JiTTree: A Just-in-Time Compiled Sparse GPU Volume Data Structure.

Area of Science:

• Computer Graphics
• Data Structures
• Scientific Visualization

Background:

• Sparse volume data structures are crucial for efficient GPU memory representation of large, sparse datasets.
• Selecting an optimal sparse data structure is challenging due to varying data characteristics like sparsity and access patterns.
• Existing hybrid structures offer adaptability but often incur significant traversal overhead.

Purpose of the Study:

• To present JiTTree, a novel sparse hybrid volume data structure designed for GPU computation and visualization.
• To address the limitations of existing sparse data structures, particularly traversal overhead in hybrid approaches.
• To enable efficient adaptation to diverse datasets with locally varying sparsity.

Main Methods:

• Developed JiTTree, a hybrid data structure combining multiple sparse representations.
• Integrated just-in-time (JIT) compilation to minimize traversal overhead.
• Evaluated performance and memory usage across various datasets with differing sparsity levels.

Main Results:

• JiTTree demonstrates superior adaptability to a wide range of sparse datasets.
• The structure significantly outperforms other sparse data structures for datasets with locally varying sparsity.
• Just-in-time compilation effectively reduces traversal overhead, enhancing computational efficiency.
• JiTTree achieves better memory usage compared to non-hybrid sparse data structures.

Conclusions:

• JiTTree offers an effective solution for sparse volume representation on GPUs, balancing memory and performance.
• The hybrid approach combined with JIT compilation overcomes the limitations of traditional sparse structures.
• This novel data structure is particularly advantageous for applications with heterogeneous sparsity and demanding computational requirements.