Raise the amdflang species bound so mechanisms above ten species build

[sbryngelson]

Closes #1848.

When case optimization is off, the USING_AMD guard substitutes a literal extent for dimension(num_species). That literal was ten, across 21 declarations in nine files, with m_checker_common.fpp stating the matching constraint:

@:PROHIBIT(chemistry .and. num_species > 10, "num_species > 10 for AMDFLang when Case optimization is off")

So a mechanism above ten species cannot be built that way, and the check meant to explain it never runs. m_thermochem.f90 is generated per mechanism with concrete bounds, so an eleven-species mechanism is a compile error long before the runtime check, and what the reader sees is

error: Actual argument array has fewer elements (10) than dummy

repeated across m_chemistry.fpp, m_cbc.fpp, m_riemann_state.fpp, m_ibm.fpp and the three Riemann solvers, rather than the sentence describing the limit. Hit while reviewing #1821, whose carbon surface mechanism has eleven species.

This raises the bound to twenty and updates the check to match, so the stated limit and the real one agree.

The reductions had to move with it

Raising the literal alone is wrong, and the first commit here was. These arrays are kernel-private and only their first num_species elements are ever written, so a whole-array reduction over one of them is correct only while the literal equals num_species — which is what ten happened to be for the mechanisms in the test suite. At twenty, sum(Xs(:)/(Gamma_i(:) - 1.0_wp)) sums ten elements of uninitialised device scratch. Zero-filled scratch gives the right answer; anything else gives a quietly wrong gamma, or a divide by zero if Cp_i lands on one.

The second commit pins every such reduction to 1:num_species:

• m_cbc.fpp — Gamma_i and the gamma sum on the gamma_method = 1 path
• m_riemann_state.fpp — h_avg_2, Yi_avg, the four Cp_avg/Cv_avg sums, Phi_avg and c_sum_Yi_Phi in the reacting Roe average

That removes the hidden dependency rather than moving it, and is the right form regardless of what the literal is. The three Riemann solvers' gamma_L/gamma_R sums were already sliced. Everything else touching these arrays indexes elementwise over 1, num_species, and the generated m_thermochem dummies are explicit-shape at the mechanism's own count, so a longer actual argument never reaches the padding.

Cost

Measured on C4EB58A8 (2D reacting mixing layer), generic chemistry GPU build, amdflang OpenMP offload on MI210. Numbers are the 5-equation pure-fluid HLLC kernel, which carries eleven of these arrays:

bound	VGPR	scratch/thread	avg time/step
ten	260	396 B	1.37e-1 s
twenty	326	1372 B	1.40e-1 s

About two percent. Occupancy does not change: waves per SIMD is floor(512 / roundup(vgpr, 8)), and 260 and 326 both give one wave, so the extra pressure costs spill traffic rather than a wave. One run each, so read the timing as "no large penalty" rather than an exact figure. A case where HLLC dominates would show more than this one does.

Note

rhoYks in m_variables_conversion.fpp spelled the same bound as (1:10) and is included here.

Raising the literal also surfaced three latent shape mismatches in #1821's new code, where surrounding code assumed the padded array was exactly num_species long. The bound is not only a ceiling; code near it depends on the padding happening to match. That is worth keeping in mind for the alternative in #1848 of retiring these guards once the amdgpu bug they work around is fixed.

Testing

All sixteen Chemistry tests pass on amdflang OpenMP offload on MI210, the configuration the guard applies to. That set covers the paths the slicing fix touches: both Reacting Roe Average cases, both Riemann Solver 1/2 -> Gamma pairs, and the two isothermal-wall CBC cases. Note that passing tests are not evidence the unsliced version was broken — zero-filled scratch reads as a correct answer, which is the whole hazard. Case-optimized builds are unaffected either way: the guard is not MFC_CASE_OPTIMIZATION and USING_AMD, so those already take dimension(num_species).

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Pull request overview

Raises the hardcoded amdflang (non-case-optimized) species array extent so chemistry mechanisms with >10 species can compile, and aligns the documented num_species limit check with the new bound.

Changes:

• Increased multiple USING_AMD/not MFC_CASE_OPTIMIZATION array extents from 10 to 20 across chemistry/CBC/IBM and Riemann modules.
• Updated the amdflang configuration checker to prohibit num_species > 20 instead of > 10.
• Normalized one previously hardcoded (1:10) extent to (1:20) in variables conversion.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/simulation/m_riemann_state.fpp	Bumps amdflang non-case-optimized species array extents to 20 in Riemann state helper.
src/simulation/m_riemann_solver_lf.fpp	Updates Lax–Friedrichs solver local species arrays to the new padded bound.
src/simulation/m_riemann_solver_hllc.fpp	Updates HLLC solver local thermochemistry/species arrays to the new padded bound.
src/simulation/m_riemann_solver_hll.fpp	Updates HLL solver local species arrays to the new padded bound.
src/simulation/m_ibm.fpp	Updates IBM path species array extent for amdflang non-case-optimized builds.
src/simulation/m_compute_cbc.fpp	Updates CBC helper interfaces taking dYs_ds to use extent 20 under the amdflang guard.
src/simulation/m_cbc.fpp	Updates CBC routine local species/thermo arrays to the new padded bound.
src/common/m_variables_conversion.fpp	Updates hardcoded species temporary arrays to extent 20 under the amdflang guard.
src/common/m_chemistry.fpp	Updates chemistry kernels’ local species arrays to the new padded bound.
src/common/m_checker_common.fpp	Aligns the enforced species limit message/constraint with the new bound (20).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

❌ Patch coverage is 72.72727% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 61.26%. Comparing base (dc0aec1) to head (e669ad0).

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

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1852      +/-   ##
==========================================
- Coverage   61.26%   61.26%   -0.01%     
==========================================
  Files          84       84              
  Lines       22330    22331       +1     
  Branches     3265     3265              
==========================================
  Hits        13680    13680              
- Misses       6207     6208       +1     
  Partials     2443     2443              

☔ 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/include/shared_parallel_macros.fpp	162	+6
src/common/m_checker_common.fpp	39	+3
src/simulation/m_riemann_state.fpp	1193	+2
src/simulation/m_riemann_solver_hllc.fpp	1319	+1

Directory	Lines	Diff
common	10362	+9
simulation	28089	+3
total	46352	+12