Module batched 

Batched sparse operations.

This module provides routines for executing many small sparse operations in a single logical batch, amortising kernel launch overhead and enabling fusion when beneficial.

Provided types

• BatchedSpMV – batched sparse matrix-vector multiply (host-side loop).
• BatchedSpMVPlan – precomputed plan with concatenated CSR arrays.
• UniformBatchedSpMV – optimised path when all matrices share structure.
• BatchedSpGEMM – batched sparse-sparse matrix multiply.
• BatchedTriSolve – batched sparse triangular solve.
• BatchScheduler – heuristic-based execution strategy selector.

Structs

BatchScheduler

Heuristic-based scheduler that decides how to execute a batch.

BatchedSpGEMM

Batched sparse-sparse matrix multiplication.

BatchedSpMV

Batched sparse matrix-vector multiplication.

BatchedSpMVPlan

Precomputed plan for batched SpMV execution.

BatchedTriSolve

Batched sparse triangular solve.

UniformBatchedSpMV

Optimised batched SpMV when all matrices share the same sparsity pattern.

Enums

Strategy

Execution strategy chosen by BatchScheduler.

Functions

batched_spmv_cpu

Batched SpMV (CPU reference): compute y_i = A * x_i for all b RHS vectors.

generate_batched_spmv_ptx

Generate a batched SpMV PTX kernel that processes all matrices via offset arrays in a single launch.

mixed_precision_spmv_cpu

Mixed-precision SpMV (CPU reference): matrix values stored in FP16 range, accumulation performed in FP64, result cast to FP32.