The Research On Program Optimization Techniques Of Embedded Image Compression System

The time performance of embedded platform is unsatisfactory since multimedia applications such as image compression procedure need large amount of computation and complicated procedures. The optimization technology used in image compression for embedded systems could reduce the run-time of procedure, which lowers the power consumption and enhances performance and reliability of embedded systems. But common optimization methods make the compiler optimize the programs automatically by optimizing the compiler. The disadvantage of this kind of methods is that the workload of optimizing the compiler is very huge, the work is very difficult, and when the compiler optimizes a program automatically, it adopts conservative optimization methods to guarantee the correctness of the program. Therefore, for a given multimedia application, it can not promote the performance to the greatest degree In order to reduce the runtime of an embedded image compression program and enhance its performance, this dissertation focuses on the most time-consuming loop structure and studies the method of how to improve the parallelism and data locality by optimizing the C language source code manually with the loop transformation techniques.This dissertation is carried out in the following aspects:(1) With regard to the loop structure in the embedded image compression programs, it imports loop transformation techniques of parallel compiling theory, and studies some loop modeling problems, such as how to describe a single nested loop with MDFG and how to describe multiple nested loops with MLDG.(2) It proposes a program optimization method of embedded image compression system which uses loop distribution, loop fusion and loop tiling synthetically. This method first uses Maximum Loop Distribution Algorithm to separate the loop bodies which are not suitable to loop fusion. Then it uses the loop fusion algorithm of ULF_IP to enhance the instruction-level parallelism and reduce the loop bound testing. Finally, for the large loop bodies generated by the loop fusion, stream-conscious loop tiling algorithm is used to enhance the data reuse.(3) It improves a tiling method which considers input and output data stream and makes orthogonal tiling on iteration space and data space synchronously. This method could avoid the time consumption generated by reordering the input and output arrays following their optimum access orders, thereby improves the time performance.At last, taking the JPEG2000 image compression program as an example, the program optimization method of embedded multimedia system proposed in this dissertation is validated on the DM6446 platform. The results of the experiments testify the validity.