CAPA: Collaborative Research: ARION: Taming Heterogeneity with DSLs, Approximation, and Synthesis

专业化和新技术的出现是推动异构系统的关键力量。异构性已经被广泛使用，公共云提供了计算能力和存储异构的实例。该项目确定了以下力量，这些力量将使系统变得异构，而不仅仅是计算和存储，使编程和编译变得更加复杂，超出了我们今天面临的挑战。该项目开发了 Arion，这是一个基于多种统一思想将程序编译到异构平台上的系统。 Arion 系统将根据实际相关的工作负载进行评估，范围从计算机视觉和虚拟现实到图形计算、机器学习和流处理。研究人员将与业界合作伙伴合作，将研究成果转化为产品，该项目开发的工具和软件将作为开源发布。该项目的研究依赖于四个统一的想法。第一个主旨探讨了调度和类型系统将程序的规范与其实现策略分开，从而实现性能可移植性，因为人们可以在不更改程序的情况下选择其并行性、局部性和硬件映射。第二个推动力使用特定于领域的语言不仅描述程序，还描述编译期间使用的工件，例如调度、资源和内存一致性模型。这允许自动合成这些工件。第三个推动力使用资源模型为大型程序带来调度和综合，因为目标程序不需要一次全部调度或综合。相反，编译器在做出低级决策之前，通过使用模型估计性能来做出高级决策。最后，研究人员将使用正式的方法将程序提升到我们的 DSL 中，并验证和综合我们的 DSL 中的程序，从而提供高度的自动化。验证器和合成器是根据 DSL 的描述自动生成的。

Specialization and the arrival of new technologies are key forces motivating heterogeneous systems. Heterogeneity is already in use widely, with public clouds offering instances that are heterogeneous in both compute capabilities and storage. This project identifies the following forces that will make systems heterogeneous beyond just compute and storage, complicating programming and compilation beyond the challenges that we face today. This project develops Arion, a system for compiling programs onto heterogeneous platforms based on several unifying ideas. The Arion system will be evaluated on practically relevant workloads ranging from computer vision and virtual reality, to graph computations, machine learning and stream processing. The investigators will work with partners in industry to transfer research results to products, and the tools and software developed by this project will be released as open source.The research in this project relies on four unifying ideas. The first thrust explores schedules and type systems separate a program's specification from its implementation strategy, enabling performance portability because one can select, without changing the program, its parallelism, locality, and hardware mapping. The second thrust uses domain-specific languages to describe not only programs but also artifacts used during compilation, such as schedules, resource-, and memory consistency models. This allows automatic synthesis of these artifacts. The third thrust uses resource models to bring scheduling and synthesis to large programs because the target program need not be scheduled or synthesized all at once. Instead, the compiler makes high-level decisions by estimating performance using a model before committing to low-level decisions. Finally, the investigators will use formal methods to lift programs into, and verify and synthesize programs in our DSLs, providing a high degree of automation. The verifiers and synthesizers are automatically generated from descriptions of DSLs.