Datapath Design and Code Generation for VLIW ASIPs

VLIW ASIP 的数据路径设计和代码生成

• 批准号: 9901255
• 负责人: Margarida Jacome
• 金额: $ 30万
• 依托单位: University of Texas at Austin
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 1999
• 资助国家: 美国
• 起止时间: 1999-09-15 至 2003-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9901255&HistoricalAwards=false
• 关键词: Datapath; Design; Code; Generation; VLIW

Very Large Instruction Word (VLIW) Application Specific Instruction-Set Processor (ASIP) cores realize performance/power/energy tradeoffs that can be very attractive in the exceedingly competitive embedded multimedia and signal processing market. In particular, specialized datapaths comprising clusters of functional units with associated (local) register files can enable significant performance/power/energy gains over those based on centralized register files. However, such specialized datapaths pose significant research and practical challenges. The proposed research focuses on jointly addressing (1) the synthesis of VLIW ASIP datapaths, and (2) the development of high-quality retargetable compilers for such specialized processors. The proposed methodology includes a novel problem phasing/decomposition that first considers the joint binding of operations and data objects to the functional units and register files of the specialized datapath. Complementary aggressive optimization heuristics will be developed to perform datapath dependent code transformations, exploit problem symmetries to dramatically reduce the search space, and to effectively bind/assign and schedule operations, data objects and data transfers onto the datapath's resources. If this research is successful, it will contribute to technologies that will increasingly provide the competitive edgefor the embedded systems industry.

超大指令字（VLIW）专用指令集处理器（ASIP）内核实现了性能/功耗/能量的权衡，这在竞争激烈的嵌入式多媒体和信号处理市场中非常有吸引力。特别地，包括具有相关联的（本地）寄存器堆的功能单元的集群的专用数据路径可以实现超过基于集中式寄存器堆的那些的显著的性能/功率/能量增益。然而，这种专门的数据路径带来了重大的研究和实践挑战。拟议的研究重点是共同解决（1）合成的VLIW ASIP数据路径，和（2）高品质的可重定向编译器的发展，这样的专门处理器。所提出的方法包括一个新的问题分阶段/分解，首先考虑联合绑定的操作和数据对象的功能单元和寄存器文件的专用数据路径。将开发互补的积极优化算法，以执行依赖于数据路径的代码转换，利用问题对称性，以大大减少搜索空间，并有效地绑定/分配和调度操作，数据对象和数据传输到数据路径的资源。如果这项研究成功，它将有助于技术，将越来越多地提供嵌入式系统行业的竞争优势。