Generate caller-shaped thermochemistry routines

[sbryngelson]

Summary

Follow-up to #1915. MFC's hot loops chain several generated calls per gas state, and each re-evaluates the same NASA7 fits or composition. The generator now also emits routines shaped like those call sites, in the spirit of #1915's fused creation/destruction routine:

routine	returns	replaces per state
get_mixture_caloric_state(T, Y, cp0_r, cp, cv, e)	species cp/R, mixture cp, cv, internal energy	get_species_specific_heats_r + ..._cp_mass + ..._cv_mass + get_mixture_energy_mass (4 NASA7 passes -> 2)
get_mixavg_transport_state(p, T, Y, MW, X, D, lambda)	molecular weight, mole fractions, mixture-averaged diffusivities, conductivity	get_mixture_molecular_weight + get_mole_fractions + get_species_mass_diffusivities_mixavg + get_mixture_thermal_conductivity_mixavg (composition computed once, not three times)
get_species_enthalpies_mass(T, h)	species enthalpies [J/kg]	get_species_enthalpies_rt + the caller's *gas_constant*T/molecular_weights loop

Callers:

• HLL, HLLC, LF, CBC: one get_mixture_caloric_state call per state. With gamma_method = 2 that halves the NASA7 evaluations; with the default gamma_method = 1 it is the same work in fewer calls. Unused m_thermochem imports are dropped.
• Mixture-averaged diffusion flux (transport_model = 1): one get_mixavg_transport_state and one get_species_enthalpies_mass per side, and the hand-written enthalpy scaling loop goes away.

Design constraints:

• Call depth. CCE has faulted when a thermochem call sits one routine deeper (see the comment in m_riemann_state.fpp), so the calls stay in the loop bodies and the shared diffusivity/conductivity bodies are Mako %defs expanded into both the existing and the fused routine, not a new callee. No device routine is nested deeper than before, and the existing routines generate identically to master.
• Arithmetic. The fused routines reuse the separate routines' operations.

Testing

• New test_fused_routines (h2o2, gri30): the fused routines are bitwise equal to the separate calls (gfortran -O0) and agree with Cantera.
• amdflang 23.2.1, OpenMP offload, MI250X (OMP_TARGET_OFFLOAD=MANDATORY): --only Chemistry - 16 passed. (One test first failed because a concurrent build in the same checkout replaced its syscheck binary; it passed on re-run.)
• gfortran 12.2 CPU: --only Chemistry - 16 passed, 0 failed.
• ./mfc.sh precheck passes.

Optimized builds do not keep bitwise equality: on the MI250X benchmark below, 13 of 17 fields differ from master at step 500, by at most 8e-11 relative (momentum; most fields 1e-15 to 1e-13), with an identical initial condition. That is inlining/FMA contraction differing between one fused routine and separate calls. All goldens pass at their existing tolerances.

Performance

MI250X (one GCD), amdflang 23.2.1, generic --gpu mp --no-mpi, examples/2D_premixed_flame_vortex (h2o2, HLLC, gamma_method = 1, mixture-averaged diffusion), 500 steps, interleaved:

	master fef3c276	this PR
s/step	0.05237, 0.05310, 0.05234	0.05224, 0.05048, 0.05215
mean	0.05260	0.05162 (-1.9%)

One run drives most of the difference, so read this as neutral to slightly faster for this case. gamma_method = 2 cases should gain more, since they save half the NASA7 passes; I did not benchmark one.

Updates after the first CI run

• Frontier CCE (OpenACC and OpenMP) faulted in the HLLC kernel: get_mixture_caloric_state passed the species cp array on to get_species_specific_heats_r from inside the kernel. It is now a leaf routine (NASA7 expressions and averaging expanded in place, no calls), so HLLC's call is shallower than master's get_mixture_energy_mass. The bitwise test still passes. (6c93dec3)
• NVHPC 23.11 to 24.7 (OpenACC) crashed in the front end (fort1, signal 11) compiling m_compute_levelset, which this PR does not touch. I reproduced it locally with nvfortran 24.1 and bisected it to m_thermochem.mod alone: nearly every file reads it through m_global_parameters, and adding the three routines trips a content-dependent front-end bug (removing any one avoids it; gri30's far larger module does not trip it). The three routines now live in a companion module, m_thermochem_state, generated into the same file and imported only by their callers. m_thermochem is byte-identical to master's for h2o2 and gri30. (3a47e7ec)

Re-checked after both: test_thermochem 17 passed; nvfortran 24.1 --gpu acc --mpi Chemistry build (previously crashed) completes; gfortran CPU and amdflang OpenMP offload on MI210 --only Chemistry 16 passed each.

Follow-up

#1821's surface residual (m_ibm.fpp) calls get_mixture_molecular_weight, get_species_mass_diffusivities_mixavg and the surface rates separately and recomputes mole fractions. Once both land, it can use get_mixavg_transport_state too.

Prepared with Claude Code.

Contribution Policy

We do not accept pull requests generated primarily by AI without genuine understanding or real-world usage context.

All contributions are expected to demonstrate:

• A clear understanding of the codebase
• Alignment with product direction
• Thoughtful reasoning behind changes
• Evidence of real-world usage or hands-on experience with the problem

If these expectations are not met, we would prefer to implement the changes ourselves rather than spend time reviewing low-effort submissions.

---

Acknowledgement

• I confirm this PR meets the above expectations and reflects my own understanding and real-world context.

---

PR template credit: junegunn

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

❌ Patch coverage is 70.00000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 61.51%. Comparing base (fef3c27) to head (3a47e7e).

Files with missing lines	Patch %	Lines
src/simulation/m_riemann_solver_lf.fpp	0.00%	2 Missing ⚠️
src/simulation/m_cbc.fpp	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1919      +/-   ##
==========================================
+ Coverage   61.48%   61.51%   +0.02%     
==========================================
  Files          86       86              
  Lines       22485    22458      -27     
  Branches     3298     3303       +5     
==========================================
- Hits        13826    13816      -10     
+ Misses       6206     6189      -17     
  Partials     2453     2453              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Lines of Code

File	Lines	Diff
src/common/m_chemistry.fpp	411	-9
src/simulation/m_riemann_solver_hll.fpp	609	-7
src/simulation/m_riemann_solver_hllc.fpp	1313	-7
src/simulation/m_riemann_solver_lf.fpp	511	-6
src/simulation/m_cbc.fpp	1115	-3

Directory	Lines	Diff
common	10412	-9
simulation	27825	-23
total	46648	-32