YHFT-Matrix编译器向量化技术的研究与实现

发布时间：2018-05-28 06:12

本文选题：YHFT-Matrix + 自动向量化　；参考：《国防科学技术大学》2013年硕士论文

【摘要】：随着多媒体和数字信号处理领域的飞速发展，需要更高性能的DSP来满足计算需求。YHFT-Matrix是一款具有自主知识产权的高性能32位DSP处理器，具有特有的体系结构和创新的指令集，属于国家核高基重大专项重点工程项目。YHFT-Matrix采用VLIW体系架构，支持SIMD（single instruction multipledata）指令，具有规模大、功能全、综合运算性能高等特点。Matrix编译器的向量部件能够有效提升YHFT-Matrix芯片的处理性能。论文实现了YHFT-Matrix体系结构的SIMD指令自动向量化技术，但由于自动向量化方法并不能完全满足YHFT-Matrix的需求，论文进一步研究并实现了基于编译指导的向量化技术和基于内联函数的向量化技术，这两项向量化技术与SIMD指令自动向量化技术一起构建了Matrix编译器的向量化技术体系。论文首先对现今编译器向量化技术的研究现状进行了分析，并分析了YHFT-Matrix体系结构和指令集特点，论文的主要工作如下： 1.实现了SIMD指令的自动向量化技术。建立了Matrix编译器的向量后端，按照其指令集的特点，描述了向量指令，对向量寄存器、机器模式和向量长度进行了扩展。本文实现的SIMD指令自动向量化有效提升了Matrix编译器的向量化处理能力，并有效降低了编译器的程序报错率。 2.实现了基于编译指导的向量化技术。针对SIMD指令自动向量化技术无法满足Matrix编译器向量化要求的情况，本文在YHFT-Matrix编译器中设计并实现了两类编译指导控制向量化函数的方法，帮助向量化相关函数识别可以进行向量化的程序信息，并引导向量化相关函数对这些程序信息进行向量化处理操作。 3.实现了基于内联函数的向量化技术，并用此技术在Matrix编译器上实现了IR算法的向量化运算。根据内联函数的向量化技术修改IIR算法的向量化程序，通过Matrix编译器编译产生其目标代码，，并将目标代码导入到YHFT-Matrix DSP芯片中执行，并将执行结果与在CCS中TMS320C6713上执行的模拟结果进行了对比，实验结果表明了IIR算法的向量化运算无误高效，同时也证明了基于内联函数的向量化技术是正确有效的。 Matrix编译器是以GCC-4.7.0为基础移植开发的，本文在Matrix编译器上实现了上述工作并进行了验证，结果表明Matrix编译器的向量化技术方法正确有效，所做工作有效提高了YHFT-Matrix编译器的向量化性能。
[Abstract]:With the rapid development of multimedia and digital signal processing, higher performance DSP is needed to meet the computing requirements. YHFT-Matrix is a high performance 32-bit DSP processor with its own intellectual property rights, with unique architecture and innovative instruction set. YHFT-Matrix, a key project of the National Nuclear High Base Project, uses VLIW architecture to support the SIMD(single instruction multiple data (SIMD(single instruction) instruction, with large scale and full function. The vector component of the matrix compiler can effectively improve the processing performance of YHFT-Matrix chip. In this paper, the SIMD instruction automatic vectorization technology of YHFT-Matrix architecture is implemented, but the automatic vectorization method can not fully meet the requirements of YHFT-Matrix. This paper further studies and implements the vectorization technology based on compiler guidance and the vectorization technology based on inline function. These two vectorization techniques are combined with SIMD instruction automatic vectorization technology to construct the vectorization technology system of Matrix compiler. Firstly, this paper analyzes the current research status of compiler vectorization technology, and analyzes the characteristics of YHFT-Matrix architecture and instruction set. The main work of this paper is as follows: 1. The automatic vectorization technology of SIMD instruction is realized. The vector back end of the Matrix compiler is established. According to the characteristics of its instruction set, the vector instruction is described, and the vector register, machine mode and vector length are extended. The SIMD instruction automatic vectorization implemented in this paper effectively improves the vectorization processing ability of the Matrix compiler and effectively reduces the program error rate of the compiler. 2. The vectorization technology based on compiler guidance is realized. In view of the fact that automatic vectorization of SIMD instruction can not meet the requirements of Matrix compiler, this paper designs and implements two kinds of compile-guided control vectoring functions in YHFT-Matrix compiler. It helps to identify the program information that can be vectorized by the vectoring correlation function, and guides the vectoring correlation function to vectorize the program information. 3. The vectorization technique based on inline function is implemented, and the vectorization of IR algorithm is implemented on Matrix compiler. According to the vectorization technique of inline function, the vectorization program of IIR algorithm is modified, and the target code is compiled by Matrix compiler to produce its target code, and the target code is imported into YHFT-Matrix DSP chip to execute. The experimental results show that the vectorization operation of the IIR algorithm is correct and efficient, and that the vectorization technique based on inline function is correct and effective. The Matrix compiler is developed on the basis of GCC-4.7.0. This paper implements and verifies the above work on the Matrix compiler. The results show that the vectorization technique of the Matrix compiler is correct and effective. The work has effectively improved the vectorization performance of YHFT-Matrix compiler.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP314

【参考文献】