论文部分内容阅读
在信号处理领域的优异表现使得Vector-SIMD结构在近年来获得了广泛的关注.Vector-SIMD结构和多核技术相结合是目前高性能DSP体系结构发展的重要方向.然而,在目前的多核VectorSIMD处理器中某些部件间的协同工作能力还比较弱,导致了系统的整体性能得不到有效发挥.本文设计实现了一款协同多核DSP YHFT-QMBase,从4个方面增强了多核Vector-SIMD体系结构的协同性:(1)采用动态耦合机制重定义了标量单元和向量单元的工作方式;(2)采用矩阵方式的通信机制增强了向量Lane间的交互能力;(3)采用非对齐向量存储访问机制解决了向量存储器的数据共享问题;(4)采用Qlink-Crossbar机制满足了多核间后台高效粗粒度数据搬移的需求.评估结果显示,本文提出的协同增强机制能够使传统的Vector-SIMD结构获得58.5%的性能提升.目前YHFT-QMBase已经成功流片,评测结果显示其峰值浮点乘加能力(单精度)达到32 GFMACS,定点运算能力(16位)为128 GMACS,典型功耗为8.65 W.
The outstanding performance in the field of signal processing has made Vector-SIMD structure gain widespread attention in recent years. The combination of Vector-SIMD structure and multi-core technology is an important direction for the development of high-performance DSP architecture.However, in the current multi-core VectorSIMD processing The cooperation between some parts of the device is still weak, resulting in the system’s overall performance can not be effectively played.This paper designed and implemented a collaborative multi-core DSP YHFT-QMBase, enhanced from 4 aspects of the multi-core Vector-SIMD system Structural synergy: (1) The dynamic coupling mechanism redefines the work of scalar elements and vector elements; (2) the matrix-based communication mechanism enhances the interaction between vectors Lane; (3) the use of non-aligned vector storage (4) The Qlink-Crossbar mechanism is adopted to meet the requirement of high-efficiency coarse-grained data migration between multi-cores in back-end.The evaluation results show that the proposed collaborative enhancement mechanism can make the traditional Vector-SIMD structure Get 58.5% performance improvement.At present YHFT-QMBase has been successful film, the results show that the peak floating-point multiply and add capacity (single-precision) To 32 GFMACS, fixed-point arithmetic capability (16) is 128 GMACS, typical power consumption of 8.65 W.