Ppa Optimizer Dc Policy

PPA Optimizer (DC-based) — Policy Reference

This is a policy skill: it defines rules and reference material consumed by the ppa-optimizer-dc agent and the ppa-optimizer-dc-orchestrator. It is not user-invocable.

Optimization Philosophy

Timing-first. EDA tools (DC compile_ultra -retime -gate_clock) already perform exhaustive optimization search; LLM RTL patches add value only where tools cannot reach: missed clock gating opportunities, ill-shaped critical paths solvable via pipelining, operand isolation on idle arithmetic blocks. Escalate to the user early when improvement stalls.

Default Weights

timing = 0.7
power  = 0.2
area   = 0.1

Loaded from requirements.json["ppa_targets"]["weights"]. Sum is normalized to 1.0 at runtime; a sum <= 0 is treated as a hard configuration error.

Default Convergence Thresholds

delta_pct          = 2.0    # normal convergence Δ band
streak_required    = 3      # consecutive iterations with |Δ|<delta_pct
early_plateau_pct  = 1.0    # early-plateau Δ band (iter 1-2)
max_cycles         = 4      # hard stop unless user overrides

All overridable via requirements.json["ppa_targets"]["convergence"].

Three-Tier Termination

State cleanup (all tiers). Every terminal tier — plus the rollback halts (equivalence / smoke / timing FAIL) — removes .rat/state/ppa-loop-state.json so the project exits ppa-loop mode and hooks/rtl-edit-tracker.sh resumes normal verify-gate staleness tracking. This removal is performed by the skill/wrapper layer (rtl-ppa-optimize-dc, rat-ultraloop-ppa) and, on rollback, defensively by the orchestrator's Rollback Cleanup Protocol.

ppa-opt-done ownership. .rat/state/ppa-opt-done is emitted ONLY by the auto-loop wrapper (rat-ultraloop-ppa) after the mandatory final full Phase 5 regression on normal convergence. The one-shot rtl-ppa-optimize-dc does NOT emit ppa-opt-done (it writes only the Rule 5 rtl-verify-done marker). This keeps a terminated loop's state unambiguous: ppa-loop-state.json absent = loop finished; ppa-opt-done present = true convergence declared (triggers the Phase 6 cascade re-review).

Timing-First Heuristic (rule priority)

Rule 1 (HIGHEST · aggressive):  Timing closure
    - WNS < 0 → critical-path pipelining, retiming, logic restructuring
    - Logic levels > 12 @ 100 MHz → register rebalancing
    - REJECT any patch worsening WNS by more than 20 ps

Rule 2 (MAIN · safe):  Clock gating coverage
    - Identify ungated register banks (always_ff without enable)
    - Add enable conditions DC could not auto-infer
    - Goal: gating_efficiency_pct  80% → 90%+

Rule 3 (SECONDARY · timing-neutral):  Operand isolation
    - Target: multiplier / divider with idle cycles > 50%
    - Accepted only when timing slack unchanged

Rule 4 (SECONDARY · timing-neutral):  Resource sharing
    - Target: duplicate operator instances
    - Accepted only when timing slack unchanged

Reject Set

• Any patch worsening WNS by more than 20 ps
• Any patch creating inferred latches
• Any patch touching frozen_scope

Scope Semantics

allowed_edit_scope:
  rtl/{target_module}/**/*.sv

frozen_scope:
  rtl/common/**
  rtl/pkg/**
  rtl/intf/**

Every iteration runs validate_patch_scope.py on the generated patch.diff. Violation → rollback (scoped git checkout -- 'rtl/<target>/') + halt.

Rollback Protocol

All of these trigger rollback + halt:

• Patch touches frozen_scope
• Equivalence check FAIL
• Smoke regression FAIL
• Timing regression > 20 ps (Δ_wns < −0.02 ns)

Rollback is implemented as git checkout -- 'rtl/<target>/' && git clean -fd -- 'rtl/<target>/' to confine reversion to allowed_edit_scope. Iteration artifacts under docs/ppa-opt/iter-{N}/ are preserved for forensics. Prior to the loop starting, the skill asserts a clean working tree.

DC Tcl Fragment

See {plugin_root}/skills/ppa-optimizer-dc-policy/templates/dc-compile-ppa.tcl ({plugin_root} = plugin root resolved from .rat/state/spawn-context.json). The PPA orchestrator composes a thin wrapper Tcl that performs DC setup/analyze/elaborate/link, sources this fragment (with {plugin_root} expanded to the absolute path, since dc_shell runs in the user project CWD), then runs it via run_syn.sh --tool dc_shell --script <wrapper.tcl> (the --script path bypasses run_syn.sh auto-generation). Wiring: agents/ppa-optimizer-dc-orchestrator.md.

PPA Brief Scaffold

When requirements.json is missing ppa_targets, the skill writes back the content of {plugin_root}/skills/ppa-optimizer-dc-policy/templates/ppa-brief-scaffold.json ({plugin_root} = plugin root resolved from .rat/state/spawn-context.json) and halts with user prompt.

Early-Plateau Escalation Report Template

# reviews/ppa-opt/early-plateau-escalation.md

## Verdict: EARLY_PLATEAU

{if data_points < 2}
> **Caution:** Only {N} iteration delta point(s) observed before plateau detection.
> Single-point plateau detection indicates the loop halted conservatively; consider
> forcing one additional iteration via `requirements.json["ppa_targets"]["convergence"]["max_cycles"]`
> before relying on this conclusion.
{endif}

EDA tool auto-optimization appears saturated; RTL-level patches yielded
< {early_plateau_pct}% weighted improvement over the observed iterations.

## Iteration history

| iter | wns_ns | power_mw | area_um2 | weighted_Δ |
|------|--------|----------|----------|------------|
| ...  | ...    | ...      | ...      | ...        |

## Current bottleneck (from latest ppa-report.json)

- Critical path: {from} → {to} (slack {slack_ns} ns, {logic_levels} logic levels)
- Power hotspot: {group} {pct}%, clock network {clock_pct}%
- Clock gating efficiency: {gating_efficiency_pct}%

## Recommended user actions

1. Review μArch (pipeline depth, algorithm variant)
2. Relax spec targets (clock / power budget)
3. Evaluate technology changes (stdcell library, Vt mix)
4. Force continuation via:
   requirements.json["ppa_targets"]["convergence"]["max_cycles"] = N  (N > 4)

Phase Integer Mapping

The ppa-opt phase is identified in phase-registry.json by the string key "ppa-opt" with "order": 5.5 (semantic: between P5 verify and P6 design-note). For POSIX shell compatibility, sctx_skill_to_phase() emits integer 8.

Rationale: the integer space 1–7 is reserved for canonical pipeline phases. Integer 8 is used for the optional post-verify stage. Any future phases beyond 7 should follow this extension pattern and also declare "integer_map": <N> in the registry entry so the dual encoding stays discoverable.

References

• Synopsys Design Compiler User Guide — compile_ultra, set_clock_gating_style
• Synopsys Power Compiler User Guide — set_power_opt, report_power -analysis_effort
• Weste & Harris, CMOS VLSI Design — clock gating methodology