Quick Clusters: A GPU-Parallel Partitioning for Efficient Path Tracing of Unstructured Volumetric Grids.

Area of Science:

• Computer Graphics
• Scientific Visualization
• Computational Geometry

Background:

• Unstructured volumetric grids present challenges in preprocessing and rendering performance.
• Existing methods often require auxiliary connectivity data and extensive offline preprocessing.
• Bounding Volume Hierarchies (BVHs) can be memory-intensive when applied to individual elements.

Purpose of the Study:

• To develop a method for clustering finite elements to improve rendering performance and reduce preprocessing times.
• To enhance the efficiency of unstructured volumetric grid visualization without auxiliary connectivity data.
• To enable high-quality interactive visualization on consumer graphics hardware.

Main Methods:

• Elements are sorted along a Hilbert curve to aggregate neighboring elements, reducing BVH memory consumption.
• On-the-fly mesh clustering and re-indexing create sub-meshes for efficient ray tracing.
• Adaptive sampling guided by maximum extinction bins and ray-aligned bins optimizes sample placement.

Main Results:

• Rendering performance on the NASA Mars Lander dataset improved by 6x (up to 6 fps, 12.4 M rps) with volumetric shadows.
• Memory consumption was reduced by 3x (from 33 GB to 11 GB).
• Performance was further boosted by 17x (up to 17 fps, 35.6 M rps) using an RTX 8000 GPU.

Conclusions:

• The proposed Hilbert curve-based clustering and on-the-fly re-indexing method effectively improves rendering speed and reduces memory footprint.
• The approach eliminates the need for offline preprocessing, enabling interactive visualization of complex volumetric data.
• This technique offers significant advantages for real-time scientific visualization applications.