CAREER: Automatic Generation of High Efficiency Graphics Systems

下一代视觉计算应用程序将需要图形系统在广泛的计算环境中执行（从嵌入式设备到超级计算机），并且比现有系统更强大，效率更高。不幸的是，为任何一个平台或任何一类工作负载创建高效的渲染解决方案目前都是一项费力的、专家级别的任务，其难度阻碍了交互图形技术和图形硬件架构的创新。作为回应，PI正在开发技术，给定虚拟场景的知识，特定的渲染任务和并行硬件平台，可以自动生成专门针对此上下文的优化图形系统实现。核心思想是为关键渲染问题（如表面可见性和阴影）的大量算法方法开发统一的表示，然后在新的智能渲染器编译系统中利用这些表示，为给定的工作负载和机器上下文识别有效的算法。在这个系统中，选择合适的算法和合成高性能代码一样，都是渲染器编译过程的一部分。定义包含这些参数化算法空间的新的核心图形编程抽象，探索它们的应用范围，并确定哪些图形领域知识是自动化图形系统优化的基础，这是该项目的重要研究组成部分。项目的第二个组成部分包括建模和描述图形系统行为，以便指导自动发现有效的解决方案。图形系统性能是新的视觉计算应用程序的基本实现因素。特别是，项目的成功将促进视觉计算在移动和嵌入式环境中的影响，在这些环境中，逼真的增强现实和高分辨率信息显示将对导航、工人培训以及新形式的数字娱乐和艺术等任务产生革命性的影响。由于为图形加速开发的硬件、软件和编译技术继续对通用并行计算的发展产生广泛的影响，因此，成功也将对科学计算领域产生跨学科的好处。高性能视觉计算主题将作为并行计算教育现代化的核心组成部分。这项工作的成果将作为现代并行计算的在线教科书和网络参考资料提供给公众。该项目的网站（http://graphics.cs.cmu.edu/projects/renderergenerator/）提供了有关该项目的更多信息。

The next generation of visual computing applications will require graphics systems to execute within a wide range of computing environments (from embedded devices to supercomputers) and be both more powerful and orders of magnitude more efficient than existing systems today. Unfortunately, creating efficient rendering solutions for any one platform or any one class of workloads is currently a laborious, expert-level task whose difficulty discourages innovation in interactive graphics techniques and graphics hardware architecture. In response, the PI is developing technologies that, given knowledge of a virtual scene, a specific rendering task, and parallel hardware platform, can automatically generate an optimized graphics system implementation specialized for this context. The core idea is to develop unifying representations for the large space of algorithmic approaches to key rendering problems such as surface visibility and shading and then to leverage these representations in new intelligent renderer compilation systems that identity efficient algorithms for a given workload and machine context. In this system, selecting appropriate algorithms is as much a part of the renderer compilation process as synthesis of high performance code. Defining new core graphics programming abstractions that embody these parameterized algorithmic spaces, exploring their application scope, and determining what graphics domain knowledge is fundamental to automating graphics system optimization are significant research components of this project. A second component of the project involves modeling and characterizing graphics system behavior in order to guide automated discovery of efficient solutions.Graphics system performance is a fundamental enabler of new visual computing applications. In particular, project success stands to catalyze the impact of visual computing in mobile and embedded environments, where photorealistic augmented reality and high-resolution display of information will be transformative to tasks such as navigation, worker training, as well as new forms of digital entertainment and the arts. Success also stands to have cross-disciplinary benefit to the domain of scientific computing, as hardware, software, and compilation technologies developed for graphics acceleration continue to have widespread impact on the advancement of general parallel computing. High-performance visual computing topics will be used as a central component of efforts to modernize parallel computing education. The results of this effort will be available to the public as an on-line textbook and web reference for modern parallel computing. The project web site (http://graphics.cs.cmu.edu/projects/renderergenerator/) provides additional information on the project.