results: A100 A/B of the compacted ConstantSplit assembly (PyAutoArray#537, 12x)

[Jammy2211]

Summary

The autolens_profiling half of the DelaunayNN split-regularization assembly
compaction (PyAutoLabs/PyAutoArray#536): the bench that priced every candidate
assembly, the real HST stencil tables it runs on, the A100 A/B submits, and the
results note that records the verdict.

The bench is what chose the design rather than confirming it. On real HST tables
it priced the alternatives — dense GEMM 10.1 ms (a wash with today's scatter),
BCOO 21 ms, dedup/segment-sum 39 ms — and swept the compact width (12: 0.58 ms,
16: 1.47, 20: 2.78, 24: 4.53, 28: 6.71, 32: 9.31, 33 uncompacted: 10.03), which
is where SPLIT_REG_COMPACT_WIDTH = 12 comes from.

A100 A/B (jobs 342334-342338, same node, same window)

measurement	before	after	speedup
assembly row (ms/call)	10.03	0.84	12.0x
params -> H prefix, vmap 16 (ms/call)	16.423	7.260	2.26x
params -> H prefix, unbatched (ms)	24.341	15.189	1.60x
full likelihood, single JIT (ms)	75.6	66.1	1.14x
full likelihood, vmap 16 (ms/call)	50.04	40.86	1.22x

Pinned likelihood 29144.581944 bit-identical on both legs, pinned_drift: [].
The pre-registered witness (assembly < 3 ms, params -> H < 11 ms) is met.

Scripts Changed

• scripts/misc/delaunay_nn/assembly_bench.py — new: V0..V6 assembly variants
  (uncompacted scatter, compact scatter + wide-row supplement, dense GEMM, BCOO,
  dedup/segment-sum), compact-width sweep, run against real tables.
• scripts/misc/delaunay_nn/build_tables.py — new: rebuilds the frozen stencil
  tables the bench consumes.
• scripts/misc/delaunay_nn/extract_mesh_points.py — new: extracts the HST mesh
  points the tables are built from.
• scripts/misc/delaunay_nn/README.md — new: how to rebuild and re-run.
• hpc/batch_gpu/submit_delaunay_nn_assembly_bench_a100 and the four A/B
  breakdown/likelihood submits — new A100 submits (assembly + control legs).
• results/delaunay_nn/{tables_hst_1500.npz, mesh_points_hst.npy, assembly_bench_a100_fp64.json} — frozen inputs and bench results.
• results/{breakdown,jit}/...delaunay_nn_hpc_a100_fp64_assembly{,_control}.{json,png}
  — the A/B results.
• results/notes/delaunay_nn_constant_split_assembly.md — the results note:
  design, the priced alternatives, the width sweep, the A/B verdict, the cap
  audit that sizes the wide-row budget.

Results and notes only apart from the new bench scripts; no existing profiling
script or result is modified.

Upstream PR

PyAutoLabs/PyAutoArray#537

Library-first gate: merge after PyAutoLabs/PyAutoArray#537. The bench's V1
variant mirrors the compacted library path, and the note's verdict describes
behaviour that only exists once the library PR lands.

Test Plan

• ruff check scripts/misc/delaunay_nn/ — all checks passed.
• ruff format --check scripts/misc/delaunay_nn/ — 4 files already
  formatted.
• assembly_bench.py executed on the A100 (job 342338) and on CPU; results
  committed under results/delaunay_nn/.
• A/B breakdown + likelihood submits ran on the same A100 node and window
  (jobs 342334-342338), pinned likelihood bit-identical.

Links

• Results note: results/notes/delaunay_nn_constant_split_assembly.md
• Refs perf: compact the ConstantSplit split-stencil scatter (DelaunayNN) PyAutoArray#536

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_011xsh8KqgiHWTfPGgWEYMQw