Description

This contains the first incremental steps toward making src/common genuinely stage-independent instead of selecting different implementations with MFC_SIMULATION, MFC_PRE_PROCESS, and MFC_POST_PROCESS.

The first slice focuses on m_variables_conversion:

• removes all stage-selection macros from the module;
• uses one host/device scalar species-to-mixture conversion kernel in every executable;
• makes primitive-to-conservative, primitive-to-flux, sound-speed, and fast-magnetosonic routines consistently available;
• moves material-property arrays, viscosity-index metadata, and direction mappings to the common global-parameter module so they have one declaration and GPU ownership point;
• replaces stage-dependent behavior with explicit initialization policies:
  • simulation enables the density floor and QBMM number preservation;
  • post-process requests mixture diagnostic storage and provides its Lagrangian beta index;
  • pre-process uses the common defaults;
• allocates and updates conversion data consistently, including Re_idx, Re_size, material arrays, and the new policy scalars.

The second slice makes shared grid-buffer population stage-independent:

• m_boundary_common now receives cell-boundary, cell-center, and cell-width arrays explicitly instead of selecting stage-global representations;
• simulation supplies its uniform buff_size cell-boundary extent, while post-process supplies its asymmetric Silo/output offsets;
• the one-dimensional physical-boundary algorithm is shared across x, y, and z instead of dispatching to direction-specific global-array routines;
• MPI grid-buffer exchange has an explicit one-dimensional array interface;
• simulation retains ownership of glb_bounds and performs the GPU update immediately after the host-side grid population.

The third slice removes the superseded grid implementations and separates minimum spacing from width-array storage:

• deletes the old direction-specific physical-boundary routines and module-global MPI grid exchange;
• introduces common dx_min, dy_min, and dz_min scalars;
• changes pre-process grid construction and smoothing calculations to use the explicitly named minimum spacings;
• leaves simulation/post-process dx, dy, and dz width arrays unchanged;
• removes the simulation/pre-process selection from STL model grid-spacing logic.

The fourth slice makes MPI I/O field binding and downsampled datatype construction stage-independent:

• passes QBMM pressure and moment fields explicitly instead of selecting pb/mv or pb_ts(1)/mv_ts(1) by stage;
• passes immersed-boundary marker storage and its MPI descriptor together, with a runtime consistency check;
• passes downsampled local extents explicitly instead of deriving different layouts from the executable stage;
• makes downsampled conservative-field pointer binding optional, so pre-process/simulation can construct file datatypes while post-process can additionally bind its input fields;
• removes six more stage guards from m_mpi_common, reducing the remaining stage-condition count in src/common from 30 to 24.

The fifth slice makes Dirichlet ghost filling an explicit runtime policy:

• replaces the simulation/post-process/pre-process implementation guard in s_dirichlet;
• initializes simulation to consume stored boundary buffers while pre-process and post-process retain boundary-cell extrapolation;
• declares and updates the policy scalar through the existing GPU portability macros so the shared GPU_ROUTINE takes the same branch on host and device;
• reduces the remaining stage-condition count in src/common from 24 to 23.

The sixth slice removes low-risk stage selection from common validation and finite differences:

• makes the total-cell check an explicit caller policy, with pre-process and post-process supplying nGlobal while simulation skips the check;
• removes empty stage guards from the common checker;
• makes callers own finite-difference coefficient allocation instead of reallocating only for post-process inside the shared coefficient routine;
• reduces the remaining stage-condition count in src/common from 23 to 19.

The seventh slice moves common global declarations, equation indexing, and GPU declarations to explicit ownership:

• generates num_dims, num_vels, weno_polyn, and muscl_polyn declarations for all executables, while preserving simulation case optimization;
• emits GPU declarations beside the generated simulation declarations that own the corresponding variables;
• keeps case-optimized parameters as compile-time constants without creating device storage for them;
• replaces the six-equation eqn_idx%alf stage guard with an explicit initialization policy, preserving simulation, pre-process, and post-process behavior;
• removes the stage branch from common dimension and velocity initialization;
• reduces the remaining stage-condition count in src/common from 19 to 14.

The eighth slice makes common MPI transport stage-independent:

• makes unified-memory preference depend only on NVIDIA unified-memory capability;
• compiles the shared stability reduction for every executable while retaining MPI/non-MPI capability selection;
• replaces the simulation/post-process chemistry halo branch with an explicit temperature-exchange policy;
• sizes the chemistry halo buffer using the same policy, including post-processing without diffusion;
• shares one staged/RDMA send-receive implementation for ordinary and Lagrangian-beta halos;
• initializes RDMA explicitly from simulation and disables it for pre-process/post-process;
• reduces the remaining stage-condition count in src/common from 14 to 9.

The ninth slice makes MPI domain decomposition stage-independent and removes its repeated directional policy code:

• passes Silo ghost-offset and local-domain-adjustment policies explicitly;
• uses thin pre-process and post-process adapters to pack stage-owned x/y/z structures into common three-direction arrays;
• computes interior Silo offsets for all active dimensions in one shared loop;
• computes local pre-process domain bounds and minimum spacings for all active dimensions in one shared loop;
• leaves topology selection, remainder distribution, parallel-I/O start indices, neighbor ranks, and boundary-rank construction unchanged;
• reduces the remaining stage-condition count in src/common from 9 to 3.

The tenth slice makes common STL model setup available in every executable:

• removes the post-process exclusion around s_instantiate_STL_models;
• keeps one shared implementation of model loading, transformation, bounding boxes, two-dimensional boundary extraction, GPU packing, and device updates;
• extends generated declaration ownership for num_stl_models and stl_models to post-process without adding either variable to its namelist;
• teaches typed generated declarations to honor declaration targets independently of input targets;
• reduces the remaining stage-condition count in src/common from 3 to 1.

The eleventh slice makes the full common phase-change module available in every executable:

• removes the post-process exclusion around the module imports, public interface, and solver implementations;
• keeps the existing pT/pTg relaxation algorithms, tolerances, failure handling, and accelerator loops unchanged;
• extends declaration ownership for palpha_eps and ptgalpha_eps to post-process without adding them to its namelist;
• preserves the current asserting behavior of the otherwise-empty relaxation dispatcher;
• removes the final executable-stage conditional from src/common.

Design choices

Stage identity is no longer used as a proxy for accelerator execution. Code inside accelerator regions selects the device-callable kernel with MFC_GPU; CPU execution uses the host wrapper, which can also populate optional post-processing diagnostics.

The scalar conversion kernel is marked with GPU_ROUTINE and is shared by CPU and accelerator paths. Device-resident module data retains explicit GPU_DECLARE, enter-data, update, and delete lifecycle operations. Material arrays and viscosity metadata are declared only once, in their common owning module.

Some variables must now be declared in all three executables because common code references them, but they should not become accepted inputs in every namelist. The parameter generator therefore separates Fortran declaration targets from namelist targets. Existing user-facing scopes for avg_state, alt_soundspeed, mixture_err, sigR, viscous, and riemann_solver remain unchanged.

This change deliberately does not remove capability guards such as MFC_GPU or backend-specific GPU macros. It removes stage-based implementation selection.

Pre-process currently uses dx, dy, and dz as scalar minimum spacings, whereas simulation and post-process use those names for width arrays. Passing the grid representation explicitly allows the common boundary implementation to compile for every executable without hiding that semantic mismatch behind another stage macro.

That mismatch is resolved in the third slice: scalar minimum spacings now have distinct common names, while per-cell width arrays retain the established dx, dy, and dz names.

MPI I/O now follows the same ownership rule: common code constructs views and binds only the objects supplied by its caller. Paired optional data are checked together, preventing a marker without a descriptor or one QBMM node field without the other. Accelerator ownership is unchanged because these routines set host-side MPI pointer/view metadata and introduce no device-data movement.

Dirichlet behavior is configured when the boundary module is initialized, rather than inferred from the executable name. The policy is device resident because s_dirichlet is called inside accelerator boundary loops.

Finite-difference routines now compute into caller-provided storage. This keeps allocation and lifetime with the executable that owns the array and avoids hiding post-process reallocation inside common numerical code.

Generated declarations are now the ownership boundary for generated simulation inputs. The generator emits exactly one GPU_DECLARE for each device-resident scalar it declares. Manually declared common state retains direct portability macros in the common module. Capability guards such as MFC_GPU, OpenACC, and OpenMP remain intact; only executable-stage selection is being removed.

MPI transport now has one implementation. Executables configure whether all chemistry temperatures need halo exchange and whether RDMA transport is enabled; the common layer owns buffer sizing and selects the corresponding transport path. OpenACC/OpenMP data movement remains expressed through the portability macros.

MPI decomposition owns the Cartesian algorithm, while executables provide only the data and policies that are actually stage-specific. Pre-process supplies mutable domain bounds and requests local-domain adjustment for generated grids; post-process supplies mutable output offsets and requests Silo overlap; simulation disables both policies. The common implementation applies both policies dimensionally, eliminating the three copies of each formula.

STL instantiation remains a single common routine. Post-process now has the declarations needed to compile that routine, but no new accepted inputs and no new call site. Generated declaration targets are therefore the compile-time ownership mechanism, while namelist targets remain the user-facing input contract.

Phase-change follows the same ownership rule: the complete numerical implementation is compiled once in common for all executables, while only the stages that currently need it invoke it. The two solver tolerances are declaration-visible in post-process solely so the common module compiles; their accepted input scope remains pre-process/simulation.

Behavior changes

This is predominantly a refactor, but three changes are not behavior-preserving and are called out deliberately.

Unifying the mixture-conversion kernel means pre_process and post_process now apply the mpp_lim clamp and volume-fraction renormalization that previously existed only in the simulation path. For mpp_lim = T this can change derived quantities wherever the stored volume fractions are slightly out of bounds or do not sum to one. The full golden-file suite, including post-process comparisons, shows no movement in any covered case.

Unifying the chemistry halo policy fixes a latent buffer overrun on master: buffer allocation used chemistry .and. chem_params%diffusion while the exchange itself used present(q_T_sf) .and. chemistry for pre-process and post-process, so a chemistry run without diffusion packed nVar + 1 variables into buffers sized for sys_size. Both sites now share one predicate.

Consequently pre_process no longer exchanges the ghost temperature layer when chemistry is enabled without diffusion. Nothing consumes it: elliptic smoothing reads only q_prim_vf, and s_compute_T_from_primitives recomputes the temperature over the interior immediately afterwards.

Type of change

• Refactor

Testing

• Built pre_process, simulation, and post_process without MPI.

• Built all three targets with MPI.

• Built a case-optimized simulation using examples/1D_euler_convergence/case.py.

• Ran focused MPI regression cases with post-processing:

  • 8EAC3DA7 (1D viscous)
  • 9DAC4DDC (two-fluid alternative sound speed)
  • CE9DBA3F (QBMM)
  • B9553426 (Lagrangian bubbles)

• Ran an end-to-end non-MPI 1D_euler_convergence example.

• Ran focused grid and boundary regressions with post-processing:

  • B49877F3 (3D grid stretching)
  • 8C7AA13B (2D, two MPI ranks)
  • DB670E50 (axisymmetric boundary)
  • 3008BA80 (periodic immersed-boundary case)

• Ran focused STL/model regressions:

  • EA8FA07E (2D IBM STL)
  • E777BE53 (3D pre-process ICPP STL)

• Ran focused MPI I/O regressions:

  • 356A5C50 (non-polytropic QBMM with parallel I/O)
  • 3008BA80 (immersed-boundary marker I/O)

• Ran a dedicated MPI downsampling smoke case through pre-process, simulation, and post-process; an 11x11x11 input was read as the expected 3x3x3 post-process domain.

• Ran 986BC1A2 (2D Richtmyer--Meshkov with Dirichlet boundary-buffer input) through pre-process, simulation, and post-process with MPI.

• Built a case-optimized simulation for examples/2D_richtmyer_meshkov/case.py.

• Ran FD891191 (two-fluid six-equation model) through pre-process, simulation, and post-process with MPI.

• Ran 1550B67E (2D chemistry with diffusion disabled) through pre-process, simulation, and post-process with MPI.

• Re-ran B9553426 (2D Lagrangian bubbles) through all three stages with MPI to exercise ordinary and beta halo transport.

• Re-ran decomposition regressions through pre-process, simulation, and post-process with MPI:

  • B49877F3 (3D grid stretching)
  • 8C7AA13B (2D, two MPI ranks, Silo output)
  • DB670E50 (axisymmetric)

• Ran a two-rank old-grid smoke from examples/1D_vacuum_restart: generated the original parallel grid, then re-read it with old_grid = T in a temporary case directory. The temporary restart copy supplied domain bounds required by current validation and disabled old_ic so the check exercised grid reuse without requiring a 7000-step restart file.

• Re-ran STL regressions:

  • EA8FA07E (2D IBM STL)
  • E777BE53 (3D pre-process ICPP STL)

• Ran phase-change regressions through pre-process, simulation, and post-process with MPI:

  • E6DBCBD9 (1D, model 5, two fluids, six-equation model)
  • 02BD9B5E (2D, model 6, three fluids, six-equation model)
  • 344FAA70 (3D, model 5, two fluids, six-equation model)

• Ran the parameter generator tests: 93 passed.

• Ran the focused Fortran-generator tests after the declaration changes: 48 passed.

• Ran the repository pre-commit gate: formatting, spelling, toolchain lint, source lint, documentation references, parameter documentation, and all example validations passed.

• Expanded the affected sources for OpenACC and inspected the resulting declarations and data lifecycle. The device conservative-to-primitive path calls the GPU_ROUTINE kernel directly, and the old stage-specific accelerator routine is absent.

• Expanded the grid changes for both OpenACC and OpenMP. The host grid population precedes acc update device(glb_bounds) / omp target update to(glb_bounds), and common grid routines do not take over device-data ownership.

• Expanded m_mpi_common for all three executables with both OpenACC and OpenMP definitions and inspected the new explicit QBMM, immersed-boundary, and downsampled interfaces. A physical accelerator build is unavailable on this macOS/GNU toolchain.

• Expanded the Dirichlet change for all three executables with both OpenACC and OpenMP definitions. The policy receives acc declare / omp declare target and acc update device / omp target update, and both buffer-copy and extrapolation paths remain in the device routine.

• Expanded the common global-parameter module for all three executables with both OpenACC and OpenMP definitions. Pre-process/post-process receive only common manual device declarations; simulation receives exactly one generated declaration for each owned GPU scalar.

• Expanded a case-optimized simulation and verified that num_dims, num_vels, weno_polyn, and muscl_polyn remain Fortran parameters with no OpenACC/OpenMP device allocation.

• Expanded m_mpi_common for all three executables with both OpenACC and OpenMP definitions after transport unification. Every stage receives the same staged and RDMA branches; runtime policy selects one, and the portability macros emit the expected host/device updates or device-address regions.

• Re-expanded m_mpi_common for every executable and both accelerator backends after decomposition unification; each expansion contains the same single policy helper and no stage-selection residue.

• Built a case-optimized MPI simulation for examples/2D_richtmyer_meshkov/case.py after the decomposition interface change.

• Expanded m_model for all three executables with OpenACC and OpenMP definitions. Every expansion contains exactly one STL-instantiation routine, its existing device updates, and no stage-selection residue.

• Expanded m_phase_change for all three executables with OpenACC and OpenMP definitions. Every expansion contains the same full phase-change solver and accelerator loop annotations, with no stage-selection residue.

• Verified that rg -n 'MFC_(PRE_PROCESS|SIMULATION|POST_PROCESS)' src/common returns no matches.

• Merged master (Remove non-operational hyperelasticity capability #1675, hyperelasticity removal) and resolved eight conflicts; m_hyperelastic.fpp is deleted, elasticity becomes hypoelasticity, and b_size/tensor_size/elasticity GPU declarations are dropped.

• Ran the full golden-file suite with post-process comparisons enabled (./mfc.sh test -a) on GNU 13.3 with MPI: 577 tests, no golden-file differences.

• Built all three targets with NVHPC 25.11 and OpenACC, and ran the suite on NVIDIA A100 hardware: 296 cases from the full run plus a 30-case set targeting every device-facing change (IBM STL, immersed-boundary markers, Lagrangian bubbles, non-polytropic QBMM, cylindrical and axisymmetric grids, phase change, mpp_lim, viscous, and the boundary-condition sweep). No failures.

• Ran both chemistry cases on GPU (1550B67E, C4EB58A8) after building the chemistry variant.

• Built a case-optimized simulation with NVHPC and OpenACC, and confirmed the generated declarations emit case-optimization parameters as Fortran parameter constants with no device storage, while the non-case-optimized branch emits exactly the twenty-two GPU_DECLARE names the deleted hand-written block contained.

• syscheck passes with MPI and OpenACC device offload on the same machine.

Cray and Intel are covered only by CI; neither was exercised locally.

Checklist

• I added or updated tests for new behavior
• I updated documentation if user-facing behavior changed

No user-facing parameter behavior changed; the declaration/namelist separation preserves the existing accepted inputs for each executable. The numerical changes are limited to those listed under Behavior changes.

See the developer guide for full coding standards.

AI code reviews

Reviews are not retriggered automatically. To request a review, comment on the PR:

• @claude full review — Claude full review (also triggers on PR open/reopen/ready)
• Or add label claude-full-review — Claude full review via label