CAREER: Design Virtualization for Mainstream Programming of Reconfigurable Computers

尽管与其他计算设备相比，可重构计算(RC)设备已被广泛显示出具有显著的性能和功率优势，但由于非主流高级综合、复杂的低级调试、通常需要数小时或数天的漫长编译时间、缺乏代码可移植性、令人望而却步的设备成本等挑战导致的应用设计复杂性的增加，这种设备的主流使用受到了抵制。为了解决这些问题，本研究引入RC虚拟化来缩小高层应用程序与拥有数十万细粒度资源的物理RC设备之间的语义鸿沟。虽然物理RC设备必须是细粒度的，以实现灵活性和成本效益，但建议的虚拟RC设备(在此类物理设备上实现)可以专门用于不同应用领域的需求，以隐藏物理设备的复杂性。因此，建议的虚拟化支持快速的RC编译，可以在几秒钟内完成，而不是几个小时。快速的RC编译反过来又支持用于调试的主流设计方法。除了使第三方工具能够针对任何RC系统实现工具可移植性之外，虚拟化还支持跨物理RC设备的应用程序可移植性，类似于虚拟机实现可移植软件代码的方式。最后，通过将虚拟化与针对异类加速器(例如OpenCL)的标准化并行语言的高级综合相结合，本研究通过支持其他主流计算设备使用的设计流程来实现主流RC的使用。

Although reconfigurable computing (RC) devices have been widely shown to have significant performance and power advantages compared to other computing devices, mainstream usage of such devices has been resisted due to increased application design complexity caused by non-mainstream high-level synthesis, complex low-level debugging, lengthy compilation times often requiring hours or days, a lack of code portability, prohibitive device costs, among other challenges. To address these problems, this study introduces RC virtualization to reduce the semantic gap between high-level applications and physical RC devices with hundreds of thousands of fine-grained resources. Whereas physical RC devices must be fine grained for flexibility and cost effectiveness, the proposed virtual RC devices, which are implemented atop such physical devices, can be specialized to the needs of different application domains to hide physical device complexity. As a result, the proposed virtualization enables rapid RC compilation that can complete in seconds as opposed to hours. Rapid RC compilation in turn enables mainstream design methodologies for debugging. Virtualization also enables application portability across physical RC devices, similar to how virtual machines achieve portable software code, in addition to tool portability that enables third-party tools to target any RC system. Finally, by combining virtualization with high-level synthesis from standardized parallel languages for heterogeneous accelerators (e.g., OpenCL), this study enables mainstream RC usage by enabling the design flow used by other mainstream computing devices.