Fix silent viscosity loss on AMD flang GPU (host-capture Re_size in the Riemann solvers)

[sbryngelson]

Closes #1587. (The branch name is stale — this no longer makes Re_size allocatable; see below.)

Summary

Fixes silently-wrong viscous results on the Frontier AMD-flang OpenMP-offload GPU build — the 2D_viscous_shock_tube regression introduced by #1556. 3 files, solver-only; Re_size stays static.

Root cause

On AMD flang OpenMP offload, a cross-translation-unit read of the static declare target Re_size inside a Riemann-solver kernel is unreliable and codegen-dependent: some solver kernels read it as 0, so if (Re_size(i) > 0) is false, Re_L/Re_R collapse to dflt_real, the interface Reynolds number becomes 1e16, and viscosity is silently switched off. CPU and NVIDIA (nvfortran) builds are unaffected — verified across all 14 NVHPC CI lanes.

Why not "make Re_size allocatable" (the previous version of this PR)

That approach only moved the bug. On AMD flang it fixed HLL/HLLC but broke the LF (riemann_solver=5) and IBM viscous paths — they pass on master and failed with the allocatable change (an inversion: static breaks some solver kernels, allocatable breaks others). The defect is per-kernel codegen, not a clean static-vs-allocatable rule, so no choice of storage class is reliable for every kernel. (Confirmed on hardware: per-solver GPU_UPDATE workarounds also only fixed some kernels.)

The fix

Keep Re_size static; in each of s_hll/hllc/lf_riemann_solver, capture it into a host-set local Re_size_loc and firstprivate that value into the viscous kernels, so they never read Re_size from declare-target device memory:

integer, dimension(2) :: Re_size_loc   ! host copy
...
Re_size_loc = Re_size
$:GPU_PARALLEL_LOOP(collapse=3, firstprivate='[Re_size_loc]', private='[...]')
...
do q = 1, Re_size_loc(i)   ! was Re_size(i)

Removing the declare-target read removes the variable the compiler bug acts on, so it's immune regardless of which kernel or solver.

Verification (Frontier AMD flang, gpu-omp)

All five previously-affected viscous tests pass with this fix:
70EC99CE viscous_shock_tube, 8EAC3DA7 1D viscous canary, 27B24CD9 + 4C812637 LF-viscous (the two the allocatable approach broke), B0CE19C5 IBM-viscous. format + precheck clean.

Upstream

This is a workaround for an amdflang bug, filed with a standalone minimal reproducer: ROCm/llvm-project#2890, llvm/llvm-project#203711 (still reproduces on the newest amdflang, AFAR 23.2.0 / LLVM 23). The always-run viscous-GPU canary in #1590 is what surfaced the inversion before merge.

[Copilot]

Pull request overview

This PR addresses an AMD flang OpenMP-offload GPU correctness regression (post-#1556 module split) where Re_size could remain zero on-device in solver translation units, effectively disabling viscous terms without an error. The fix makes Re_size allocatable so it is handled through the OpenMP runtime mapping machinery (consistent with Re_idx / Res_gs) and thus behaves correctly across translation units.

Changes:

• Change Re_size from a static integer, dimension(2) to integer, allocatable, dimension(:).
• Allocate Re_size(1:2) during global-parameters initialization before first use and keep initialization to zero.
• Deallocate Re_size in the module finalization path.

[codecov]

Codecov Report

❌ Patch coverage is 42.85714% with 12 lines in your changes missing coverage. Please review.
✅ Project coverage is 60.94%. Comparing base (ac5774e) to head (da4b6e3).
⚠️ Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/simulation/m_riemann_solver_hllc.fpp	50.00%	3 Missing and 2 partials ⚠️
src/simulation/m_riemann_solver_lf.fpp	42.85%	0 Missing and 4 partials ⚠️
src/simulation/m_riemann_solver_hll.fpp	25.00%	3 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1588   +/-   ##
=======================================
  Coverage   60.94%   60.94%           
=======================================
  Files          82       82           
  Lines       19922    19925    +3     
  Branches     2924     2924           
=======================================
+ Hits        12141    12144    +3     
  Misses       5805     5805           
  Partials     1976     1976           

☔ 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.