Developing efficient and advanced rendering and lighting code is difficult in itself but becomes almost impossible when it should take good advantage of the increasingly different hardware architectures and specialized hardware components. However, best exploiting available resources becomes increasingly important across the device continuum from mobile to large compute centers. Some companies provide highly optimized ray tracing libraries (e.g. Optix and Embree) but each one is optimized only for a specific hardware architecture and took years to reach its current performance levels.
In this talk we present AnyDSL, a new compiler technology that allows for specifying advanced algorithms in simple, high-level code and for easily mapping it to very different HW architectures while taking full advantage of each of them without changing the high-level code. In AnyDSL a developer can simply define his own domain specific architecture through a hierarchy of abstractions. At each level, the abstractions are refined in terms of lower-level and appropriate information is added (e.g. which spatial index to use, how implement vector instructions, etc.). Complete implementations can be exchanged for each other by simply substituting different libraries of abstractions, including taking advantage of hardware-specific features at lower levels. Through its novel design, the AnyDSL compiler can then specialize code across all levels, taking into account the information added at each of them. This is done by aggressively specializing code at compile time via interpreting everything that is known and only emiting the residual code.
We will present the AnyDSL compiler, show how it enables us to beats even highly hand-optimized code from different domains across all major hardware architectures despite writing only a few lines of high-level code. We then show how to apply it to rendering and discuss how it can be used to design highly optimized ray tracing and lighting simulation architectures that map optimally to existing and upcoming HW architectures.
Philipp Slusallek is Scientific Director at the German Research Center for Artificial Intelligence (DFKI), where he heads the research area “Agents and Simulated Reality” since 2008. He is also Director for Research at the "Intel Visual Computing Institute”, a central research institute at Saarland University founded in 2009 in collaboration with Intel, DFKI, and the two local Max-Planck-Institutes. At Saarland University he has been a professor for Computer Graphics since 1999 and a Principle Investigator at the German Excellence-Cluster on “Multimodal Computing and Interaction” since 2007. Before coming to Saarland University, he was a Visiting Assistant Professor at Stanford University, USA. He studied physics in Frankfurt and Tübingen (Diploma/M.Sc.) and got his PhD in Computer Science from Erlangen University. His research interests are focused on novel service-oriented architectures for 3D-Internet technology, integrating research in areas such as real-time realistic graphics, artificial intelligence, high-performance computing as well as security by design for creating distributed, immersive, collaborative environments for simulation, analysis, visualization, and training.
The research department Agents and Simulated Reality (ASR) at the DFKI conducts research in the areas visual computing, multi-agent systems and formal methods for safe and secure systems in the context of interactive 3D simulations and visualizations as well as other applications. Using Simulated Reality, real-world scenarios can be accurately depicted and predictions about their behavior can be made. Research in ASR allows for visualizing and simulating complex scenarios, such as industrial production lines, cities, or even biological cells. Semantic information is tightly integrated with the models and provides the context for the simulations and visualizations. A key focus is on real-time realistic rendering and lighting simulation by exploiting today's hardware capabilities to the fullest extend. Highly optimized software platforms for multi- and many-core systems enable physically-correct representations as well as direct interaction with complex 3D simulations. A special focus is on Web-based solutions directly in HTML-5 using our XML3D technology to guarantee a universal access and wide distribution.
Simulated Reality allows for accurately reproducing real world scenarios in the computer and making predictions about their behavior. Today, reliable predictions of the development and the behavior of complex systems are essential for the planning, development and operation of industrial plants, in production, trade, automotive and aerospace industries, as well as many other areas. The DFKI research department "Agents and Simulated Reality (ASR)" combines competencies from the core fields of Visual Computing and Artificial Intelligence, particularly in multi-agent systems and formal methods for safe and secure systems.