Module tensor 

Sparse tensor operations for GNN (Graph Neural Network) workloads.

This module extends SpMV/SpMM primitives with operations tailored for graph neural network message passing and aggregation, including:

• scatter_reduce – Scatter-reduce with configurable aggregation
• gather – Index-based gathering from source arrays
• sparse_message_passing – GNN message passing over adjacency matrices
• sparse_attention_message – Attention-weighted message passing (GAT-style)
• compute_degree_matrix – Node degree computation from CSR offsets
• symmetric_normalize – D^{-1/2} A D^{-1/2} normalization (GCN-style)
• add_self_loops – Add identity self-loops to adjacency matrix
• sparse_row_softmax – Row-wise softmax over sparse matrix values
• generate_message_passing_ptx – PTX kernel generation for GPU message passing

(C) 2026 COOLJAPAN OU (Team KitaSan)

Structs

EdgeFeatures

Edge features for attention-weighted message passing.

GnnSparseConfig

Configuration for GNN sparse operations.

Enums

MessagePassingOp

Aggregation operation for GNN message passing.

Functions

add_self_loops

Add self-loops to an adjacency matrix: A’ = A + I.

compute_degree_matrix

Compute the degree of each node from CSR row offsets.

gather

Gather operation.

generate_message_passing_ptx

Generate a PTX kernel for GPU-accelerated GNN message passing.

scatter_reduce

Scatter-reduce operation.

sparse_attention_message

Attention-weighted message passing.

sparse_message_passing

Sparse message passing over an adjacency matrix in CSR form.

sparse_row_softmax

Row-wise softmax over sparse matrix values.

symmetric_normalize

Symmetric normalization of an adjacency matrix: D^{-1/2} A D^{-1/2}.