YHFT-Matrix2 Compiler’s Technologies Related To SIMD Optimization Research And Implementation

YHFT-Matrix2 DSP is a kind of 64-bits high-performance Digital Signal Processor, which is researched and developed by the institution of Computer Science Microelectronics of the National University of Defense Technology. It is mainly used in multimedia computing. To support the multimedia computing characteristics of large amount of data, narrow data type and high degree of parallelism, YHFT-Matrix2 DSP designs a large number of SIMD instructions. Therefore,the implementation of SIMD instructions on YHFT-Matrix2 compiler has an important significance in improving the execution efficiency of apps and the speed of processor. Based on analysis of YHFT-Matrix2 DSP architecture, this paper focuses on the implementation of SIMD vectorization from manual mode, automatic vectorization and human-computer interaction. The followings are the main contents and contributions of this paper :1)The built- in function method of SIMD vectorization is realized. This paper accomplishes SIMD vectorization and achieves the expected SIMD code on YHFT-Matrix2 compiler by three steps, including founding built- in function and parameter type initialization,adding instruction template and processing the compiler option. Then the YHFT-Matrix2 compiler implements the SIMD instructions by calling the built- in function and gets a pretty noticeable effect on performance.2)The auto-vectorization mapping on YHFT-Matrix2 compiler of the SIMD instruction is achieved. This paper analyzes the main factors restricting the development of SIMD vectorization technology: data dependence, control dependence, the continuity requirements for memory access and data memory alignment. Then on YHFT-Matrix2 compiler, data dependencies is solved by using conventional data dependence testing method; control dependence is solved by using if-conversion method; the continuity requirements for memory access is solved by using data reorganization method; data memory alignment is implemented by modifying the back-end macro definition. The implementation of SIMD auto-vectorization largely reduces programmers’ workload and requirements.3)The compiler-directive of SIMD vectorization is accomplished. This paper analyze and research the two compilation guidelines in GCC, including SIMD directive position information and based on the built- in function. Then a whole derective mapping thought combining front-end, middle-end and back-end is presented on YHFT-Matrix2 compiler. The YHFT-Matrix2 compiler can analyze and process the compiler directives, and then call the built- in function to achieve SIMD instrctions. The SIMD instruction comp iler-direct method can use programming skills and experience of programmers, give full play to the advantages of auto-vectorization.Finally, analyze the benefit cost of all SIMD optimization method s, and propose the prospect of the next step of work.