Backends & API reference

AddingDoublingRT provides three backends that implement the same algorithm and produce matching results. Choose the one that fits your workload:

Backend	Best for	Precision	Parallelism
C++ (CPU)	Single spectral points, one profile at a time, temperature Jacobians.	double	Template-specialised fixed-size kernels (\(N = 2, 4, 8, 16, 32\)).
CUDA (GPU)	Whole spectra, retrieval inner loops.	double LU, mixed	One CUDA thread per wavenumber, batched.
JAX (CPU/GPU)	Differentiable pipelines, research, prototyping.	float64	Vectorised across wavenumbers; JIT-compiled; autodiff.

All three share the same physics and configuration model (User guide). The pages below document each backend’s public API and any backend-specific data-layout or precision considerations.

• C++ (CPU) backend
  • Solver entry points
  • ADConfig
  • RTOutput
  • SolverWorkspace
  • Exposed utilities (for testing)
• CUDA (GPU) backend
  • Configuration
  • Device data layouts
  • Entry points
• JAX (CPU/GPU) backend
  • Single-wavenumber solver
  • Batched solver
  • Differentiation
  • Module-level helpers