Codesign Methodology of Application Specific DSPs Based on a Retargetable Compiler

在这个项目中，我们对特定应用dsp的软件和硬件的协同设计方法进行了研究，重点是弥合“数据路径配置和指令集之间的差距”的问题。具体来说，我们已经开发了一个可重目标编译器、一个代码压缩器和一个控制部分生成器，它们针对具有不规则配置的用户定义数据路径。可重定向编译器接受C程序并为数据路径生成水平代码。代码压缩器采用基于字典的方法减少代码长度，在某种意义上取代了指令集设计。控制部分发生器为数据通路合成了一个控制器，使其成为一个完整的DSP。我们还在G.723语音压缩机上对我们的“面向数据路径”协同设计方法进行了案例研究。至于可重目标编译器，我们已经开发了指令选择、绑定和调度的方法，以便即使对于包含100多个操作的大型基本块，也能在可行的时间内生成代码。特别是，我们开发了即使对于不规则数据路径也不会失败的溢出代码插入机制和寄存器生命周期序列化算法，以最大限度地减少溢出代码插入计数。在为G.723及其两个变体的专用数据路径生成代码的实验中，我们发现我们可以检查硬件成本和执行周期之间的权衡。

In this project, we conducted a research on codesign methodology for software and hardware of application specific DSPs focusing on the problem of birdging "a gap between datapath configurations and instruction sets". Specifically, we have developed a retargetable compiler, a code compressor, and a control part generator, which target user-defined datapaths with irregular configurations. The retargetable compiler takes C programs and generates horizontal codes for the datapath. The code compressor reduces the code length by a dictionary based method, which is in a sense a replacement of instruction set design. The control part generator synthesizes a controller for the datapath to make it a complete DSP. We also conducted a case study of our "datapath oriented" codesign methodology on a G.723 voice compressor.As for the retargetable compiler, we have developed methods for instruction selection, binding, and scheduling so that code would be generated within a feasible amount of time even for large basic blocks consisting of more than 100 operations. Especially, we have developpeda spill code insertion mechanism that never fails even for irregular datapaths and a register lifetime sequentialization algorithm to minimize the spill code insertion count. In a experiment where codes were generated for a dedicated datapath for G.723 and their two variations, we found that we can examine trade-offs between hardware cost and execution cycles.