ICCAD 2026 FloorSet Challenge — Problem C

Deterministic heuristic optimizer for data-driven SoC floorplanning (21–120 blocks).

📊 Presentation / overview: docs/SUMMARY.md — problem, solution, results, methodology, key insights, and honest assessment (each section ≈ one slide). Start here.

📦 Submission: SUBMISSION_PLAN.md — organizer requirements (PyInstaller executable via op_wrapper.py), the verified torch-free package (packaging/, build with packaging/build_submission.sh), and the rebuild gate.

🤝 Handoff: HANDOFF.md — complete state, workflows, and ranked open leads for whoever continues the work.

🏆 Winning strategy: docs/WINNING_PLAN.md — authoritative beta-to-final research plan, promotion gates, architecture, and execution calendar. This supersedes narrower tactical conclusions.

Current Result (submission candidate)

Current best = v32: v31 plus a zero-solve-cost final gate that reuses the already-computed first pass of the heavy strong-pin candidate. The rest of the solver remains the exact-area dissection portfolio, constraint-preserving polish, and feasibility-gated selector developed through v31 — contest_solution/my_optimizer.py + contest_solution/dissect.py + contest_solution/topology_polish.py; result snapshot results/integrated_v32.json. v32 exposes an intermediate layout that the incumbent previously discarded and compares it only after all primary selection and boundary repairs. It adds no dissection pass. Previous locked entry: sprint5_v9 (2.7182, results/v9_locked.json) — still the per-case fallback inside solve().

Metric	Value
Validation score (RF=1.0)	1.615379 (v31: 1.616638; v9: 2.7182)
Feasible cases	100 / 100
Average runtime	0.196s/case paired (v31 control: 0.196s)
Runtime-adjusted total @ median 1s	1.1951 (paired v31: 1.1967)

Note on the score: the contest cost is runtime-adjusted — cost · max(0.7, (rt/median)^0.3) with a hard 0.7× floor for being ≥~3× faster than the field median. The current solver has 40% better raw quality than v9. v32 improves all five clean and all five raw source-disjoint heavy folds: pooled clean 1.778134 → 1.767565 and raw 1.848364 → 1.834588, with 1,050/1,050 feasible. Its paired runtime-adjusted score improves at every tested field median from 0.25s through 3s. Runtime discipline is a standing rule: changes are judged runtime-adjusted at median ∈ {0.25,0.5,1,2,3}s (HANDOFF.md §5.3), never on the raw score alone.

Validation-set results only. Final ranking uses hidden test data.

Frozen pre-beta release: v32-prebeta-20260711 (tagged source plus the verified AMD64 submission archive).

Repository Structure

contest_solution/
├── my_optimizer.py          # THE solver: shelf + dissection portfolio + selector
├── dissect.py               # Exact-area dissection engine (CAMPAIGN_GOLDEN)
├── topology_polish.py       # Fixed-topology weighted-median HPWL polish (n≤90)
├── learned_order.py         # Input-only, permutation-equivariant order features
├── sequence_pair_sa.py      # Dormant SP-SA floorplanner (historical)
├── iccad2026_evaluate.py    # Adapted local convenience copy (not the official source of truth)
└── *dataset*.py             # Dataset loaders + tests
packaging/                   # Submission executable sources + build script + organizers' op_wrapper
scripts/                     # Maintained and legacy tools; see scripts/README.md
tests/                       # Regression + unit tests
results/                     # Curated evidence and retention policy; see results/README.md
├── folds/                   # Immutable heavy folds, baselines, selector audits
└── models/                  # Reproducible learned-model artifacts and provenance
external/FloorSet/           # Contest framework + datasets (gitignored; see HANDOFF bootstrap)
docs/
├── README.md               # Documentation index and source-of-truth map
├── WINNING_PLAN.md          # AUTHORITATIVE first-place strategy + execution gates
├── CAMPAIGN_GOLDEN.md       # Prior dissection campaign: evidence and milestones
├── extracted/               # Problem statement v10 + Q&A 2026-06-18 + submission guidelines
└── archive/                 # Superseded plans/reports (history)

HANDOFF.md                   # Complete handoff: state, workflows, ranked open leads
CLAUDE.md                    # Agent onboarding + hard rules
SUBMISSION_PLAN.md           # Submission mechanics + package verification + rebuild gate
MASTER_PLAYBOOK.md           # History (exact failed variants remain dead ends)
PLAN_EXECUTION_LOG.md        # Chronological experiment log (newest at top)
PROJECT_STATUS.md            # Frozen v32 status snapshot (historical link target)

The tracked contest_solution/iccad2026_evaluate.py is an adapted convenience copy. Official verification uses the evaluator in the pinned organizer checkout at external/FloorSet/iccad2026contest/iccad2026_evaluate.py.

Quick Start

my_optimizer.py targets the FloorSet contest framework — it is not a standalone module (it imports iccad2026_evaluate → cost/utils, which live in the FloorSet checkout). Run it by copying it into the official contest dir, with the FloorSet root on PYTHONPATH (the PYTHONPATH=.. below). One-shot setup + eval: scripts/setup_and_evaluate.sh.

# Validate
cp contest_solution/my_optimizer.py contest_solution/dissect.py \
   contest_solution/topology_polish.py /path/to/FloorSet/iccad2026contest/
cd /path/to/FloorSet/iccad2026contest
PYTHONPATH=.. python iccad2026_evaluate.py --validate my_optimizer.py --quick

# Evaluate
PYTHONPATH=.. python iccad2026_evaluate.py --evaluate my_optimizer.py

(The solver is my_optimizer.py + dissect.py + topology_polish.py; sequence_pair_sa.py is only needed if the dormant SP-SA path is re-enabled. Full environment bootstrap and all workflows: HANDOFF.md §5.)

Release Checks

.venv/bin/python -m pytest                          # all tests pass (torch-dependent tests skip if torch absent)
.venv/bin/python scripts/check_public_release.py    # PASS (defaults + hashes: results/release_manifest.json)

Documentation

Doc	Purpose
HANDOFF.md	Complete handoff — state, workflows, ranked open leads (start here)
docs/WINNING_PLAN.md	Authoritative winning plan — research tracks, kill gates, beta/final calendar
docs/README.md	Documentation index — active sources of truth vs snapshots/history
docs/CAMPAIGN_GOLDEN.md	Prior dissection campaign — evidence, measured milestones
docs/SUMMARY.md	Presentation overview — problem, solution, results, insights
results/README.md	Evidence catalog — release, rollback, folds, timing, model provenance, retention
scripts/README.md	Tool catalog — supported workflows and deprecated legacy utilities
MASTER_PLAYBOOK.md	Historical strategy and decision trail
PLAN_EXECUTION_LOG.md	Chronological experiment log
PROJECT_STATUS.md	Frozen v32 status snapshot; current status is HANDOFF.md
docs/extracted/	Contest problem statement + official Q&A
docs/archive/	Superseded plans/reports (history)

Contest

• Problem: FloorSet Challenge
• Repository: IntelLabs/FloorSet
• Contest dir: iccad2026contest/
• Scoring: Cost = (1 + 0.5*(HPWL_gap + Area_gap)) * exp(2*V_rel) * max(0.7, RuntimeFactor^0.3)