Architecture¶
This section explains how RayON is structured internally: how a scene is built and handed to the two GPU back-ends, how the CUDA kernel is launched, and how the interactive renderer accumulates samples and keeps its frame rate.
The "build once, render anywhere" design: one neutral SceneDescription converted into flat GPU structs for CUDA and into acceleration structures and a shader binding table for OptiX.
32×8 thread blocks, thread→pixel mapping, random states, the float accumulation buffer, and the C++/CUDA boundary.
The accumulation loop, adaptive sampling, the motion regulator and denoiser, the display pipeline and live parameter updates.