Fix two GPU kernels that share a scalar every thread writes

[sbryngelson]

Two GPU kernels write a scalar that is not in their private clause, so every thread shares one copy. Under OpenMP offload that is a data race; under OpenACC it is hidden, because the spec predetermines scalars in a compute region as private. The symptom is a run that gives a different answer every time on Cray OpenMP while every other backend looks fine.

The two sites

m_riemann_solver_hllc.fpp emits its kernel twice through #:for HYPO in [True, False], once specialised for hypoelasticity and once for pure fluids, so that the pure-fluid copy is not pinned at the register ceiling by variables it never uses. Each copy carried its own hand-written list of about eighty private names, and the two drifted: c_sum_Yi_Phi, written by the chemistry block that is common to both, is private in the hypoelastic list and missing from the other.

Measured on Frontier, three runs of 1D -> Chemistry -> Inert Shocktube -> Reacting Roe Average -> HLLC on one binary:

0.07593517165078
0.07593517172084
0.07593517170526

m_riemann_solver_lf.fpp has the same defect independently: Gamm_L and Gamm_R are assigned inside the kernel under chem_params%gamma_method == 2 and are absent from its private list.

The fix

The Lax-Friedrichs list gains the two names. The HLLC lists are no longer written twice: one list holds the shared names, another the elastic-only ones, and the two emissions are composed from them, so adding a variable to the shared body cannot leave one copy behind.

Verified against a pristine build of the same file: three of the four generated kernels are byte-identical, and the pure-fluid one gains exactly c_sum_Yi_Phi plus the collapse indices and an unused iterator that were already private in the sibling list.

How to find this class

Reading the source shows one loop body, so a missing name looks handled. It is only visible in the generated Fortran, where there are two directives and one of them is short. The check is to take each !$omp target directive and confirm every scalar assigned in its body appears in the clause.

Testing

Chemistry set passes, 15 of 15. Format and precheck clean. No golden files move: on a serial CPU build a private clause is a no-op, which is also why this never showed up outside GPU offload.

[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

Fixes GPU offload data races caused by missing private scalars in two Riemann-solver kernels, making OpenMP offload behavior consistent with other backends.

Changes:

• Add missing Gamm_L/Gamm_R to the Lax-Friedrichs kernel private list.
• Refactor HLLC private-variable lists into shared vs hypoelastic-only fragments to prevent drift between the two generated kernels.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/simulation/m_riemann_solver_lf.fpp	Adds missing chemistry-only scalars to the OpenMP private clause to avoid per-team sharing/races.
src/simulation/m_riemann_solver_hllc.fpp	Deduplicates the HLLC private list definition by composing shared + elastic-only fragments to prevent future omissions.

---

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

[github-actions]

Lines of Code

File	Lines	Diff
src/simulation/m_riemann_solver_hllc.fpp	1291	+23

Directory	Lines	Diff
simulation	27691	+23
total	45461	+23

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 62.26%. Comparing base (8010057) to head (fc8c570).
⚠️ Report is 5 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1823   +/-   ##
=======================================
  Coverage   62.26%   62.26%           
=======================================
  Files          84       84           
  Lines       21558    21558           
  Branches     3188     3195    +7     
=======================================
  Hits        13423    13423           
  Misses       5937     5937           
  Partials     2198     2198           

☔ 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.