DPL Two-Pass with toggle

[PrecisEDAnon]

The following content was AI generated.

Two-Pass Congestion-Aware Global Swap (DPL) — Optional “Extra DPL” Path

Summary

This PR adds an optional alternative implementation of the gs (global swap) phase used by improve_placement (DPO / detailed placement optimization). The new path is designed to help high core-utilization designs where routability-driven metrics (e.g., DRWL) become limiting.

Expected outcome (high-util targets): ~2–3% DRWL reduction near the highest utilization where the design still completes the full flow; outliers up to ~8% were observed.

Motivation / When It Helps

At high utilization (typically ~85%+ and/or “near max util that still converges”), a pure HPWL-driven global swap can “clump” cells into locally congested regions. This often harms routability and increases DRWL. The new algorithm explicitly trades bounded HPWL for congestion relief.

What This PR Changes

• New algorithm: a two-pass congestion-aware global swap (“profile then optimize”).
  • Pass 1 (profiling): run HPWL-only optimization to estimate a near-optimal HPWL target.
  • Rollback: restore the exact initial placement using Journal undo.
  • Pass 2 (guarded optimization): re-run global swap using a congestion-aware objective while enforcing an HPWL budget derived from Pass 1.
• Runtime selection (no rebuild needed): legacy vs. new global swap is chosen at runtime via a toggle.
• Tcl configurability: key hyperparameters are set via Tcl (improve_placement -global_swap_args {...}), not compile-time.

Usage / Toggle Behavior

This PR is intentionally modular:

• If your flow does not call improve_placement, nothing changes.
• If improve_placement runs but the toggle is off, behavior matches the legacy global swap.
• If improve_placement runs and the toggle is on, the new two-pass global swap runs for the gs phase.

Enable / Disable

Enable per call:

improve_placement -enable_extra_dpl 1

Or enable via environment (useful for flows/CI harnesses):

set ::env(ENABLE_EXTRA_DPL) 1

Disable:

improve_placement -enable_extra_dpl 0

Hyperparameters (Tcl)

Hyperparameters are set through -global_swap_args as key/value pairs. They are reset each improve_placement call, so flows can A/B test cleanly.

Example:

improve_placement \
  -enable_extra_dpl 1 \
  -global_swap_args {
    -tradeoff 0.4
    -profiling_excess 1.10
    -budget_multipliers {1.50 1.25 1.10 1.04}
    -area_weight 0.4
    -pin_weight 0.6
    -congestion_user_weight 35.0
    -sampling_moves 150
    -normalization_interval 1000
  }

Recommended tuning choices (2–4 each)

Tune on any circuit that struggles at high utilization near its “max util that still completes”.

Knob	Meaning	Suggested values
-tradeoff	Exploration vs. wirelength-optimal moves	0.2, 0.4, 0.6
-profiling_excess	HPWL budget slack from Pass 1	1.05, 1.10, 1.15
-budget_multipliers	Budget schedule across stages	{1.30 1.15 1.07}, {1.50 1.25 1.10 1.04}, {1.20 1.10 1.05}
-area_weight	Area term in utilization map	0.3, 0.4, 0.5
-pin_weight	Pin term in utilization map	0.5, 0.6, 0.7
-congestion_user_weight	Global scaling for congestion term	20.0, 35.0, 50.0
-sampling_moves	Samples for adaptive weighting	50, 150, 300
-normalization_interval	Utilization renormalization cadence	500, 1000, 2000

Validation Recommendation

1. Pick a circuit that becomes routing-limited at high utilization.
2. Push utilization to the highest value that still completes (or just below).
3. Run baseline vs. new path under identical conditions (seed, constraints, etc.).
4. Sweep a small subset of the tuning grid above and compare DRWL; a consistent ~2–3% DRWL reduction is the expected outcome in this regime.

CI Note

For regression detection, it is recommended to run CI with the alternate path enabled (so the new code path is exercised). After merge, this feature should remain available as an alternate setting with default = off to preserve legacy behavior unless explicitly enabled.

Related Work

Private communication.

[github-actions]

clang-tidy made some suggestions

[PrecisEDAnon]

@maliberty - PR now shows merge CI pass, and the linting should all be fixed as well. The head shows bazel weirdness, but since this resolves post-merge, it should be alright (no changes were made on the bazel end, and the tests pass locally, so I would guess this is a case of uploading the head that resolves post-merge). Please re-review and confirm if OK for merge.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[maliberty]

Linting / formatting is not resolved. Bazel is likely a known issue so I won't worry about that.

[PrecisEDAnon]

Fixed on the flagged files, and re-pushed.

[PrecisEDAnon]

@maliberty Let me know if we can run the lint checks.