SDM Refactor

[debog]

Refactors the warm SDM implementation to reduce per-tile boilerplate, consolidate GPU memory management, improve numerical precision in templated code, and add GPU-direct distribution sampling. No physics changes; all 8 SDM ctests pass with identical results.

New files

• Source/Particles/ERF_SuperDropletPCProcess.H: Defines SDProcess::ProcessContext and SDProcess::ParticlePointers structs that bundle geometry/species metadata and SoA attribute pointers, respectively. These replace the repeated manual extraction of plo, dxi, num_species, num_aerosols, species/aerosol mass pointer arrays, etc. at the top of every particle kernel.
• Source/Utils/ERF_InterpolationUtils.H: Generic multi-field cloud-in-cell (CIC) interpolation helper ERF::Interpolation::interpolateFields() templated on an enum of field indices. Replaces duplicated per-field interpolation code in advection, mass change, and boundary routines.

Core infrastructure (ERF_SuperDropletPC.H, ERF_SuperDropletPCDefinitions.H)

• forEachParticleTile(): Three overloads (with ProcessContext, lightweight, serial) that replace manual #pragma omp parallel + ParIter loops throughout the codebase. The serial variant (forEachParticleTileSerial) is used where thread safety is required (e.g., DenseBins in coalescence).
• buildProcessContext(lev): Consolidates extraction of geometry arrays, species/aerosol counts and indices, water density, and device property pointers into a single ProcessContext struct.
• setupParticlePointers(): Consolidates extraction of SoA attribute pointers (velocity, mass, radius, multiplicity, terminal velocity, species masses, aerosol masses) into a ParticlePointers struct.
• updateParticleAttributes(): Static AMREX_GPU_DEVICE method that recomputes effective radius and total mass from species/aerosol mass arrays. Replaces inline SD_effective_radius + SD_total_mass calls scattered across boundary, recycle, and add-particle routines.
• Persistent device vectors for material properties (m_sp_density, m_sp_solubility, m_sp_ionization, m_sp_mol_weight, and corresponding m_ae_* arrays): allocated once via initializeDeviceProperties() and invalidated on setSpeciesMaterial/setAerosolMaterial. Replaces per-tile Gpu::DeviceVector allocations + host-to-device copies that previously occurred inside every particle loop.
• particleToMeshHelper(): Template for particle-to-mesh deposition with support for non-uniform vertical grids (m_zlevels_d). Refactors SDNumberDensity, numberDensity, massDensity, massFlux, speciesMassDensity, and the new cloudRainDensity to delegate to this helper.
• ridx_a/ridx_s refactoring: These runtime-offset index functions are now the base implementations (without SuperDropletsRealIdxSoA::ncomps offset); idx_a/idx_s are redefined to call them with the compile-time offset added.
• SDTerminalVelocityType converted from enum struct to AMREX_ENUM, enabling direct ParmParse parsing and getEnumNameString() for terminal velocity model selection.
• SuperDropletsIntIdxSoA typedef added for integer SoA component access.
• m_term_vel_type renamed to m_term_vel_type_w; m_idx_w initialized to -1.

Distribution sampling (ERF_SDInitialization.H/.cpp, ERF_SuperDropletPCAddParticles.cpp)

• SDDistributionType AMREX_ENUM replaces the SupDropInit::attrib_init_* string constants. All distribution type comparisons, ParmParse queries, and printing now use the enum directly instead of string matching.
• SDDistributionParams struct: GPU-compatible POD holding distribution type, mass/radius bounds, pre-computed CDF bounds for truncated log-normal, and multiplicity parameters.
• SD_erfinv_gpu(): AMREX_GPU_HOST_DEVICE inverse error function approximation for GPU-side log-normal inverse CDF sampling.
• SD_sample_mass_gpu(): AMREX_GPU_HOST_DEVICE function that generates a mass sample from SDDistributionParams + a RandomEngine, supporting constant, exponential, and log-normal (including autorange) distributions with both sampled and constant multiplicity modes.
• getSpeciesDistParams(), getAerosolDistParams(), makeDistributionParams(): Build SDDistributionParams from the existing member vectors, pre-computing log-space ranges and CDF bounds on the host.
• addParticles(): Replaced host-side getSpeciesDistribution/getAerosolDistribution calls (which allocated O(num_species * np) host vectors, sampled on CPU, then copied to device) with a single ParallelForRNG kernel that calls SD_sample_mass_gpu directly into the particle SoA. Per-tile device vectors for material properties replaced with persistent device arrays.
• setDefaults() now checks m_is_water to set species-specific defaults (water gets tiny initial mass; non-water species get zero).

Per-file refactoring

• ERF_SuperDropletPCAdvection.cpp: Rewritten with forEachParticleTile + buildProcessContext + interpolateFields. InterpFieldsAdv enum for density/pressure/temperature interpolation. Post-Redistribute k-index update loop added.
• ERF_SuperDropletPCBoundaries.cpp: Rewritten with forEachParticleTile + buildProcessContext. Uses updateParticleAttributes instead of inline radius/mass recalculation.
• ERF_SuperDropletPCMassChange.cpp: Rewritten with forEachParticleTile. InterpFieldsLV enum for saturation/pressure/temperature interpolation. Solver objects (drsqdt, newton_solver) moved outside the tile loop.
• ERF_SuperDropletPCRecycle.cpp: Rewritten with forEachParticleTile for all three iteration blocks (recycle, post-redistribute location update, remove inactive). Division-by-zero guard added for deac_frac computation.
• ERF_SuperDropletPCDiagnostics.cpp: Individual ReduceMin/ReduceMax/ReduceSum calls replaced with a single fused ReduceData<>/ReduceOps pass. Individual MPI reductions replaced with batched array ReduceRealMin/ReduceRealMax/ReduceRealSum calls.
• ERF_SuperDropletPCInitializations.cpp: Terminal velocity parsing uses AMREX_ENUM directly instead of string comparison. AMREX_ASSERT changed to AMREX_ALWAYS_ASSERT for species/aerosol count checks. initializeDeviceProperties() called in define(). Z-levels reading added for non-uniform grids. SetAttributes and DensityScaling rewritten with forEachParticleTile.
• ERF_SuperDropletPCUtils.cpp: computeMeshVar refactored with early returns and cleaner dispatch. SDNumberDensity, numberDensity, massDensity, massFlux delegate to particleToMeshHelper. cloudRainDensity added (previously was a 4-argument overload of speciesMassDensity).

Precision and constant fixes

• ERF_SuperDropletPCMassChange.H: Include guard fixed (COALESCENCE -> MASSCHANGE). Namespace renamed SDMassChangeUtils -> SDMassChangeUtils_LV. Named constants (one, zero, myhalf, two, three) replaced with RT() casts throughout ODE functions. std::exp(-myhalf*std::log(R_sq)) replaced with RT(1.0)/std::sqrt(R_sq). Helper methods added to TimeIntegrator: computeTimestep, computeTau, limitTimestep, isTimestepTooSmall, evalRHS, printStepInfo, printStepInfoNewton.
• ERF_SuperDropletPCCoalescence.H: amrex::Real(...) replaced with RT(...) throughout all data tables and kernel functions. Named constants replaced with RT() casts.
• ERF_TerminalVelocity.H: m_rho renamed to m_rho_w. Viscosity calculation consolidated into viscCoeff(a_T) method (was inline in CloudRainShima). All amrex::Real(...) replaced with RT(...).
• ERF_Constants.H: four_thirds_pi constant added. Used throughout in place of Real(4.0)/three*PI or (amrex::Real(4.0)/three)*PI.

Misc

• ERF_SuperDropletsMoist.H: GetPlotVar refactored with try_copy helper lambda.
• ERF_SuperDropletsMoistUtils.cpp: speciesMassDensity 4-argument calls updated to cloudRainDensity.
• Backward-compatibility block removed from SDInitProperties::readInputs (the per-species keyed queries supersede the old condensate_* keys).