perf(cuda): skip the far LOD pass when the far band is empty

[bam93]

What

API_computeSES_lod runs computeSlicedSES for both the near and far bands unconditionally. When the molecule lies entirely in the near band, the far pass still pays a full pass of setup (grid build + atom thrust::sort) + the slab-enumeration loop + aabbInFrustum tests, only to produce 0 geometry. The per-slab visible-only cull (mode==0) already skips the heavy refine/MC kernels, but not this setup/enumeration cost.

This adds a pass-level pre-check: test the model's padded world AABB (getMinMax, expanded by maxAtomRad + 2*probe to match the grid extent, so the test is conservative and never skips a band that could contribute) against farPlanes using the existing aabbInFrustum. The far computeSlicedSES is skipped when the far band is provably empty.

Correctness — output-identical (verified byte-exact)

The skip only removes a pass that produced nothing (when the far AABB is outside the far frustum, the near pass already covered every visible slab). Verified on 1CRN and 3EAM, both the empty-far and the real-split cases: NVert/NTri are identical with vs without the skip.

case	structure	NVert (no-skip)	NVert (skip)
empty far band	3EAM	391977	391977
real split	3EAM	485153	485153

Performance (dgx, RTX A1000, reso 0.5, coarseMul 2)

case	structure	no-skip	with-skip	delta
empty far band	3EAM (12625 atoms)	129 ms	115 ms	~11%
empty far band	1CRN (327 atoms)	4.0 ms	3.6 ms	~noise
real split	3EAM	146 ms	150 ms	unchanged (skip correctly does not fire)

A modest win, larger on big multi-slab systems; the skip never fires when the far band has real work. Tested via a small harness calling API_computeSES_lod directly (the CLI doesn't exercise the LOD API); harness not included in the PR.

Provenance

This mirrors the same empty-far-pass skip applied in the UnityMol Unity-ComputeShader port (where it was ~2×, since that path re-ran the whole GPU dispatch; CUDA's in-loop cull already makes the empty pass cheaper, hence the smaller win here).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR optimizes the CUDA LOD SES pipeline by avoiding the far-band computeSlicedSES pass when it can be proven to generate no geometry, reducing unnecessary setup and slab-enumeration overhead for near-only cases.

Changes:

• Added a pass-level frustum pre-check using a conservatively padded model AABB against farPlanes.
• Skips the far computeSlicedSES call entirely when the padded AABB is outside the far frustum (i.e., far band is empty).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[bam93]

Good catch — fixed in a7ad2e4. The comment now states the relationship accurately: the per-side pad maxAtomRad + 2*probeRadius is >= the grid's actual reach (the SES grid origin offsets each side by maxAtomRad + probeRadius via originGridNeighbor, and computeMaxDist adds 2*maxAtomRad + 4*probeRadius along the longest axis), i.e. a deliberate over-cover that never under-covers — so the skip can never drop a band that could contribute. Comment-only change; the pad value and logic are unchanged, so the verified parity/perf results still hold.

[bam93]

Added a small related parity tidy-up (8eadb93): computeClosestAtom now uses squared distance for the per-vertex nearest-atom argmin (sqrt is monotonic, so the argmin is identical without the per-candidate sqrt) — matching the Metal port which already does this. Byte-identical verified: md5-identical OBJ on 1CRN/1UBQ/3EAM (baseline build vs change, same -v 0.5). Correctness-neutral; brings the two backends into agreement. Happy to split it into its own PR if you'd prefer.