gko::batch

Batched linear algebra: solve many small independent systems at once on the same executor, sharing kernel launches and data movement to keep accelerators saturated. Every type in this namespace has the uniform-batch constraint — all items share the same dimensions and the same sparsity pattern (CSR/ELL store the column indices and row pointers once, common across the whole batch).

For background on when batched solvers help, see the Batched concept page.

Right-hand sides and matrices

• MultiVector — the right-hand-side container: one dense block per batch item, all sharing the same shape.

• matrix::Csr — batched compressed sparse row.

• matrix::Dense — batched column-major dense.

• matrix::Ell — batched equal-length-rows.

• matrix::Identity — batched identity LinOp.

Solvers

• solver::Cg — batched conjugate gradient (SPD).

• solver::Bicgstab — batched stabilised biconjugate gradient (non-symmetric).

• solver::BatchSolver — common runtime setters (tolerance, max-iterations, preconditioner) shared by every batched solver.

Preconditioners

• preconditioner::Jacobi — batched block-Jacobi.

Loggers

• log::BatchConvergence — per-batch-item iteration count and residual norm; the batched equivalent of gko::log::Convergence.