摘要
魂芯DSP是一款采用VLIW和SIMD架构的针对高性能计算领域而设计的32bit静态标量数字信号处理器.为了满足数字高性能计算的性能要求,魂芯DSP提供了丰富的复数指令,而编译器不能直接利用这些复数指令来提升编译性能.因此针对魂芯DSP芯片提供了大量的复数类操作指令的特点,在传统开源编译器Open64的编译框架基础上进行研究,实现了复数作为编译器基础类型和复数运算操作的支持.同时,通过识别特定的复数类操作的模式利用魂芯DSP上的复数类指令对程序编译优化.实验结果表明,该实现方案在魂芯DSP编译器上对复数程序优化后能够取得平均5.28的加速比.
BWDSP is a 32 bit static scalar digital signal processor with VLIW and SIMD features, which is designed for high performance computing. In order to meet the performance requirements of digital high-performance computing, the soul core DSP provides a rich set of complex instructions, and the compiler cannot directly use these complex instructions to improve the compilation performance. Since BWDSP has a wealth of complex type of instructions, and it has high performance demands in the radar digital signal field, the implementation is researched according to the characteristics of BWDSP features based on the traditional open-source Open64 compiler framework to achieve the complex data type and complex operations support operations, and further optimization of complex instruction is realized by identifying a specific type of complex operation of a series of patterns. The experimental results show that the implementation on BWDSP compiler can achieve 5.28 time performance improvement on average.
出处
《计算机系统应用》
2017年第9期40-45,共6页
Computer Systems & Applications
基金
"核高基"重大专项(2012ZX01034-001-001)