Hector: An Efficient Programming and Compilation Framework for Implementing Relational Graph Neural Networks in GPU Architectures

Keyphrases

• Access Scheme 20%
• Accessible Graphs 20%
• Art Publics 20%
• Backward Propagation 20%
• C++ Code 20%
• Code Generator 20%
• Compiler Framework 100%
• CUD A 20%
• Data Access 20%
• Data Copy 20%
• Data Heterogeneity 20%
• Data Layout 40%
• Deep Graph Library 20%
• Design Challenges 20%
• Forward Propagation 20%
• General Matrix 20%
• GPU Architecture 100%
• Graph Benchmark 20%
• Graph Neural Network 100%
• Heterogeneous Graph 40%
• High-throughput 20%
• Intermediate Representation 20%
• Interoperator 20%
• Latency 20%
• Linear Operator 20%
• Lines of Code 20%
• Matrix multiplication 20%
• Memory Access 20%
• Memory Efficiency 20%
• Model Semantics 20%
• Neural Network Model 20%
• Operator Scheduling 20%
• Out-of-memory 20%
• Outer Product 20%
• Performance-based Design 20%
• Programming Effort 60%
• Programming Framework 100%
• Programming Interface 20%
• Public System 20%
• Real-world Application 20%
• Redundant Data 20%
• Relational Graph 100%
• Semantic Data 20%
• System Design 20%
• Traversal 20%
• Two-level 20%

Computer Science

• Access Schemes 20%
• Code Generator 20%
• Data Access 20%
• Data Layout 40%
• Forward Propagation 20%
• Graph Neural Network 100%
• Graphics Processing Unit 100%
• High Throughput 20%
• Intensive Computation 20%
• Intermediate Representation 20%
• Linear Operator 20%
• Lines of Code 20%
• Matrix Multiply 20%
• Memory Access 20%
• Neural Network Model 20%
• Redundant Data 20%
• Relational Graph 100%
• Relational Graph Convolution Network 20%
• Speed-up 40%
• World Application 20%