acceptodds
Under review as a conference paper at ICLR 2027

Coarse-to-Fine Macro Placement via Evolutionary Search and Critical Macro Tuning

Abstract

Macro placement is a critical stage in chip physical design that substantially affects downstream implementation quality. Recent search-based methods improve existing layouts through partial reconstruction, but quality-biased or spatially restricted macro selection can limit the diversity of reconstruction proposals, potentially hindering escape from local optima. Moreover, coarse-grid representations restrict placement precision. To address these challenges, we propose C2FPlace, a Coarse-to-Fine macro Placement framework that integrates population-based evolutionary search with fine-grained refinement. During coarse-grained optimization, tournament selection chooses promising parents from randomly sampled groups of layouts, and stochastic partial rip-up and re-place generates offspring by sampling macro subsets across the entire layout. A two-phase schedule samples reconstruction ratios from a higher range early in the search and a lower range later, supporting broad exploration followed by more conservative refinement. During fine-grained optimization, critical macro tuning enables positional adjustments beyond the coarse grid to obtain additional half-perimeter wirelength (HPWL) reduction. Experiments on the ISPD2005 benchmark show that C2FPlace reduces HPWL by 17.82% over EGPlace and 17.86% over RollPlace on average. On the ICCAD2025 benchmark, C2FPlace achieves the best average ranking among the compared methods under the evaluated power, performance, and area (PPA) metrics.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Physical Design, Macro Placement, Evolutionary Search

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_7UUrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes C2FPlace, a three-stage macro placement framework: (1) a population of 20 initial layouts built by repeated WireMask-guided greedy placement with random tie-breaking; (2) coarse-grained steady-state evolutionary search using tournament selection (size 5), random-subset partial rip-up-and-replace across the whole canvas, a two-phase rip-up ratio schedule ([0.4, 0.7] → [0.1, 0.4]), and elitist worst-replacement; (3) fine-grained “critical macro tuning” that identifies macros providing bounding-box extremes of nets and refines them via structure-preserving coarse-grid displacement combined with fine-grid offset search. Although I believe the paper contains substantial substance, I find several flaws that prevent it from meeting the bar required for ICLR.

Critical strengths
  1. EGPlace's fitness-softmax parent selection and score-guided macro subset, and RollPlace's rectangular-region selection, both bias reconstruction proposals; C2FPlace's tournament selection + spatially-unconstrained random subset + scheduled ratio is a coherent response.
  2. The 1000-minute unified budget with convergence curves rerun on the authors' own hardware (Fig. 2) is commendable and often missing in this literature.
  3. OpenROAD v2.0 + ASAP7 PPA evaluation (WNS/TNS/NVP/power) on ICCAD2025 goes beyond HPWL-only claims; the authors honestly report that Stage 3 HPWL gains do not always translate to timing gains.
Critical weaknesses
  1. The motivation is that selection mechanisms drive diversity, yet no ablation isolates them. I find that tournament vs. fitness-softmax selection, random vs. score-guided/region-based macro subset, and the diverse-population vs. single initialization, but all of them are within the same framework. This could come from the initialization diversity or budget rather than the proposed selection scheme. I strongly believe that this is the most significant experimental gap.
  2. All five evaluated circuits contain 0 macros, ariane with 136 macros. However, the paper does not provide a clear justification. Only saying "its placement setting differs" is not enough and too vague. I assume applying the method to standard-cell placement is a different problem regime, the scalability claim should be reframed and ariane should included.
  3. Baseline numbers are taken from EGPlace and ROllPlace paper rather than rerun. This is unprofessional and it is unclear whether the headline 17.8% improvements hold under identical compute, especially since C2FPlace allocates its budget explicitly? Please explain.
  4. I believe stage 3 is a marginal contributor. The paper yields only .31% average HPWL on ISPD2005 and on ICCAD2025 it degrade WNS/TNS on ac97_top and pci_bridge 32. I do believe the work as claimed effectiv contradicts the paper's own numbers, either the claim or the title emphasis should be tempered.
  5. The novelty of the paper is incremental . For example, Tournament slection, annealed rip-up ratio, bounding box critical refinement, and fine grid offsetting are each standard techniques. The contribution is a careful recombination rather than a new principle.
  6. I do not know if this is intend. There is no significance testing. Several stage 3 deltas are within noise. The objective is HPWL under non-overlap only. No congestion, routability analysis on ISPD2025 is provided. The HPWL ->PPA disconnect is noted but no analyses are provided as well? Any reasons for this?

several typos I would not list them all, but name a few. In line 402-406, the sentence is broken. Line 430, "rwith" what is this? ... Line 131, there are two schedule...

Questions for authors
  1. Can you provide the selection-mechanism ablations (tournament vs. softmax; random vs. score-guided subset) under identical budgets?
  2. Why exactly is ariane excluded, and can C2FPlace handle genuine mixed macro + standard-cell designs?
  3. What defines the fine-offset candidate set in Stage 3, and how sensitive is it to resolution?
  4. Why does DREAMPlace still win on adaptec4 — any circuit characteristics explaining this?
Reviewer_wvDqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes C2FPlace, a population-based search framework for macro placement. Its central motivation is that quality-guided or spatially restricted selection of macros for reconstruction may limit exploration. C2FPlace combines tournament selection with random-subset rip-up-and-re-place, using larger reconstruction ratios early and smaller ratios later. A final refinement stage adjusts macros that determine net bounding-box boundaries, including positional refinement beyond the coarse grid. Experiments include HPWL comparisons on ISPD2005, convergence curves under a common runtime budget, and standard-cell placement experiments with downstream timing and power estimates on ICCAD2025. The results support the competitiveness of the overall framework, although attribution of the gains to the central search mechanism and their downstream implications remain incompletely established.

Critical strengths

1.The paper addresses an important optimization problem and reports substantial empirical improvements. Macro placement materially affects downstream physical design. Table 2 reports average HPWL reductions of approximately 17.82% over EGPlace and 17.86% over RollPlace, with sizeable improvements on several circuits. The proposed separation between quality-aware parent selection and unrestricted random-subset reconstruction is intuitive and potentially useful: stronger heuristic guidance need not always produce better exploration. The framework is also straightforward and does not require model training. 2.The experimental evidence supports the value of the coarse search procedure, beyond an additional refinement pass. Figure 2 reruns selected baselines on common hardware under a unified 1000-minute budget and reports variability across five runs. Moreover, most of the improvement remains when Stage 3 is removed: the average reduction relative to the Table 2 EGPlace results is still approximately 17.58%. Thus, additional post-processing is insufficient to explain the main results. The ICCAD2025 experiments also provide useful supplementary evidence of applicability to standard-cell placement and downstream timing and power estimates.

Critical weaknesses

1.The experiments do not sufficiently separate the proposed search insight from changes in resolution and population configuration. C2FPlace uses a 512×512 grid and a population of 20, whereas EGPlace defaults are a 224×224 grid and a population of 5. Section 5.2.2 states that rerun baselines retain their original hyperparameters. Matching runtime therefore does not establish whether the gains primarily arise from random-subset reconstruction, finer positional resolution, or a different population configuration. Appendix B provides useful partial evidence: C2FPlace remains competitive at a 224 grid, so resolution alone cannot simply explain away the results. Nevertheless, a direct comparison under shared settings is needed to establish the central algorithmic contribution. This matters because tournament selection and staged perturbation are conventional, while Stage 3 contributes only about 0.31% additional HPWL reduction on ISPD2005. The substantive novelty therefore rests primarily on the coarse search mechanism. My concern is attribution of that contribution, rather than the absence of an exhaustive ablation study. 2.The transfer of macro-stage improvements to downstream full-design placement quality remains unverified. The ISPD2005 results demonstrate improvements in the reported macro-placement objective, but the paper does not report global-HPWL after feeding these macro placements into a common downstream standard-cell or mixed-size placement flow. Given the missing timing and technology information in the original ISPD2005 Bookshelf data, using ICCAD2025 for timing and power evaluation is reasonable; I do not expect direct OpenROAD PPA validation on ISPD2005. However, the five evaluated ICCAD2025 circuits contain no macros, and the method is applied directly to standard-cell placement. Those experiments therefore do not establish whether the improved ISPD2005 macro placements remain beneficial after subsequent placement and legalization. A focused comparison of downstream full-design HPWL, using a shared placement backend and clearly specified macro-movement policies, would address this gap without requiring a complete PPA or routing flow. This limits the evidence for downstream practical benefits rather than invalidating the macro-stage optimization results.

Questions for authors

1.Comparison settings and attribution: Please clarify the grid size, population size, initialization procedure, and inclusion of initialization time for each method rerun in Figure 2, as well as how those runs relate to the baseline values quoted in Table 2. Under shared grid, population, initialization, and runtime settings, does C2FPlace retain a substantial advantage over EGPlace? Within that shared setup, a focused comparison replacing score-guided macro selection with random-subset selection would be particularly informative. This would materially affect my assessment of the central insight; a full factorial study is unnecessary. 2.Downstream full-design placement: Have the macro placements generated on ISPD2005 been evaluated through a common downstream standard-cell or mixed-size placement flow? If so, please report the resulting global- HPWL and specify the backend version, benchmark variant, and whether macros remain fixed or can move during subsequent optimization. A focused evaluation on representative circuits would help establish whether the macro-stage gains survive downstream processing. If this evidence is unavailable, please clarify the distinction between the demonstrated macro-stage improvements and the separate standard-cell results on ICCAD2025, and bound the downstream-benefit claims accordingly.

Reviewer_3SdWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

C2FPlace combines evolutionary search with local refinement for macro placement. It searches by removing and greedily reinserting subsets of macros, using larger subsets early and smaller ones later, then refines macros that determine net bounding boxes at coarse and fine resolutions. The paper reports roughly 17.8% average HPWL reductions over EGPlace and RollPlace on ISPD2005. It also includes runtime comparisons, ablations, and post-placement timing and power results on five standard-cell-only ICCAD2025 circuits.

Critical strengths
  • The reported HPWL gains over recent baselines are substantial, with the best mean result on seven of the eight ISPD2005 circuits (Table 2).
  • The runtime comparison uses the same hardware and time budget across five seeds, providing useful evidence of search efficiency (Figure 2).
Critical weaknesses
  • The ablations do not clearly establish how much of the gain over EGPlace comes from the proposed search changes rather than grid resolution and other configuration choices. The final refinement adds only about 0.31% average HPWL reduction on ISPD2005 (Table 2).
  • The downstream evaluation uses five circuits with no macros, leaving the timing and power benefits on mixed macro/standard-cell designs unclear. The refinement also worsens timing on two of these cases despite reducing HPWL (Tables 3–4).
Questions for authors
  • How much of the gain over EGPlace remains when grid resolution and population size are matched under the same runtime budget?
  • Does Stage 3 outperform simply continuing the coarse search for the same additional time?
  • What differences in the placement setup led to excluding Ariane, and do you have timing and power results on any mixed macro/standard-cell design?
Reviewer_H2voOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes C2FPlace, an evolutionary search-based macro placement framework that contains coarse-grained placement exploration and fine-grained macro tuning. It first generates a population of WireMask-guided greedy placements, and then performs evolutionary search using tournament selection, random macro-subset rip-up and re-placement, and a two-phase perturbation-ratio schedule. Finally, it refines macros whose pins define the boundaries of net bounding boxes through coarse-grid displacement and fine-grid offset adjustment.

Experiments on the ISPD2005 benchmark show that C2FPlace achieves lower HPWL on most designs. On the ICCAD2025 benchmark, C2FPlace obtains better average ranks in HPWL, timing, and power than the baseline macro placement methods.

Critical strengths
  1. The idea of rip-up and replacement is novel.

  2. C2FPlace achieves the lowest HPWL on seven of the eight ISPD2005 circuits and reports substantial average improvements over EGPlace and RollPlace.

Critical weaknesses
  1. The experimental comparisons are primarily based on HPWL, which has limited relevance to final physical design quality (e.g., PPA). Although the experiments on the ICCAD 2025 benchmark report downstream metrics such as timing and power, the choice of benchmark cases is questionable. The selected cases contain only standard cells, making them unsuitable for evaluating macro placement methods. If the authors intend to evaluate C2FPlace as a cell placement method, analytical placers such as DREAMPlace should be included in the comparisons, rather than relying solely on macro placers such as WireMask-EA. Moreover, the deliberate exclusion of Ariane cases, which contain macros, is particularly puzzling in the context of macro placement experiments.

  2. The proposed evolutionary framework lacks effective mechanisms for maintaining population diversity. Specifically, the population is updated through elitist selection and the rip-up-and-replace operator primarily focuses on local modifications, which limits the ability to discover substantially different solutions. As a result, the population may rapidly converge toward similar solutions, increasing the risk of becoming trapped in local optima.

Questions for authors
  1. What does the HPWL exactly denote in the reported experiments? Please clarify whether it is only macro-related HPWL or involves standard cells.

  2. Why are standard cell-only designs from ICCAD 2025 used to evaluate macro placement methods? Why is ariane, the only design containing macros, excluded? Can the authors evaluate downstream benefits on mixed-size benchmarks that contain macros?

  3. Could the authors provide experiments with different population sizes? Could you analyze how the population evolves during the search, particularly in terms of solution diversity?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.