Crate oxicuda_gnn 

oxicuda-gnn — Graph Neural Network primitives for OxiCUDA.

Pure-Rust implementation of GNN building blocks suitable for CPU simulation and PTX kernel generation for GPU execution.

Architecture

oxicuda-gnn
├── graph/          — Sparse graph representations (CSR, COO, Heterogeneous, Sampling)
├── message_passing — Aggregate, Scatter, Update primitives
├── layers/         — GCN, GAT, GATv2, GraphSAGE, GIN
├── pooling/        — Global pool, Top-K pool, DiffPool
├── readout/        — Set2Set
├── error           — GnnError / GnnResult
├── handle          — GnnHandle (SmVersion + LcgRng)
└── ptx_kernels     — GPU PTX kernel strings

Modules

conv

GNN convolution primitives (self-contained, weight-bearing layers).

error

Error types for the oxicuda-gnn crate.

graph

Graph representations for GNN operations.

handle

Session handle for oxicuda-gnn.

layers

GNN layer implementations.

message_passing

Message passing framework for graph neural networks.

ops

Graph neural network operation primitives.

pooling

Graph pooling operations.

prelude

Convenience re-exports for common GNN types.

ptx_kernels

PTX GPU kernel sources for GNN operations.

readout

Graph readout functions.

sampling

Graph sampling algorithms for mini-batch GNN training.