SCC26: Read WENO stencils directly from v_vf instead of an identical copy - #2092

Closes #2091.

Summary

For WENO3/5/7, s_pack_weno_input_arr copies v_vf into v_rs_weno with the same (j, k, l) layout before each reconstruction. This PR makes the WENO3/5/7 reconstruction kernels read v_vf(i)%sf directly. v_rs_weno is kept only for MP-WENO and THINC, which still take it as an argument. Background and the trade-off analysis are in #2091.

On one H200, grind time drops 5.9–6.5% on three example cases with NVHPC 26.3 and 25.9. Outputs are bit-identical to master on our tested cases, and the full GPU and CPU test suites pass.

This PR is from Team NTHU SCC, participating in the SC26 Student Cluster Competition.

Changes

Only src/simulation/m_weno.fpp changes:

• In the WENO3/5/7 reconstruction loops, v_rs_weno(${SF(...)}$, i) becomes v_vf(i)%sf(${SF(...)}$). The arithmetic and the loop structure are unchanged.
• v_rs_weno is allocated, packed and deallocated only when mp_weno .or. int_comp > 0, for s_preserve_monotonicity and s_thinc_compression.
• WENO1 (which already skipped the copy) and MUSCL (which never calls s_weno) are unchanged.

Performance

Grind time (ns/gp/eq/rhs), master → this PR. Each entry is the mean of three interleaved runs within one allocation. All runs use H200, double precision, --case-optimization and the default build flags. Two-GPU runs use two MPI ranks on one node.

Maximum run-to-run spread was 0.81% (NVHPC 26.3, 1 GPU), 0.61% (NVHPC 25.9, 1 GPU), and 2.21% for the two reported 2-GPU cases.

*The 2-GPU 3D_performance_test is omitted: on our node its run-to-run spread (up to 7.8%) was comparable to the effect.

Where the time goes (NVHPC 26.3, Nsight Systems and Nsight Compute):

• The copy accounted for 7.7% of GPU kernel time in both 3D_performance_test and 2D_shockdroplet. Removing it reduces total kernel time by 6.1% and 6.5%, respectively.
• WENO5 reconstruction kernels become 5–6% slower due to additional address computation and descriptor accesses, but the eliminated packing kernels more than offset this overhead. L2/DRAM traffic remains within about 1% of master. See Avoid redundant WENO input packing in the common reconstruction path #2091 for details.

Trade-off for MP-WENO / THINC

With mp_weno or int_comp > 0 the copy is still needed, so the small reconstruction overhead is not offset. On 2D_richtmyer_meshkov (WENO5 + MP-WENO, 431 steps), total kernel time is +0.8%, and grind time is +0.45%, within run-to-run variation (up to 2.3%). Avoiding this overhead would require separate reconstruction kernels selected at runtime. We chose not to introduce this duplication for the small observed regression.

Unchanged paths behave as expected: total kernel time is +0.06% for 2D_orszag_tang (WENO1) and +0.03% for 2D_shockdroplet_muscl (MUSCL + THINC), both with bit-identical outputs.

Testing

Baseline: master at 42c44fd.

• GPU full suite (NVHPC 26.3, OpenACC, H200): 725 passed, 0 failed, 39 skipped. This includes the three mp_weno=T tests and the five WENO5 int_comp tests, which exercise the remaining copy path.
• CPU full suite (NVHPC 26.3, Intel Xeon Platinum 8480+): 722 passed, 0 failed, 39 skipped.
• Bit-identical outputs: in our runs, the final-step restart files of every case above match master, including case-optimized builds. No golden files changed.

CPU builds also use the new reads. CPU performance was not benchmarked.

Contribution Policy

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.