Low-Energy Datapath Design for Programmable Digital Signal Processors

本研究集中在低能耗应用程序特定的数据路径，低能耗的时间表，设计权衡数据路径的数量，流水线和调度方法的水平。应用程序是低能量的数字信号处理和密码系统的实现，使用可编程数字信号处理器的数据路径设计。假设VLIW（超长指令字）类型的可编程DSP（PDSP）。在这样的PDSP中，不同的数据路径可以通过VLIW指令字被独立地激活。本研究旨在修改现有的PDSP中的数据路径，并设计新的数据路径，不存在于当前的低功耗实现的PDSP。这里的重点是同步数据路径和调度或软件共同选择。重要的是，不仅要设计最佳的数据路径，而且要设计最佳的调度方法，以减少数据路径中的切换活动，以减少能量消耗。在固定和自适应FIR数字滤波器和均衡器的背景下，目标是找到最适合于转置形式FIR滤波器的数据路径，使用强度降低的并行结构来减少时钟周期的数量，并通过切换活动来执行调度。在无线通信中用于空时自适应处理和多用户检测问题的RLS自适应滤波器中，正在探索一种分层调度方法。这里采用了全局并行的局部顺序分区方法，其中每个分区被映射到不同的CORDIC数据路径。在每个分区内，Givens的旋转操作被调度以减少切换活动。在椭圆密码系统设计中，目标是通过适当的软硬件协同设计和参数折衷分析来降低能量消耗。

This research concentrates on design of low-energy application-specific datapaths, low-energy schedules, and design tradeoffs with respect to number of datapaths, level of pipelining and scheduling approaches. Applications are to datapath design for low-energy implementation of digital signal processing and cryptography systems using programmable digital signal processors. A VLIW (very long instruction word) type programmable DSP (PDSP) is assumed. In such a PDSP, different datapaths can be independently activated through the VLIW instruction word. This research seeks modifications of datapaths in existing PDSP and design of new datapaths that do not exist in current PDSPs for low power implementations. The emphasis here is on simultaneous datapath and scheduling or software co-selection. It is important not only to design the best datapaths but also to design best scheduling approaches to reduce switching activity in the datapaths to reduce energy consumption. In the context of fixed and adaptive FIR digital filters and equalizers, the goal is to find datapaths which are best suited to transpose-form FIR filters, to use strength-reduced parallel structures to reduce the number of clock cycles and to perform scheduling by switching activity. In RLS adaptive filters for space-time adaptive processing and multi-user detection problems in wireless communications, a hierarchical scheduling approach is being explored. Here a globally parallel locally sequential partitioning approach is adopted where each partition is mapped to a distinct CORDIC datapath. Within each partition Givens' rotation operations are scheduled to reduce switching activity. In elliptic cryptosystem design, the goal is to reduce energy consumption by appropriate hardware-software codesign and parametric tradeoff analysis.