HyperGef

Efficient kernel fusion for hypergraph neural networks on GPUs.

Hypergraph neural networks run two rounds of irregular gather/scatter per layer — vertex-to-edge, then edge-to-vertex. Executed as separate kernels, that pattern spends most of its time moving intermediate results through memory.

HyperGef is a framework that fuses those stages, with a workload-balancing scheme that keeps the fused kernel from stalling on skewed hyperedge degrees.

Published at MLSys 2023.