Scalable Realistic Rendering

Improving and unifying the human visual experience across the wide range of available devices will benefit users across a wide range of scientific fields and consumer applications. In the scalable rendering project we aim to improve this visual experience by creating models of these computational devices and then using those models to enhance the realistic images that are generated.

The current range of computational devices includes hand held devices such as tablet computing, desktop systems and continues all the way up to supercomputers, which are accessible via networks.
These platforms cover an ever increasing range of hardware architectures including parallel and heterogeneous, with CPU and GPU like properties.

Making use of these complex architectures is challenging, ensuring good performance and utilization, even more so. This project will build models that predict the performance of a particular algorithm on a particular system and select the correct rendering algorithm for that platform.
We will also break down existing graphics algorithms to understand how they map to varying processing models and create new scalable realistic rendering algorithms. We will start with realistic techniques such as our recent work into analytical motion blur and move into traditional image rendering techniques such as ray tracing.

Hardware architectures are either designed for highly efficient single instruction performance or highly parallel instruction performance.
Highly parallel processors group operations in a SIMD like fashion.
Both architecture types are supported by a hierarchy of cache architectures and performance is determined by the use of computational power and memory bandwidth. Memory bandwidth usage is based on the complex usage of caches that hide memory latency based on a programs behavior. By analysing a program's properties and estimating it's performance on a specific architecture, decisions between competing program optimization and algorithm choices can be made to achieve the best performance.

Project team

Principal Investigator
Prof. Dr.-Ing. Michael Doggett

Per Ganestam