NVIDIA Researchers Demonstrate New Raytracing Algorithm That Can Render Direct Lighting from Millions of Dynamic Light Sources

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,786
Almost Billions of Dynamic Light Sources! Almost!

"In recent years, Monte Carlo path tracing has been widely adopted for offline rendering [Christensen and Jarosz 2016; Fascione et al. 2017] and is seeing increasing use in real-time applications [Schied 2019] with the arrival of specialized hardware support for ray intersection tests [Parker et al. 2010; Wyman et al. 2018]. Even in offline rendering, without the constraints of real-time, direct lighting with many emissive objects remains challenging; it’s not feasible to trace shadow rays to all of the lights, and finding the lights that contribute most at a given point depends on each light’s visibility to that point, the distribution of the scattering function (BSDF or phase function) at the point, and the light source’s power and emissive characteristics. Real-time rendering adds even more challenges: the scenes to be rendered are dynamic and the renderer generally has no future knowledge of how the scene will change, as that may be affected by user interaction. Furthermore, only a few rays can currently be traced at each pixel, so finding important lights is even more critical, yet there is a limited amount of time to build and update data structures to aid light sampling [Moreau et al. 2019]. This is true even for the restricted case of direct lighting at the first camera vertex, which we consider in this paper. These constraints have spurred research in denoising and reconstructing images from noisy low-sample-per-pixel rendered images. While great strides have been made in this area in both offline [Vogels et al. 2018] and real-time [Schied et al. 2018] rendering, a limited amount of processing time is available for real-time denoisers since time spent filtering takes away from the available frame time. Denoising is particularly challenging with low sample-count images; as shown in Fig. 2, improving the quality of samples provided to a denoiser can significantly increase its effectiveness. We introduce a method to sample one-bounce direct lighting from many lights that is suited to real-time ray tracing with fully dynamic scenes (see Fig. 1). Our approach builds on resampled importance sampling (RIS) [Talbot 2005], a technique for taking a set of samples that are from one distribution and selecting a weighted subset of them using another distribution that better matches the function being integrated. Unlike prior applications of RIS, we use a small fixed-size data structure—a “reservoir” that only stores accepted samples—and an associated sampling algorithm (used frequently in non-graphics applications [Efraimidis and Spirakis 2006]) to help achieve stable, real-time performance. Given the reservoir, our approach does not use any data structures more complicated than fixed-size arrays, yet it stochastically, progressively, and hierarchically improves each pixel’s direct light sampling PDF by reusing statistics from temporal and spatial neighbors. In contrast to modern real-time denoising algorithms that reuse pixel colors across temporal and spatial neighborhoods, our reuse informs the sampling probabilities used within the renderer, which in turn makes an unbiased algorithm possible. Our unbiased mode can be modified to be biased, which further reduces noise at the cost of some over-darkening near geometric discontinuities. We demonstrate our algorithms running interactively on a single GPU with scenes that have thousands to millions of dynamic lights, obtaining one to two orders of magnitude speedup for the same error compared to state-of-the-art methods implemented on the same hardware. We cover the mathematical preliminaries of the techniques we build upon in Section 2 before describing our work in the subsequent sections. We discuss related work in Section 7, for better context when comparing with our results."

https://news.developer.nvidia.com/rendering-millions-of-dynamics-lights-in-realtime/

An interpretation exists by WCCFTech: https://wccftech.com/nvidia-researc...hting-from-millions-of-dynamic-light-sources/
 
I need AI to read me all that!!😵
Any versions in less than a paragraph?
(I hear being lazy is new healthy, so here taco am..)

Math and smarts make telling computer to draw light go vroooom

Imagine tires squealing noise at this point, the marks on the ground when the smoke clears looks suspiciously like the letters "RTX"
 
2070_super_duper.jpg
 
Does this requiuire Tensor units to pull-off in real-time, or will any RT-capable card be able to do this with shaders?
 
Back
Top