Development And Study Of Compiler And Parallel Programming Model For Multi-core DSP

With the development of speech communication, speech signal processing is becoming more and more complicated, general DSPs have been unable to meet the application requirements. Therefore, we need to design specific mutil-core DSP to realize it. In the process of DSP development. A variety of engineering difficulties exsit in the compiler development by using conventional framework, and the widely used parallel programming model can not make full use of the multi-core computing resources. As a result, the compiler development and programming model design have become key issues in the study of the current multi-core processors.Focusing on the above problems. The study of compiler and parallel programming model launches based on MAD which is developed by ourself. The building of a compiler is achieved by transplanting a MAD backend based on LLVM compiler framework. In this paper, the architecture of MAD, the structure and composition of LLVM framework, the retargetable character of LLVM, and the theory and steps of the transplanting of MAD backend are introduced in details. A hybrid programming model which takes the advantage of both the message passing and the shared memory communication programming model is proposed. The specific implementation details of task scheduling, communication methods between cores, synchronous operations, and lock operations are also described.Finally, the fuction of the compiler and the performance of the parallel programming model are tested. The compiler function test is achieved mainly by running the automatic generated test files, and the result shows that most of the instructions are well supported. The parallel proogramming model performance test is achieved by Livermore loops program test set, and the result shows that by using the model the test program can obtain a high speed up ratio. At last, the 300 bps low bit rate speech codec algrithm is optimized in parallel and the real-time communication is realized on the MAD platform by using the compiler and the model. In addition, the performances of MAD and TI C5535 DSP are compared. It indicates that the compiler and the model can accelerate the program in parallel efficiently to meet the performance requirements.