Dynamic Scheduling for Data-Parallel Path Tracing of Large-Scale Instanced Scenes.

Area of Science:

• Computer Graphics
• High-Performance Computing

Background:

• Data-parallel ray tracing is essential for rendering large scenes beyond single-node memory.
• Communication overhead is a major performance bottleneck in distributed rendering.
• Current methods using spatial partitioning cause data redundancy in instanced scenes.

Purpose of the Study:

• To develop a novel dynamic scheduling algorithm for efficient data-parallel ray tracing.
• To address the challenge of redundant data storage in large-scale instanced scenes.
• To improve rendering performance by reducing communication overhead.

Main Methods:

• Introduced the first object-space-based dynamic scheduling algorithm using object groups as scheduling units.
• Developed two data access frequency prediction methods for asynchronous data prefetching.
• Eliminated redundant storage of base geometry in instanced scenes.

Main Results:

• Achieved an average rendering speedup of 77.6% compared to state-of-the-art methods.
• Demonstrated a maximum rendering speed improvement of up to 146.1%.
• Increased scene memory consumption by only 5%.

Conclusions:

• The proposed object-space dynamic scheduling algorithm effectively reduces communication overhead and data redundancy.
• This approach significantly enhances rendering efficiency for large-scale instanced scenes.
• The method offers substantial performance gains with minimal memory overhead.