GAN-Place: Advancing Open Source Placers to Commercial-quality Using Generative Adversarial Networks and Transfer Learning

Recently, GPU-accelerated placers such as DREAMPlace and Xplace have demonstrated their superiority over traditional CPU-reliant placers by achieving orders of magnitude speed up in placement runtime. However, due to their limited focus in placement objectives (e.g., wirelength and density), the placement quality achieved by DREAMPlace or Xplace is not comparable to that of commercial tools. In this article, to bridge the gap between open source and commercial placers, we present a novel placement optimization framework named GAN-Place that employs generative adversarial learning to transfer the placement quality of the industry-leading commercial placer, Synopsys ICC2, to existing open source GPU-accelerated placers (DREAMPlace and Xplace). Without the knowledge of the underlying proprietary algorithms or constraints used by the commercial tools, our framework facilitates transfer learning to directly enhance the open source placers by optimizing the proposed differentiable loss that denotes the “similarity” between DREAMPlace- or Xplace-generated placements and those in commercial databases. Experimental results on seven industrial designs not only show that our GAN-Place immediately improves the Power, Performance, and Area metrics at the placement stage but also demonstrates that these improvements last firmly to the post-route stage, where we observe improvements by up to 8.3% in wirelength, 7.4% in power, and 37.6% in Total Negative Slack on a commercial CPU benchmark.


INTRODUCTION
With the ever-increasing technology scaling driven by Moore's Law, modern Very Large-Scale Integration (VLSI) designs easily comprise millions, if not billions, of instances that must be placed onto constrained layouts.Unfortunately, existing commercial Electronic Design Automation tools heavily rely on CPU-intensive heuristics or analytical objectives to solve the placement 32:2 Y.-C.Lu et al. problem, which is often inadequate in generating high-quality placements within a reasonable amount of runtime.To address this issue, open source GPU-acceleratable analytical placers, DREAMPlace [11] and Xplace [13], are proposed to exploit the parallelization capabilities of GPU kernels to significantly reduce placement runtime.Particularly, the original CPU-intensive objective functions are reformulated into GPU-acceleratable cuda kernels that leverage GPU-specific blocks and threads to expedite analytical computations without sacrificing solution quality.
Despite the significant success achieved by DREAMPlace and Xplace, the placement solution quality of these GPU-accelerated placers is still inferior to those of commercial tools.The key reason is that these placers, as there CPU counterparts, the ePlace/RePlace family [3,14], only have limited placement objective focus on wirelength and density.Particularly, they cannot consume many other essential constraints (e.g., timing, power) as commercial Physical Design (PD) tools, leading to inferior placement quality in terms of full-chip Power, Performance, and Area (PPA) metrics.This motivates us to ask the following question: Is there any way to advance DREAMPlace toward commercial-quality without knowing the secret sauces of those black-boxed commercial engines?In this article, we prove that the answer is yes, particularly via generative adversarial learning [20].
In this work, we present the first-ever learning-driven placement optimization framework named GAN-Place that can be directly integrated with any GPU-accelerated placer (e.g., DREAM-Place or Xplace) to advance the placement quality toward commercial-quality using generative adversarial learning.The key idea is that although we do not know the underlying algorithms or constraints used by the tools, we can "quantify placement similarity" between tool-optimized placements and DREAMPlace-or Xplace-generated placements using generative learning, and by optimizing the differentiable similarity scores computed from the proposed discriminators, we can narrow the distribution gap between open source and commercial placers.Our is a simple yet highly effective, which is greatly motivated by the success of Generative Adversarial Networks (GANs) [5] in real-world applications, where signals in different domains (e.g., texts and images) can be converted to each other by parameterizing target distributions using differentiable frameworks.
Figure 1 presents a high-level overview of GAN-Place, where the underlying GPU-accelerated placer is considered as a "generator" whose goal is to generate the placements that follow similar distributions as the tool-optimized ones in the databases (which are optimized for different PPA objectives).To achieve this, two discriminators built upon Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN) are developed to quantify the placement similarity between two types of placement (i.e., gpu placer-generated and tool-optimized).The goal of the discriminator is to make correct judgements by telling whether its input is coming from commercial databases or DREAMPlace, whereas one of the objectives of DREAMPlace (in our GAN-Place settings) is to "fool" the discriminator by generating similar (or realistic) placements as the ones in the database.Specifically, the GNN-based discriminator is dedicated to encode netlist connectivity with cell locations as node attributes, and the CNN-based discriminator is dedicated to encode bin-density maps resulted from the underlying cell locations.
The goal of this work is to demonstrate that the placement quality (or style) of a placer (e.g., commercial placer) can be transferred onto another placer without knowing the underlying algorithms or constraints through generative adversarial learning.Particularly, we show that any GPU-accelerated placer that serves as a differentiable placement engine can be elegantly considered as a generator that generates tool-alike placements of the underlying gate-level input netlists.The greatest strength of GAN-Place is that it enables any differentiable placer to optimize the underlying placement toward a commercial tool-verified (and optimized) direction without explicitly knowing the black-boxed algorithms.Although DREAMPlace or Xplace leverages fundamentally GAN-Place 32:3 Fig. 1.High-level overview of our GAN-Place framework that performs placement optimization using generative adversarial learning.Note that GAN-Place can be integrated with any GPU-accelerated placer to improve placement quality.In this work, we particularly showcase the integration with DREAMPlace [11] and Xplace [13].
different algorithms compared with the tools, in this work, we demonstrate that the placement quality of these open source placers can be efficiently improved with the proposed framework GAN-Place.

RELATED WORKS AND MOTIVATIONS
Analytical placers [2,3,7,8,14] have brought tremendous success to the semiconductor industry since the early 2010s [10].Nevertheless, recent advancements in ML Theory and its applications have further pushed modern placers to an unprecedented frontier [9].Based on the characteristics of learning strategies, existing learning-driven frameworks for VLSI placement can be categorized into the following streams: -Supervised Placement Quality Prediction: These categories include the works who strive to predict crucial placement-related metrics in early stages of the design flow, which typically rely on pre-built databases with comprehensive data.A recent work [4] developed an ensemble framework to predict post-place congestion and timing metrics based on a massive database that contains millions of instances from 72 industrial designs.To extend the prediction scope from single-stage to full-flow modeling [15,18], developed learning-driven flow-based graph modeling techniques using GNNs to predict end-of-flow PPA metrics for each intermediate placement stage, allowing designers to perform efficient design space exploration.-Unsupervised Placement Optimization: This category refers to the works who aim to develop ML frameworks that directly perform placement optimization without a pre-built database.
In other words, they often directly consider ML models as optimization rather than prediction frameworks.Reference [19] presented the PL-GNN framework that generates cell clustering constraints as placement guidance using GNNs to advance commercial placers.
The key idea is to drive commercial placers to spend additional effort in placing the cells belonging to the same ML-predicted cluster closer to each other to improve the overall PPA metrics.However, the framework in Reference [19] is not "goal-directed" as the graph learning and the clustering steps are not end-to-end-differentiable, which prohibits GNN models from utilizing feedback from the achieved optimization results.To overcome this problem, the authors of Reference [21] developed differentiable PPA-inspired ML loss functions to improve the GNN learning process with direct optimization feedback, which includes timing, power, and congestion evaluations of the clustering results.-Reinforcement Learning (RL) Placement Optimization: RL is a promising paradigm that has shown superior optimization results in high-dimensional control problems.The authors of Reference [22] presented a seminal work that leveraged RL for macro placement to replace human labor, which significantly improves the chip design turn-around time.Another work [1] utilized RL to efficiently tune the placement parameters of commercial placers through thousands of iterations.Still another work [24] further combined simulated annealing and RL in a cyclic fashion to conduct iterative placement optimization.However, given that RL algorithms typically demand significant training time, this article focuses on placement optimization techniques that do not incur excessive overhead to enhance placement quality.
In this work, we take a brand new perspective to solve the VLSI placement optimization problem using generative adversarial learning [5].In the realm of PD, previous work [16] has shown that the generative learning conducted by GAN can elegantly optimize the Clock Tree Synthesis (CTS) process of commercial tools without knowing the algorithms inside the black-boxed CTS engine.Motivated by Reference [16], in this article, we aim to leverage GAN-based models to demystify commercial black-boxed placers so as to improve open source placers: DREAM-Place [12] and Xplace [13] toward commercial-quality.Particularly, we consider any vanilla open source placer as a generator in a conventional GAN-based framework whose goal is to generate commercial-quality placements from an unplaced input netlist.To achieve this, we first build a relevant database containing commercial-quality placements with different objectives (e.g., timingdriven, power-driven, congestion-driven, etc.) using Synopsys ICC2, an industry-leading commercial tool.We then develop a discriminator that serves as a teacher to improve DREAMPlace by optimizing the similarity scores between DREAMPlace-generated placements and the ones in the commercial database as shown in Figure 1.By minimizing the similarity loss, the underlying cell locations of DREAMPlace-or Xplace-generated placements will be updated through gradient descent, which effectively become more similar to the tool-generated placements in the database.

GAN-PLACE OVERVIEW
It is widely acknowledged that generative adversarial learning is a promising paradigm that effectively captures complicated distributions using generative and discriminative models that have "opposite" objectives.Generally speaking, the goal of the generator is to generate target distribution alike data from random (or non-meaningful) distributions, while the goal of the discriminator is to distinguish the source of its inputs (i.e., from the generator or from the target distribution).Note that both generator and discriminator can be realized by any differentiable system (i.e., not necessarily neural networks).As aforementioned, in this article we consider DREAMPlace, a differentiable placement system, as a generator whose goal is to generate placements that follow similar distributions as the ones in a tool-optimized database.
Our discriminator leverages CNNs and GNNs to determine the origin of its input source that alternates in each iteration of GAN training, where CNNs are responsible to encode cell bin-density maps and GNNs are responsible to encode netlist connectivity.The rationale behind using GNNs for netlist encoding is that netlists are essentially hypergraphs whose node connectivity is critical to placers.The motivation behind using CNNs for bin-density map encoding is shown in Figure 2.  We observe that under the same global density target, the commercial tool has extra intelligence on locally aggregating or loosening cells to improve design PPA metrics globally, where DREAMPlace naively strives to make every local bin to have the same local density target as the global density target.This density variation is proven to be critical to the success of placement optimization in Reference [21].The observation shown in Figure 2 strongly motivates us to leverage bin-density map as one the indicators of placement similarity, and CNNs thus become the second-to-none choice to perform encoding on it as they are well known for grid signals (e.g., images) classification.
Figure 3 shows the key objective difference between the proposed framework GAN-Place and the vanilla GPU-acceleratable (i.e., differentiable) placer.Aside from optimizing the wirelength (WL) and density objectives as in the vanilla placer on the left, in each placement iteration, GAN-Place further computes a differentiable similarity loss using GNN-based and CNN-based discriminators.Note that the in our generative settings, the input design of GAN-Place and the designs in the commercial database are not necessarily the same.That is, GAN-Place facilitates transfer learning to perform placement optimization on unseen netlists, which is the greatest strength of the proposed method.float ([0.7,0.9])density of global placement target_route_density float ([0.7,0.9])density of early global routing

Commercial Database Construction
A high-quality placer is helpful to demonstrate the proposed placement optimization technique using generative adversarial learning.In this article, we take Synopsys ICC2, an industry-leading commercial tool, as our reference tool for database construction.Particularly, we sweep around essential placement parameters offered by ICC2 as shown in Table 1.The combinations of these parameters form a high-dimensional space, leading to a variety of placements that have distinct PPA focus, including performance-driven placements, low-power placements, routability-driven placements, and so on.Nonetheless, we want to emphasize that our framework GAN-Place is not limited to any objective focus.It can be equipped with any database to transfer the placement quality (or style) within onto the targeted placer (e.g., DREAMPlace or Xpalce).More importantly, the designs in the database and the designs that the targeted placer is optimizing do not have to be the same as GAN-Place facilitates transfer learning for placement optimization.The goal of this work is to demonstrate that GAN-Place can perform efficient placement optimization on unseen netlists using a tool-optimized database where the placements are not necessarily coming from the same design.

GAN-PLACE ALGORITHMS
The key idea of GAN-Place is to transfer the placement quality (or more precisely, style) of one placer onto another through generative adversarial learning.To achieve this, a discriminator built upon GNNs and CNNs is developed to quantify similarity of placements between different placers.Since the similarity scores have to be differentiable to update the underlying cell locations (so as to improve DREAMPlace), a differentiable graph pooling methodology is adopted, and a differentiable bin-density map transformation technique named Soft-Bin is proposed.

VLSI Netlist Encoding Using GNNs
Graph representation learning conducted by GNNs is an effective technique to encode netlist connectivity and the underlying attributes into a meaningful graph-or node-level representations [17] In the realm of PD, such encoded representations have been leveraged to successfully solve many tasks that are once considered extremely hard to solve [23].In this article, we specifically leverage GNNs to encode netlist graph G = (V , E) of the underlying placement while considering cell locations as initial attributes.The results graph-level vectors are further taken as the input of the proceeding GNN-based discriminator network to decide the similarity between different placements.
To apply GNNs for solving VLSI problems, a netlist transformation technique is required to transform the netlist hypergraphs into normal graphs.A naive transformation that assigns a GNN message passing edge between each cell on the same net will introduce O (n 2 ) edges, which may GAN-Place 32:7 Fig. 4. Illustration of GNN netlist transformation and graph pooling.First, following Reference [21], we perform timing-aware netlist transformation by only taking timing arcs as GNN message passing edges.Then, we perform attention graph pooling to construct the final graph-level vector that characterizes the input placement.Note that the poolings are differentiable (i.e, learnable), meaning that the clustering results (e.g., {a, c, e} is now a cluster) will change across training iterations.The obtained graph-level vector is fed to downstream networks to quantify placement similarity.
damage the quality of graph learning as redundant edges limit the expressiveness of GNNs [25].To overcome this issue, in this work, we adopt the netlist transformation technique proposed in Reference [21] that for every net in the original netlist graph, only driver-to-load connections are added in the transformed graph, and beyond that, artificial edges are introduced between every start points and end points of timing paths.
Figure 4 illustrates the graph learning process of the proposed GAN-Place framework.Starting from an input placement, we first perform timing-aware netlist transformation as in Reference [21].The transformation only preserves timing arcs as GNN message passing edges.Then, following from GraphSAGE [6], an inductive based message passing process is leveraged to transform the initial features of each node into better representations through neighborhood aggregation as where k denotes the transformation level, σ denotes the sigmoid function, Neiдh(v) denotes the neighbors of node v, and W aдд k and W pr oj k denote the aggregation and projection matrices at the kth layer of the GNN model.Note that Equation ( 1) is repeated for every cell in the design.In the implementation, our GNN has three layers (K = 3), and each of them has a 32 hidden dimensions.Hence, the final node embeddings {h 3 v , ∀v ∈ V } has 32 dimensions.At each level k of the node representation learning, a differentiable graph attention pooling mechanism [26] is applied to coarsen the graph via a soft clustering assignment C k , where nodes belonging to the same cluster will be merged into a super-node through mean pooling, which can be derived as where x |V | k denotes the mapping of nodes between level k and level k + 1, where a node v may be merged into a super-node or stay as itself.At the end of the entire graph representation learning, a mean pooling is applied across the remaining nodes to obtain the final graph-level vector д, which is taken as one of the indicators of placement similarity.A graphical illustration of the entire graph learning process is shown in Figure 4.With the graph pooling strategy, netlist information will be iteratively condensed into a graph-level vector д that characterizes the underlying input placement G.  Finally, we want to emphasize that both GAN-Place generated placements and tool-optimized placements go through the same graph learning process as shown in Figure 4. Let D дnn denote the discriminator network parameters that are related to graph representation learning, the graph representation learning objective L дnn can be derived using cross-entropy as Equation ( 3) reflects the adversarial nature in our GAN-Place framework that depending on the sources of input, we train the discriminator to make corresponding correct judgements.With enough training iterations, our GNN module is able to extract the information that tells the key difference between DREAMPlace-generated placement and the tool-optimized ones.

GNN-based Discriminator
Figure 5 demonstrates the adversarial learning conducted by GNNs in the proposed framework GAN-Place.By applying node representation learning on the input placement, which is either from the commercial database or from the differentiable placement engine, we train the GNN model to differentiate various placement sources based on netlist connectivity and assigned cell locations.Note that the goal of GAN-Place is to generate "realistic" placements that are similar to the ones in the database so as to "fool" the discriminator to predict its generated placements as coming from the database.Nonetheless, GNN alone is not enough to characterize and differentiate different placements.For example, to truly encode the difference of the bin-density distribution as shown in Figure 2, a more direct approach that quantifies the layout information is needed.Hence, in this work, we leverage CNN-based networks to encode such information.

Soft-Bin: Differentiable Two-dimensional Bin-Density Map Transformation
So far, we have introduced how GAN-Place encodes netlist connectivity using GNNs by leveraging a differentiable attention pooling mechanism with cell locations as initial node features.Now, we present the details of how GAN-Place leverages bin-density maps to characterize and differentiate different placements with CNNs, where the key motivation behind is illustrated in Figure 2. It should first be mentioned that most common bin-density calculation method using the simple formula of dividing the total cell area in a bin by the total bin area is not differentiable, as for j = 0; j < num_h; ++j do 7: convert expected area to expected density the act of assigning a cell to a particular bin is deterministic.Although such computation is "exact" and accurate, a probabilistic calculation method is needed to compute gradients based on cell (x, y) locations so as to improve the underlying placement through gradient descent.To achieve this, we develop a differentiable bin-density transformation technique named Soft-Bin as shown in Algorithm 1.
The key idea behind Algorithm 1 is that instead of deterministically assigning each cell to an exact bin purely based on its location, we can probabilistically distribute the area of a cell onto its neighboring bins, which can be achieved by any activation function that maps real values into probabilities.In this article, we use softmax for such computation.The beauty of our probabilistic bin-density calculation approach is that it enables the underlying cell locations to be updated along with any operation (i.e., maximization or minimization) of the computed density.That is, with the proposed Soft-Bin technique, cell locations can be directly updated using gradient descent by optimizing the similarity loss computed from the CNN-based discriminator.
The proposed Soft-Bin algorithm works as follows.In Lines 1-4, we initialize the bin density map M based on the specifications from inputs.The cells corresponding to each bin can be obtained using Lines 7 and 8. Now, as shown in Lines 10-15, for each cell that deterministically belongs to a target bin b, we first identify its neighboring bins (adjacent or diagonal) neiдh_bins and then compute a distance vector dist_vec that denotes the Euclidean distance between the target cell to  each bin (including b and neiдh_bins).Finally, we transform the distance vector dist_vec into a probability vector prob_vec using the softmax function as shown in Line 16, and the "expected" area contribution can be calculated using Line 17.After updating the bin-density map M by the expected area contribution of each cell in the design, we obtain the final density of each bin using Line 19.

CNN-based Discriminator
With the proposed Soft-Bin technique for differentiable bin-density map transformation, we now describe the adversarial learning conducted by the CNN-based discriminator as shown in Figure 6.Note that the key motivation for using CNNs to encode and differentiate different placements is originated from Figure 2, where we clearly observe that the commercial tool is having additional intelligence on locally spreading or coarsening cells for PPA optimization that DREAMPlace is lacking of.Therefore, to advance DREAMPlace toward commercial quality, we aim to improve the DREAMPlace generated locations by following the bin-density distributions as the ones generated by the commercial tool.
As shown in Figure 6, we use the proposed Soft-Bin technique to transform the DREAMPlace generated placement into a differentiable "soft" bin-density map, while using exact (i.e., deterministic) computation to obtain the bin-density map from the commercial database, as it does not need to be differentiable.By considering the input density maps as grid signals, CNNs are used to perform convolution to extract key information that can be used to decide whether the input is coming from DREAMPlace or the database.

End-to-End GAN-Place Training
Figure 7 shows the detailed architecture of the CNN-based and the GNN-based discriminators.Note that placements originating from GAN-Place do not have to be the same as the ones from the commercial database, since our framework does not explicitly take design information such as number of cells or number of nets as features (i.e., not memorizing).Instead, GAN-Place learns to parameterize placement distributions in terms of netlist connectivity and the achieved bin-density maps.All the computations involved are "size-independent, " meaning that netlists in different sizes GAN-Place 32:11 will be encoded to vectors in the same dimensions, which is true for both CNN-based and GNNbased discriminator networks.
The key difference between the proposed GAN-Place and a vanilla GPU-accelerated (differentiable) placer is shown in Figure 3. Beyond the conventional placement objectives that are wirelength and density, we consider the underlying placement engine in our GAN-Place framework as a generator to maximize the probability of the generated placement being classified as "tooloptimized" by the GNN-based and CNN-based discriminators.The similarity loss L sim computed by the entire discriminator (GNN based and CNN based) can be obtained as where D дnn and D cnn denote the GNN-based and CNN-based discriminators, respectively; (G, x, y) denotes the sampled (either from the database or from DREAMPlace) netlist graph with annotated cell locations; exact_map denotes the deterministic bin-density map; and So f t − Bin denotes the proposed transformation technique.Finally, the similarity loss L sim will be jointly optimized with the original wirelength and density objectives (L W L and L density ) in any vanilla GPU-accelerated placer (e.g., DREAMPlace or Xplace) as where L W L and L density denote the wirelength and density objectives computed from the vanilla DREAMPlace and {λ} denote the hyper-parameters for loss weighting.At each placement  optimization, DREAMPlace will optimize Equation ( 5) to improve the underlying placement toward commercial-quality.

Full-Flow Integration
Figure 8 shows a graphical illustration of integrating the proposed framework GAN-Place into an industrial design flow implemented by Synopsys ICC2.To perform fair comparisons between vanilla placers (e.g., DREAMPlace and Xplace) and the proposed GAN-Place, we use the same ICC2 parameters and seeds to implement the rest of the design flow after global placement to remove run-to-run In the experiments, we demonstrate that GAN-Place efficiently improves both DREAMPlace [11] and Xplace [13] in critical PPA metrics at each major stage of the PD flow.Note that in this work, we not only demonstrate that GAN-Place immediately improves the PPA metrics measured at the placement stage but also show that the improvements last firmly to the post-route stage.

EXPERIMENTAL RESULTS
In this article, we validate GAN-Place with three commercial CPU benchmarks and five Open-Core benchmarks using the TSMC 28-nm technology node.The commercial database is generated by randomly sampling the parameters listed in Table 1 using Synopsys ICC2, where per benchmark, 100 placements are generated with different PPA objectives.In the experiments, we not only demonstrate that GAN-Place significantly improves vanilla DREAMPlace in single-design optimization but also show that it achieves superior optimization results on unseen designs that are not in the database.

Single-design Optimization Results on DREAMPlace [11]
In this experiment, we perform "single-design" optimization to demonstrate the considerable placement quality enhancements offered by the proposed GAN-Place framework by taking DREAMPlace [11] as a generator.The term "single-design" implies that the design that GAN-Place optimizes is same as the design in the target database.Our objective in this experiment is GAN-Place 32:13 GAN-Place optimizes the same design as in the database.
to establish the efficacy of employing generative adversarial learning for placement optimization.Subsequent sections will provide evidence of transfer learning experiments that further support our claims.Table 3 demonstrates the detailed optimization results, where we clearly observe that GAN-Place consistently outperforms vanilla DREAMPlace at each major PD stage across all four industrial and OpenCore benchmarks.Although the same design is used in the database during training, we still believe the achieved results are highly remarkable as GAN-Place is NOT using any "memorization" technique such as explicit net-matching, cell-alignment, and so on.The superior results are purely achieved by optimizing placement similarity scores via generative adversarial learning.Figure 9 further shows the bin-density map comparison between GAN-Place and ICC2 on the design CPU-2 that shows the largest PPA improvements.We observe that although the underlying placements are visually different, the achieved bin-density maps are arguably similar, which demonstrates the effectiveness of the proposed CNN-based discriminator.

Transfer Learning on Unseen Designs
Using DREAMPlace [11] and Xplace [13] Due to the intrinsic characteristics of the PD implementation process, wherein almost each realization necessitates a long runtime for completion in industrial design flows, the employment of transfer learning is the most favored approach in leveraging ML models to optimize critical PPA metrics.In this article, we refer transfer learning as the paradigm that aims to harness the knowledge acquired from training designs into unseen ones to perform PPA optimization.As  The "DREAMPlace+GAN" and "Xplace+GAN" columns are achieved by using different designs in the target database (i.e., each underlying design is unseen during GAN optimization).All metrics are reported by Synopsys ICC2.
aforementioned, the proposed GAN-Place framework does not require the target database contains the exact same design as the one being optimized by the GPU-accelerated placer.In the experiment, we demonstrate the enablement of transfer learning ability of GAN-Place using non-macro designs.Table 4 demonstrates our transfer learning optimization results.Note that each design is unseen during the optimization.The target database is built by other non-macro designs in the benchmarks as shown in Table 2.We clearly observe that the proposed GAN-Place framework GAN-Place 32:15 Fig. 10.Stage-by-stage wirelength and power metrics on AES.Note that each implementation starts from a different global placement, but all rely the exact same script to complete the full-chip design flow.
immediately improves each critical PPA metric at the global placement stage across every benchmark.Figure 10 further demonstrates the full-flow impact on wirelength and power, where we clearly observe that GAN-Place significantly improves not only the global placement metrics but also the post-route ones.

Discussion of Optimization Results
We believe the tremendous success of the proposed framework, GAN-Place, has three major implications: (1) "Placement style" can be transferred from one placer to another.This argument is supported by Table 3, which show that by using single-design optimization, GAN-Place can significantly improve the vanilla GPU placer on industrial benchmarks in consistent.(2) "Placement style" is more related to a placer itself than the designs being placed.This argument is validated by Table 4, where we observe that placement quality can not only be transferred in the same designs but also in completely different ones.(3) Without knowing the underlying algorithms or constraints, the "placement style" of a black-boxed placer can still be parameterized through generative adversarial learning.
Furthermore, in the experiments, we observe that GAN-Place not only improves the wirelength significantly but also introduces notable improvements in power.We think this is because by following the placement distribution of ICC2 in terms of cell locations, GAN-Place will introduce less buffers and logic fixing than the vanilla DREAMPlace during many optimization steps throughout the flow, which effectively results in less power consumption.Finally, we would like to emphasize that the runtime difference between the vanilla DREAMPlace and the proposed GAN-Place is no more than 10 minutes across all benchmarks, which is practically negligible compared with the global placement runtime of ICC2.

Why Use Generative Adversarial Learning to Advance VLSI Placement?
The placement problem in PD has been a central focus of extensive scientific investigation for several decades.A multitude of research groups have devoted considerable resources to this area, necessitating continuous innovation in methodological approaches.This persistent development is essential to address the emerging complexities inherent in design paradigms that arise as a consequence of advancing technology scaling.Over the years, heuristic-based and analytical placement methodologies have substantially advanced the VLSI placement domain; however, limited research has been conducted on the analysis of placement distributions across various placement algorithms and the subsequent utilization of such information for optimizing placement quality.In this study, we highlight the significance and applicability of such information for enhancing placement quality in an efficient manner, employing generative adversarial learning techniques.
Notably, the primary advantage of GAN-based learning lies in its ability to perform optimization without necessitating explicit objectives.This allows us to concentrate solely on the acquisition of high-quality designs for constructing the database.In our specific context, despite lacking knowledge of the proprietary algorithms employed by commercial tools, we successfully enhance the placement performance of open source placers to approach commercial-quality standards.We achieve this by guiding them to emulate the placement distribution in terms of netlist connectivity and bin-density maps, using the tool-verified, high-quality placements present in our database.

CONCLUSION
In this article, we have presented GAN-Place, which, as far as we know, is the first-ever learningdriven framework that improves an open source placer toward commercial-quality using generative adversarial learning.We showcase that GAN-Place can be easily integrated with state-ofthe-art open source GPU-accelerated placers DREAMPlace and Xplace to significantly improves the achieved placement quality.In the experiments, we demonstrate that GAN-Place not only improves the optimization results in single-design optimization but also facilitates transfer learning to perform optimization on unseen designs.The main assumption of this work is that "placement style (quality)" is born with a placer itself, which can be parameterized and transferred to another placer through generative adversarial learning.We have provided comprehensive experiments and analyses to prove this assumption empirically.We believe this work shall open the gate for future endeavor on advancing PD using generative adversarial learning.

Fig. 2 .
Fig. 2. Comparison of cell density maps between DREAMPlace and Synopsys ICC2 under the same global placement density target.It is observed that the commercial tool has extra intelligence on where to locally aggregate or spread out cells to optimize crucial PPA metrics while satisfying the global density constraints.

Fig. 3 .
Fig.3.Difference in placement objectives between a vanilla GPU-accelerated (i.e., differentiable) placer and GAN-Place at each placement iteration.The cell locations generated in GAN-Place are encouraged to follow the ones in commercial database by optimizing the similarity scores.

Fig. 5 .
Fig.5.GNN-based discriminator that outputs "fake/real" decision using node representation learning based on different input placement sources.

Fig. 6 .
Fig.6.Illustration of our CNN-based discriminator using the Soft-Bin transformation technique that transforms cell locations to differentiable bin-density maps.The goal of DREAMPlace is to generate placements that have similar bin-density maps as the ones in the database.

Fig. 7 .
Fig. 7. Detailed architecture of GAN-Place that leverages CNN-based and GNN-based discriminators to quantify placement similarity between different placement sources.

Fig. 8 .
Fig. 8. Illustration of integrating open source placers into an industrial design flow using Synopsys ICC2, where the global placement stage is replaced with proceeding stages remain the same.

Fig. 9 .
Fig. 9. Visualization of the proposed Soft-Bin technique in placements of ICC2 default placer, vanilla DREAM-Place, and the proposed GAN-Place on the AES benchmark.We observe GAN-Place produces a visually more similar bin-density map distribution to ICC2 than the vanilla DREAMPlace.

Table 1 .
Parameters We Leverage for Database Generation

Table 2 .
Eight Benchmarks Used in This Work and Their Attributes in TSMC 28 nm