An overview and implementation of reservoir-based spatiotemporal importance resampling (ReSTIR) for real-time ray-tracing

Abstract

Simulating real-world lighting in real-time has consistently been a challenging task, even for the latest generation of computers. Ray-tracing, an algorithm that simulates the physics of light, has long been recognized as the highest quality method for rendering realistic images, while the cost of performance is too high when compared to other methods like Rasterization. With the advent of new-generation Graphics Processing Units (GPUs), it has become feasible to leverage these devices’ highly parallel processing capabilities to accelerate ray-tracing algorithms. This advancement offers significant potential for enhancing real-time rendering by utilizing the parallelism inherent in modern GPUs. However, despite these technological advances, real-time ray-tracing continues to create significant challenges across all GPU architectures. This process involves complex calculations for reflection, refraction, shadow, and diffusion, all of which must be computed in real-time to achieve the desired visual effects. In response to these challenges, my role at Futuremark involves exploring and implementing new methods to determine their viability for enhancing our rendering capabilities. By experimenting with different techniques and optimizing the use of GPU architectures, we aim to push the boundaries of what can be achieved with real-time ray-tracing. This includes not only improving performance but also ensuring that the visual quality meets the high standards required for realistic rendering. The goal of this thesis is to find a balance where we can utilize ray-tracing in real-time applications effectively without compromising neither performance or visual fidelity. This thesis has been conducted to evaluate the benefits of a new method called Reservoir-based Spatiotemporal Importance Resampling (ReSTIR) for real-time ray-tracing with dynamic direct lighting in our engine.