Clean correct state subroutine

[danieljvickers]

Since we are planning to extend the immersed boudnary method to include more features and EOS, I have moved out some of the subroutines involved in this calculation to clean up the s_ibm_correct_state subroutine. Mostly adding abstractions and making it clear where modifications for various EOS need to go.

I have a few branches open right now as I am still waiting for pull requests to go through. This will heavily conflict with my current debug branch that is still uncovering new bugs. When that branch eventually gets merged, I will have to return and do similar abstraction for getting temperature and density and setting energy.

This is not a draft, but a full PR ready for merge. I am leaving it as a draft for now by request of @sbryngelson. I will let him determine when to open a PR

[github-actions]

Claude Code Review

Head SHA: 9ae77c5

Files changed:

• 1
• src/simulation/m_ibm.fpp

Findings:

• src/simulation/m_ibm.fpp (new s_compute_ghost_point_pressure, ~line 166-172): the case default branch for state-dependent EOS (Mie-Gruneisen/JWL/Vinet) only calls s_mpi_abort under #ifndef MFC_GPU. On a GPU build, hitting this branch (moving IB + a fluid using a non-ideal/non-stiffened-gas EOS) silently skips the pressure contribution for that fluid instead of erroring — pres_GP stays at its 0._wp initializer contribution for that species, and if every fluid uses a state-dependent EOS the ghost-point pressure is silently set to 0. Since src/simulation/ is the only GPU-accelerated target and this guard exists specifically to prevent an unimplemented/incorrect physics path from running, the CPU-only abort leaves GPU runs to silently produce wrong energy/pressure fields for this combination rather than failing loudly, which is the exact "silent wrong answer" failure mode this codebase treats as most severe.

[danieljvickers]

@sbryngelson This is read for PR

[sbryngelson]

Reviewed 5c48618. The velocity extraction is a faithful move - s_compute_ghost_point_velocity reproduces the slip/no-slip branches, the rotation superposition and the v_blow block statement for statement, in the same order. The other two cleanups are behaviour-preserving too: physical_loc = [x_cc, y_cc, 0]; if (num_dims == 3) ... matches the old if (p > 0) form, hoisting radial_vector is required for it to be defined at the call and it is only read under moving_ibm /= 0, and accumulating the pressure into a real(wp) local instead of into q_prim_vf(E)%sf removes a per-iteration round trip through stp storage - identical by default and under --single, strictly better under --mixed.

One thing is not behaviour-preserving.

The new select case (eoss(q)) silently zeroes the moving-IB ghost pressure on GPU

src/simulation/m_ibm.fpp:160-172. The old body applied the correction unconditionally for q = 1, num_fluids. The new one wraps it in select case (eoss(q)) with case (eos_ideal_gas, eos_stiffened_gas), and a case default whose entire content is:

#ifndef MFC_GPU
    call s_mpi_abort('s_compute_ghost_point_pressure: moving IB pressure correction is only ' &
                     & // 'implemented for eos_ideal_gas/eos_stiffened_gas')
#endif

So with moving_ibm >= 2 and any fluid whose eoss(q) is eos_mie_gruneisen, eos_jwl or eos_vinet:

• CPU: hard abort mid-solve.
• GPU: the case default is empty, pres_GP keeps its 0._wp initialisation for that fluid, and q_prim_vf(eqn_idx%E)%sf(j,k,l) is set to zero (or a partial sum) at every ghost point. No warning, no NaN - a zero-pressure IB surface.

That is a CPU/GPU divergence producing a silent wrong answer on the GPU path, which CLAUDE.md ranks first among review priorities. There is no input-time guard: nothing in src/*/m_checker*.fpp or case_validator.py ties moving_ibm to eos.

Per the project's own rule that validation belongs in case_validator.py and the m_checker* files rather than scattered through the solver, the restriction wants to be a startup check rejecting moving_ibm > 1 together with a state-dependent EOS, leaving the kernel unconditional as it was. As written the abort is unreachable in exactly the configuration where it matters.

Minor

• Dead locals. vel_norm_IP, v_blow_eff, norm, buf and rotation_velocity (declared around m_ibm.fpp:252-281) are no longer referenced in s_ibm_correct_state after the extraction, but remain declared and remain in the GPU_PARALLEL_LOOP private='[...]' clause. Harmless, but shrinking that routine was the point, and it keeps the private list honest - see the c_sum_Yi_Phi entry in the pitfalls file about private lists drifting from kernel bodies.
• The two new routines are not in the module's private :: list (m_ibm.fpp:28-30). There is no bare private statement, so they default to public while every other internal helper here is explicitly listed.
• s_compute_ghost_point_velocity adds a nesting level - a seq GPU_ROUTINE calling s_cross_product, where the call site used to be directly in the parallel-loop body. The argument is a fixed dimension(1:3), not a device-global bound, so ftn-7066 does not apply; still worth watching the Frontier CCE-ACC lane, since this module is where that class has bitten before.

CI is green (58 pass, 0 fail).

[danieljvickers]

Addressed review comments by simply removing the select statement. But the sentiment of just prohibiting the case is correct, and I have added a prohibit to the python checker to compensate. Uneeded variables were removed, as is correct.

[sbryngelson]

Triaging CI across the open PRs: all 21 failures on this branch are the model_eqns=3 set that every branch cut after Sept 8 inherits from master, not anything this PR does.

Root cause is in #1878: NVHPC's two-pass IPO miscompiles the six-equation pressure-relaxation kernel in OpenACC device code and silently drops the internal-energy update. It emits no diagnostic, so the build succeeds and the answer is wrong. That is why a branch can fail 23 model_eqns=3 tests without touching any of that code.

Nothing to do here. Once #1878 lands, rebasing should clear all 21 without a change on your side.

[github-actions]

Lines of Code

File	Lines	Diff
src/simulation/m_ibm.fpp	1267	+14

Directory	Lines	Diff
simulation	27918	+14
total	46204	+14