Batch solving

cubie.batchsolving

The Solver interface is the main entry point. It brings together batch grids, array managers, GPU program builds, and system details into a coordinated integration pipeline, while solve_ivp() offers a shortcut for common workflows. Supporting modules prepare the GPU programs, describe solver outputs, and provide data checks used by attrs-based containers.

Subpackages

• Arrays

Core API

• Solver – high-level manager that drives CUDA kernel launches.

• solve_ivp – convenience wrapper for single-run solver configuration.

• SolveResult – captures state, summaries, and diagnostic metadata.

• SolveSpec – checked configuration describing a solver invocation.

• Calibration – races solver configurations on a representative batch and picks the fastest (Solver.calibrate()).

• Optimisation – times buffer placement, loop unrolling and launch shape for one configuration and keeps the fastest (Solver.optimize()).

Supporting infrastructure

• BatchSolverConfig – attrs container that tracks solver metadata.

• BatchSolverKernel – compiles device kernels for batch integration.

• SystemInterface – adapts cubie.odesystems.baseODE.BaseODE instances for CUDA execution.

Dependencies

• Batch kernel compilation extends cubie.CUDAFactory and relies on cubie.integrators.algorithms and cubie.integrators.step_control for loop callables, and cubie.array_interpolator for driver interpolation.

• Array managers size buffers with cubie.outputhandling.output_sizes and allocate memory through cubie.memory managers.

• Solver results surface summary metrics registered via cubie.outputhandling.summarymetrics.