OAC Core: OAC Core Projects: GPU Geometric Data Processing

OAC 核心：OAC 核心项目：GPU 几何数据处理

• 批准号: 2403239
• 负责人: Jonathan Ragan-Kelley
• 金额: $ 60万
• 依托单位: Massachusetts Institute of Technology
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2024
• 资助国家: 美国
• 起止时间: 2024-07-01 至 2027-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2403239&HistoricalAwards=false
• 关键词: OAC; Core; Projects; GPU; Geometric

Geometric data processing is vital across various domains, from architecture and design to medical imaging, robotics, animation, and entertainment. The integration of machine learning has propelled advancements in the field and unlocked applications for computer vision and generative AI-driven design. However, the reliance on CPU-based serial processing limits the speed of processing complex and detailed geometries. This project seeks to leverage Graphics Processing Units’ (GPUs’) parallel computing capabilities to improve geometric data processing speed, efficiency, and scalability. It introduces GPU-optimized algorithms and data structures, offering significant speed improvements for unstructured mesh processing, a common but challenging-to-optimize task essential to many applications.This project addresses the inefficiencies in 3D geometric data processing by designing data structures that maximize the potential of GPU and multi-GPU systems. It focuses on three primary research directions: 1) Dynamic Unstructured Mesh Processing, to process dynamic changes in mesh topology entirely on the GPU; 2) Parallel Intrinsic Mesh Processing, managing extrinsic and intrinsic mesh representations simultaneously to ensure optimal parallel efficiency of theoretically robust algorithms; and 3) Neural Mesh Compression, to compress mesh attributes using neural networks, maximizing the size of meshes that may be processed. Tailoring these strategies specifically for GPU architectures aims to accelerate the entire spectrum of geometric data processing applications, from large-scale simulations to interactive computer graphics. Additionally, the project seeks to make these advancements broadly accessible by releasing them in high-quality, open-source libraries, encouraging the adoption of GPU-accelerated geometric processing techniques, and providing developers and researchers with a comprehensive set of tools to explore new possibilities within their fields.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

几何数据处理在各个领域都至关重要，从建筑和设计到医学成像，机器人，动画和娱乐。机器学习的集成推动了该领域的进步，并为计算机视觉和生成式AI驱动设计解锁了应用程序。然而，对基于CPU的串行处理的依赖限制了处理复杂和详细几何形状的速度。该项目旨在利用图形处理单元（GPU）的并行计算能力，以提高几何数据处理的速度，效率和可扩展性。它引入了GPU优化的算法和数据结构，为非结构化网格处理提供了显著的速度提升，这是一项常见但具有挑战性的优化任务，对许多应用至关重要。该项目通过设计最大化GPU和多GPU系统潜力的数据结构来解决3D几何数据处理中的低效问题。它主要集中在三个主要的研究方向：1）动态非结构化网格处理，完全在GPU上处理网格拓扑的动态变化; 2）并行内在网格处理，同时管理外在和内在网格表示，以确保理论上鲁棒算法的最佳并行效率;以及3）神经网格压缩，使用神经网络压缩网格属性，最大化可以处理的网格的大小。专门为GPU架构定制这些策略旨在加速从大规模模拟到交互式计算机图形的整个几何数据处理应用程序。此外，该项目旨在通过在高质量的开源库中发布这些进步，鼓励采用GPU加速的几何处理技术，该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准。