Research And Implementation Of The Vectorization For YHFT-Matrix Compiler

With the rapid development in the field of multimedia and digital signal,high-performance DSP is required to meet the computing needs. YHFT-Matrix is ahigh-performance32-bit DSP processor with independent intellectual property rights,which belongs to the national key projects. YHFT-Matrix uses VLIW architecture andsupports SIMD instructions, it applies unique architecture and Innovative instruction set,Large-scale, full-featured and high computing performance. The vector instructions ofMatrix compiler can enhance the performance of DSP chip effectively.Because of the auto-vectorization method of SIMD instructions in theYHFT-Matrix compiler can’t meet the unique architecture and instruction setcharacteristics of YHFT-Matrix, other compiler technologies need to be studiedextensively. According to the compiler vectorization characteristics of YHFT-Matrix,vectorization technologies based on compiler directive and intrinsics were studied andachieved in this paper. These two vectorization methods and the auto-vectorizationmethod of SIMD instructions formed the vectorization system of Matrix compiler.The modern vectorization technologies, the internal structure of YHFT-Matrix andthe vectorization process of YHFT-Matrix compiler were analysed. Taking modernvectorization technologies and the features of YHFT-Matrix, the major works of thispaper are as follows.1. The auto-vectorization method of SIMD instruction was accomplished. Thevector backend of YHFT-Matrix compiler was built, according to thecharacteristics of instruction set, vector instructions were described; vectorregisters, machine modes and vector length were extended too. This programeffectively reducing the compiler error rate.2. The vectorization method of compiler directive was achieved. Because of theinsufficient of the existing compiler, in this paper, the Matrix compiler studiedand implemented two types of compiler directive. This method help torecognize vectorization information of the program and guide these proceduresinformation to vectorization processing operation.3. The vectorization method of intrinsics was achieved, with the methodvectorization operation of IIR algorithm were achieved in Matrix compiler.Matrix compiler generates object code and object code was implementationin YHFT-Matrix DSP chip, the Results were compared with theimplementation of CCS TMS320C6713simulation results. The comparedresults show that vectorization of the IIR algorithm is correct and efficient.With archiving and verifying the above works based on GCC-4.7.0, the resultsshow that the vectorization method of the YHFT-Matrix compiler is correct and effective, and the works can improve the vectorization level of the YHFT-Matrixcompiler effectively.