User Guide

The user guide explains the runtime choices that define a PM++ experiment. Start with configuration and initial conditions, then add evolution, analysis, gradients, or multiple devices as needed.

The examples run on multiple GPUs and exercise PM++’s distributed ownership, communication, and FFT paths.

• Configuration: geometry, units, precision, schedules, compilation, and static capacities.

• Precision, reproducibility, and gradients: float32 atomic variation, float64 guidance, deterministic execution, and adjoint validation.

• Initial conditions: Gaussian, phase-only, and nested noise; transfer functions; and LPT.

• Evolution and analysis: N-body runs, observers, density projections, and spectra.

• Differentiation: supported gradient targets and validation strategy.

• Multi-GPU execution: device meshes, mesh_halo, ownership, distributed FFTs, and capacity planning.

• Optimizations: measured forward and AD settings and portable fallbacks.

• Pallas CIC kernels: the paired particle-mesh kernels, qualification, and reference fallback.

• CUDA routing: optional FFI setup for accelerated local mesh-halo particle routing.

• Troubleshooting: overflows, decompositions, compilation, and memory failures.

The notebook gallery pairs these explanations with pre-executed examples.