization,loop unrolling ranks of soft water,transposed,DMA double buffer hidden transmission,circulation processing methods such as reduce the time-consuming algorithm in the data access stage,synchronous data transmission,calculation and filling time gap,to solve the problem of parallel discrete unit filling dissatisfaction,and data access,algorithm efficiency is improved effectively.The correctness and performance of the optimized image enhancement algorithm are tested on FT platform.The experimental results show that the performance of the optimized Laplacian,Boxfilter and Customfilter is improved by 1.841-2.674 times,1.806-3.036 times and 2.005-3.708 times,respectively.Compared with TI-TMS320C6678 DSP processor,the optimized algorithm based on FT platform can achieve 1.472-2.502 times,3.479-4.056 times and 3.154-3.863 times of acceleration respectively,which fully takes advantage of the efficient computing advantage of domestic DSP platform. |
PDF Full Text Request