A deep learning driven high-density standard cell global router

Abstract

The application discloses a deep learning driven high-density standard cell global router, and relates to integrated circuit design automation technology. The router comprises a preprocessing module, a model inference module, a parallel routing scheduling module and a post-processing module. The preprocessing module is used for extracting relevant features of a layout from an intermediate result of the high-density standard cell global router and constructing a heterogeneous hypergraph. The model inference module is used for inputting the heterogeneous hypergraph into a multi-modal congestion prediction model for prediction. The parallel routing scheduling module is used for scheduling the first maze routing stage to perform early routing on the predicted overflow network. The post-processing module is used for screening and replacing the routing result of the first maze routing stage, and submitting the updated three-dimensional congestion graph and overflow network to the second maze routing stage for final routing. The application solves the problems of long global routing running time, serious repeated maze routing and insufficient practicality of traditional congestion prediction model in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated circuit design automation, and more particularly, to a deep learning driven high-density standard cell global router. BACKGROUND

[0002] (I) Integrated circuit physical design and global routing process.

[0003] With the continuous evolution of process technology and the continuous pursuit of performance (Performance), power (Power) and area (Area) (PPA) of chip design, standard cell libraries under advanced processes increasingly provide high-density (High Density) standard cells to improve area utilization and circuit integration. In a typical digital circuit physical design process, the layout and routing stage usually includes clock tree synthesis layout, global routing and detailed routing, etc. Among them, global routing determines the rough path of each network (Net) routing on a coarse grid, generates a three-dimensional congestion map (3D Congestion Map) and overflow network (Overflow Net) information; detailed routing completes the final landing of line length, via and design rules on a finer grid. The quality of global routing directly determines the difficulty of subsequent detailed routing, and the upper limit of final routability and routing quality.

[0004] Existing mainstream academic global routers (such as CUGR2, FastGR, SPRoute2, GAMER, DGR, HELEM-GR, etc.) generally abstract the layout as a three-dimensional grid graph (3D Grid Graph), each grid cell is called GCell, and the connection between each GCell is GEdge. Each GEdge has capacity (Capacity) and demand (Demand) attributes. The pins (Pin) of the standard cell are mapped to the corresponding GCell, and the global routing algorithm needs to assign a path for all pins of each network on the grid graph. The congestion overflow of a single GEdge “e” is defined as:

[0005] overflow(e) = max(0, demand(e) capacity(e))。

[0006] ​In practical implementation, the aforementioned global router typically first splits the multi-terminal network into multiple two-terminal networks based on the Rectangular Steiner Minimum Tree (RSMT), and then sequentially employs L / U / Z pattern routing, monotonic routing, and three-fold bend routing (3... Algorithms such as Bend Routing and Maze Routing (which has higher accuracy but higher computational complexity) have been iteratively optimized. Figure 1 As shown. Taking CUGR2 as an example, its workflow sequence adopts: L-shaped routing (L... The routing process consists of three stages: Shape Pattern Routing (LPR), Augment Pattern Routing (APR), and Maze Routing (MR). Statistical results show that in high-density standard cell scenarios, the average single-net routing time is lowest in the LPR stage and highest in the MR stage, with a difference of nearly two orders of magnitude. Figure 2 As shown.

[0007] (ii) Challenges brought about by high-density standard cells.

[0008] In advanced process nodes, Process Design Kits (PDKs) often provide various standard cell libraries. High-density standard cell libraries, in particular, improve area utilization by reducing the number of routing tracks to compress cell height. However, fewer routing tracks mean a significant decrease in available routing resources, leading to more severe cabling congestion. In high-density standard cell scenarios, existing global routers often repeatedly route the same network across multiple stages, resulting in inefficiency. For example, when using 9... When tracking the high-density standard cell library, CUGR2 statistics show that approximately 40.4% of the networks that need routing in the APR stage will be forced to be rerouted in the MR stage; and of the networks that need routing in the MR stage, approximately 73.3% have already been routed in the APR stage. This indicates that a significant amount of runtime is wasted on repeatedly routing the same batch of "difficult-to-route networks" in a maze-like manner, and the overall runtime is severely hampered by a small number of overflowing networks.

[0009] Meanwhile, in high-density standard cell layouts, due to the large size of the netlist, complex geometry, and insufficient cabling resources, congestion is prone to occur simultaneously in multiple metal layers and in multiple directions (horizontal / vertical), posing a significant challenge to path selection for global cabling. Traditional analysis methods that only focus on single-layer or two-dimensional congestion prediction cannot accurately reflect three-dimensional congestion distribution, limiting their guiding value in practical routers.

[0010] (III) Existing congestion prediction-based auxiliary routing methods and their shortcomings.

[0011] To alleviate the inefficiency problem in the layout and routing stage, a large number of deep learning-based "Shift Left" methods have emerged in recent years, which predict congestion, timing and other indicators at an early stage of design to guide subsequent optimization. For example, MEDUSA, PROS and other methods use convolutional neural networks (CNN) to predict congestion distribution from layout geometric features, and increase the routing cost of GCell with potential congestion in the global routing cost function, so as to guide the router to bypass the potential congestion area; LayNet, TCAD'25 and other works combine graph neural networks (GNN) with Transformer models to build a multi-modal congestion prediction model that considers both geometric and topological information.

[0012] However, the existing methods generally have the following shortcomings:

[0013] 1. Most models only output two-dimensional or single-layer congestion heat maps, which are difficult to provide three-dimensional congestion distribution prediction for all metal layers, and have limited practicality.

[0014] 2. Only the routing cost is adjusted heuristically at the cost function level, still following the original serial workflow, and cannot parallelize the most time-consuming maze routing stage from the architecture level.

[0015] 3. Many multi-modal models use full connection with quadratic complexity in graph construction, resulting in excessive storage / computational overhead, making it difficult to implement on large-scale netlists.

[0016] 4. Most existing methods can only predict the congestion level of GCell, lack of fine-grained prediction of "which network will become an overflow network", and are difficult to directly drive parallel maze routing scheduling for specific networks. SUMMARY

[0017] The technical problem to be solved by the present application is to solve the problems of long running time of global routing, serious repeated maze routing and insufficient practicality of traditional congestion prediction model under high-density standard cell, and to provide a deep learning-driven global router for high-density standard cell.

[0018] The deep learning-driven global router for high-density standard cell according to the present application comprises:

[0019] The preprocessing module is used to extract layout geometric features and netlist topological features from the intermediate results of the global router for high-density standard cell, and to construct a heterogeneous hypergraph according to the layout geometric features and the netlist topological features.

[0020] a model inference module, configured to input the heterogeneous hypergraph into a multi-modal congestion prediction model to perform prediction, to obtain a predicted three-dimensional congestion graph and a predicted overflow network;

[0021] a parallel routing scheduling module, configured to, in an enhanced mode routing stage, schedule in parallel early routing of a first maze routing stage according to the predicted three-dimensional congestion graph and the predicted overflow network;

[0022] a post-processing module, configured to, after the enhanced mode routing stage ends, screen and replace the early routing result based on an actual routing result of the enhanced mode routing and an early routing result of the first maze routing stage, to obtain an updated three-dimensional congestion graph and overflow network, and submit the updated three-dimensional congestion graph and overflow network to a second maze routing stage for final routing;

[0023] a routing result output module, configured to output a final routing result.

[0024] Preferably, a specific process of the multi-modal congestion prediction model for predicting the heterogeneous hypergraph is as follows:

[0025] obtain a geometric feature tensor in the heterogeneous hypergraph;

[0026] perform block embedding and position coding on the geometric feature tensor to obtain a serialized Token feature;

[0027] input the geometric feature tensor into a convolutional neural network to construct a multi-scale pyramid feature;

[0028] input the serialized Token feature, the multi-scale pyramid feature and the heterogeneous hypergraph into a plurality of three-feature fusion modules connected in sequence for fusion processing, to obtain a fused graph node feature and an updated pyramid feature;

[0029] input the updated pyramid feature into a decoder to perform step-by-step upsampling and fusion on features of each layer, to obtain a predicted three-dimensional congestion graph;

[0030] input the fused graph node feature into a fully connected layer to obtain an overflow probability of each network, and perform a Sigmoid or thresholding operation to obtain a binary label as a predicted overflow network.

[0031] Preferably, a fusion method of the three-feature fusion module for the serialized Token feature, the multi-scale pyramid feature and the heterogeneous hypergraph is as follows:

[0032] input the heterogeneous hypergraph into a relational graph convolutional network to model topological structures and inter-network relationships in the heterogeneous hypergraph, and output an updated graph node feature;

[0033] The multi-scale pyramid features are enhanced by a multi-receptive field feature pyramid module, and then the enhanced multi-scale pyramid features and the updated graph node features are input into a CGI module for reconstruction processing.

[0034] The reconstructed multi-scale pyramid features and the serialized Token features are jointly modeled to obtain updated pyramid features.

[0035] Preferably, the reconstruction method of the CGI module for the multi-scale pyramid features is:

[0036] The multi-scale pyramid features are upsampled by transposed convolution and fused to the original resolution of the layout, and then the multi-scale pyramid features processed by transposed convolution are mapped to the graph node space to perform feature alignment and fusion with the updated graph node features, and then reconstructed back to the multi-scale pyramid features through several convolution layers and multiple skip connections.

[0037] Preferably, the relationship graph convolution network models the topological structure and inter-network relationship in the heterogeneous hypergraph through the following expression:

[0038] ,

[0039] In the formula, is the graph node feature of node i in the lth layer, is the updated graph node feature, ={r GC-GC , r Net-GC , r GC-Net} is a relationship set, is a neighbor node set under the relationship r, , are trainable parameters, c i,r is a normalization coefficient, and σ is a nonlinear activation function.

[0040] Preferably, in the post-processing module, the specific method for obtaining the final routing is:

[0041] First step, after the end of the enhanced mode routing stage, obtaining the actual routing result, traversing all routing grids of the grid graph in the actual routing result, and counting a real overflow network set based on the actual routing result;

[0042] Second step, for each network in the real overflow network set, if a feasible maze routing result corresponding to the network has been generated in the early maze routing stage, the path of the network in the grid graph is replaced by the result generated in the first maze routing stage.

[0043] Third step, re-count the demand and overflow of each GEdge after replacing the grid map, get the updated three-dimensional congestion map and overflow network;

[0044] Fourth step, submit the updated three-dimensional congestion map and overflow network to the second maze routing stage, perform maze routing to obtain the final routing.

[0045] Beneficial effects

[0046] The advantages of the present application are:

[0047] 1. On the premise of maintaining or improving the overflow and routability indicators, significantly reduce the total running time of global routing, especially effectively accelerate the maze routing stage;

[0048] 2. Through the multi-modal feature fusion model, output the three-dimensional congestion map of multiple metal layers and the network level overflow network prediction simultaneously, realize the joint modeling from "GCell level geometric congestion" to "Net level topological congestion";

[0049] 3. Based on the overflow network prediction result, transform the traditional serial pattern routing (PR) and maze routing (MR) stages into a partially parallel execution workflow, so that a part of the network predicted to overflow can be routed in high precision in the PR stage in advance, thereby reducing the number of networks that need to be rerouted subsequently;

[0050] 4. Through the graph pruning based on depth-first search (DFS) and the netlist DAG (directed acyclic graph) retention strategy, effectively reduce the model input size and improve the scalability on large-scale high-density standard cell circuits. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a schematic diagram of the traditional workflow of the mainstream global router in the prior art;

[0052] Figure 2 It is a statistical diagram of the running time and rerouting of each stage of CUGR2 in a high-density standard cell scenario;

[0053] Figure 3 It is a schematic diagram of the graph representation and E CC pruning of the present application based on heterogeneous hypergraph;

[0054] Figure 4 It is a schematic diagram of the workflow comparison between HDGR and CUGR2 of the present application;

[0055] Figure 5The overall structure schematic diagram of the multi-modal congestion prediction model of the application; in Figure 5 The lower left corner is the TFBlock structure schematic diagram of the application; the lower right corner is the CGI structure schematic diagram of the application;

[0056] Figure 6 The second metal layer congestion prediction visualization contrast schematic diagram of a typical reference circuit;

[0057] Figure 7 The running time and parameter quantity contrast diagram under different models;

[0058] Figure 8 The E CC Pruning algorithm pseudocode schematic diagram. DETAILED DESCRIPTION

[0059] The application will be further described below in conjunction with the embodiments, but does not constitute any limitation on the application, and any limited number of modifications made by anyone within the scope of the claims of the application is still within the scope of the claims of the application.

[0060] The deep learning driven high-density standard cell global router of the application comprises a preprocessing module, a model inference module, a parallel routing scheduling module, a post-processing module and a routing result output module. The specific flowchart is shown in Figure 4 In Figure 4 , the solid line part is the work flow of the high-density standard cell global router of the application, and the dashed line is the work flow of CUGR2. From Figure 4 , it can be seen that in the traditional CUGR2 work flow, APR and MR are executed in series; and the application calls a multi-modal congestion prediction model after the end of the LPR stage, obtains a predicted three-dimensional congestion map and a predicted overflow network based on the input of the APR stage, and then performs post-processing, which can effectively reduce the number of networks that need to be rewired subsequently.

[0061] Specifically, the preprocessing module is used to extract layout geometric features and netlist topological features from the intermediate results of the high-density standard cell global router (such as CUGR2), and construct a heterogeneous hypergraph according to the layout geometric features and the netlist topological features.

[0062] Specifically, the specific implementation process of the preprocessing module is as follows:

[0063] A. Grid graph abstraction:

[0064] The layout in the intermediate results of the high-density standard cell global router is abstracted as a three-dimensional grid graph with a height of H, a width of W and a number of metal layers of L, and the layout geometric features V GC , the number of which is N GC= H x W; each GEdge corresponds to the connection between GCells and has capacity capacity(e) and demand demand(e), and the network table topology characteristics are obtained: V Net , CC , CN , NC .

[0065] wherein (1) V GC is a set of GCell nodes, and the feature vector of each GCell node includes:

[0066] a. Pin number feature: the number of pins contained in the GCell;

[0067] b. Capacity feature: the available routing capacity of each metal layer;

[0068] c. LPR post-demand feature (Post LPR Demand): demand distribution after the end of the L-shaped pattern routing stage;

[0069] d. APR pre-demand feature (Pre APR Demand): demand distribution based on Post LPR Demand, after removing the demand corresponding to the overflow network identified in the LPR stage;

[0070] e. Overflow count feature: the number of overflow networks on each GEdge after the LPR stage.

[0071] (2) V Net is a set of network nodes, and the feature of each network node includes:

[0072] a. Height, width and area of the network bounding box;

[0073] b. Number of E CC edges after pruning inside the network;

[0074] c. Network routing order in the LPR stage (Net Order);

[0075] d. Network routing order in the APR stage (Net Order, set to 1 and normalized if there is no overflow in the APR).

[0076] (3) E CC is a set of GCell GCell edges, used to describe the topological connection relationship between GCells in the same network.

[0077] Traditional methods usually connect all GCells within the same network pairwise, resulting in E CC The number of edges grows in a quadratic order, which not only increases the memory and computation overhead, but also aggravates the Over Smoothing problem in graph convolution. The present application provides a DFS-based E CC Pruning strategy: based on the Steiner tree structure inside the router and the directed acyclic routing order, using the depth-first search (DFS) algorithm, starting from the Steiner tree root node and the Access Points, only keeping the GCell edges that have a substantial impact on congestion along the actual routing path direction GCell edges, while adding a self-loop (Self Loop) to the GCell nodes that are not connected to any network, significantly reducing the number of edges while ensuring the integrity of the netlist DAG topology. The specific pruning algorithm pseudocode is shown in Figure 8 .

[0078] (4) E CN : GCell Net edge set, used to describe the association between GCell and its belonging network.

[0079] (5) E NC : Net GCell edge set, which is the reverse edge of E CN .

[0080] B. Constructing a heterogeneous hypergraph:

[0081] Constructing a heterogeneous hypergraph based on V GC , V Net , E CC , E CN , E NC : G h = (V GC , V Net , E CC , E CN , E NC ).

[0082] The model inference module is used to input the heterogeneous hypergraph into the multi-modal congestion prediction model for prediction, obtaining the prediction results of the three-dimensional congestion graph and the prediction results of the overflow network.

[0083] The specific prediction process of the multi-modal congestion prediction model will be introduced below.

[0084] In this embodiment, the input and output forms of the deep learning model used by the multi-modal congestion prediction model can be represented as:

[0085] .

[0086] Among them, X geom ∈R H×W×C For V GC The resulting geometric feature tensor, G h This is a heterogeneous hypergraph. Furthermore, the model's output includes:

[0087] (1) : Predicted 3D congestion map, showing the congestion level of each GCell and each metal layer;

[0088] (2) The network-level overflow network prediction result is a binary classification label.

[0089] In this embodiment, the multimodal congestion prediction model mainly includes four sequentially connected three-feature fusion modules (TFBlock) and an UpperNet-based decoder. The following will describe the process of the multimodal congestion prediction model in predicting heterogeneous hypergraphs, based on the above model framework and using one of the three-feature fusion modules as an example.

[0090] like Figure 5 As shown, the specific process of the multimodal congestion prediction model is as follows:

[0091] (1) Transformer Token branch:

[0092] For X geom After performing patch embedding and position encoding, a serialized token feature P is obtained, which serves as the input to the Vision Transformer (ViT) backbone to capture long-range dependencies globally. This branch is mainly responsible for global information modeling and multi-head self-attention feature extraction in two-dimensional space.

[0093] (2) Geometric features and image branches:

[0094] X geom The input is fed into a convolutional neural network to construct a multi-scale pyramid feature, and then processed through a multi-receptive-field feature pyramid (Multi... The Receptive Field Feature Pyramid (MRFP) module constructs multi-scale pyramid features. The MRFP module enhances the ability to model long-range geometric information. MRFP reduces dimensionality through fully connected layers, reconstructs image features by scale partitioning, then performs depthwise separable convolutions, re-tokenizes the features, and restores the channel dimension through fully connected layers, thereby expanding the effective receptive field while maintaining computational efficiency.

[0095] (3) Graph Neural Network Branch:

[0096] heterogeneous hypergraph G h Chinese V GC With V Net The input is fed into a Relational Graph Convolution Network (RGCN). RGCN is used in GCell. GCell, Net GCell and GCell The three types of relationship subgraphs in the Net share / distinguish parameters, enabling efficient modeling of the topology and relationships between networks. It outputs updated graph node features for subsequent bidirectional interaction between CNN-GNN and CNN-Transformer. The formula for the RGCN module can be expressed as:

[0097] ,

[0098] in, Let i be the graph node feature of node i at layer l. = {r GC-GC , r Net-GC , r GC-Net} represents a set of relations. Let r be the set of neighboring nodes. , For trainable parameters, c i,r σ is the normalization coefficient, and σ is the nonlinear activation function. High-dimensional node representations that simultaneously encode geometric, topological, and wiring order information can be obtained through multi-layer stacking.

[0099] Among them, the updated graph node features output by the last-level three-feature fusion module are used as the fused graph node features.

[0100] (3) CNN–GNN bidirectional fusion module (CGI):

[0101] The CNN-GNN Bidirectional Fusion Interaction (CGI) module first up-samples and fuses the multi-scale pyramid features to the original resolution HxW through transposed convolution, and then maps the feature to the graph node space, aligns and fuses the updated graph node features output by the RGCN, and then reconstructs the multi-scale pyramid features {F3, F4, F5} through several convolution layers and multiple skip connections to realize the bidirectional flow of geometric and topological information between the pixel / grid level and the network level.

[0102] (4) CNN-Transformer Bidirectional Fusion Interaction (CTI):

[0103] The CNN-Transformer Bidirectional Fusion Interaction (CTI) module uses a multi-scale deformable attention (MSDA) mechanism to jointly model the multi-scale pyramid features {F3, F4, F5} from the CGI module and the serialized Token feature P. After normalization (LayerNorm) and feedforward network (FFN), the CTI adjusts the fusion strength between the Transformer output and the CNN feature through a learnable scaling factor a to obtain the updated Token feature P' and the pyramid feature F', thereby compensating for the lack of hierarchical structure in the pure ViT model and strengthening the ability of CNN in long-distance modeling. The specific expression is:

[0104]

[0105] In the above expression, FFN(·) represents the feedforward network, MSDA(·) represents the multi-scale deformable attention mechanism, and NORM(·) represents the layer normalization. The learnable scalar a is used to promote training stability. P i and F i ' represent the updated Transformer token and CNN feature at level i, respectively.

[0106] Finally, the updated pyramid features are input into the UperNet-based decoder to perform level-by-level up-sampling and fusion of each layer feature, and finally output a three-dimensional congestion map CongM pred with a size of HxWxL, realizing congestion map prediction. The specific expression is:

[0107] ,

[0108] In the above expression, B h,w ​CongM (h, w) is the congestion prediction average value of the corresponding coordinates. As shown in pred (h, w) is the congestion prediction average value of the corresponding coordinates. As shown in Figure 6 Fig. 2 shows a typical benchmark circuit second metal layer congestion prediction visualization comparison diagram.

[0109] The fused graph node features pass through a fully connected layer to obtain the overflow probability of each network, and a Sigmoid or thresholding operation is performed to obtain the predicted overflow network, thereby realizing overflow network prediction. The specific expression is:

[0110] .

[0111] In this embodiment, the parallel routing scheduling module is used to schedule the first maze routing stage (MR1) in advance based on the predicted three-dimensional congestion graph and the predicted overflow network in the enhanced mode routing (APR) stage. In the enhanced mode routing stage (APR), the first maze routing stage (MR1) is scheduled for parallel routing while the conventional network continues to use mode routing / enhanced routing. APR and MR1 run simultaneously on different threads and share / synchronize the partial updates of the grid graph demand.

[0112] The post-processing module is used to filter and replace the routing results of MR1 based on the actual routing results of the enhanced mode routing and the early routing results of the first maze routing stage after the APR stage ends, and finally obtain the remaining predicted three-dimensional congestion graph and overflow network. This part of the three-dimensional congestion graph and the overflow network is the updated three-dimensional congestion graph and the overflow network. These updated three-dimensional congestion graphs and overflow networks are submitted to the second maze routing stage (MR2) for final routing.

[0113] In this embodiment, the specific steps of the post-processing module are as follows:

[0114] Step 1, terminate the further expansion of MR1, traverse all routing grids in the grid graph obtained by the enhanced mode routing stage, and count the real overflow network set OFNet based on the actual results of APR APR .

[0115] Step 2, for each network in OFNet APR , if MR1 has output a feasible maze routing result for the network, replace its path in the grid graph with the result of MR1.

[0116] Step 3, re-count the demand and overflow of each GEdge for the grid graph after replacement to obtain the updated three-dimensional congestion graph and overflow network.

[0117] Step 4: Submit the updated 3D congestion graph and overflow network to the second maze routing stage (MR2) to perform the final fine maze routing, ensuring that overflow is eliminated as much as possible and the line length and via number are optimized while meeting the design rules and constraints.

[0118] The routing result output module is used to output the final routing result, the predicted 3D congestion map, and related statistical indicators (such as overflow quantity, line length, number of vias, running time, etc.).

[0119] The high-density standard cell global router of the present invention will be described below with reference to specific experiments.

[0120] The experimental environment for the high-density standard cell global router of the present invention is shown in Table 1 below:

[0121]

[0122] (1) Based on the above experimental environment, this invention was compared with existing methods MEDUSA, LayNet, and PROS on 10 high-density reference circuits (i8t6–i8t10, i9t6–i9t10) of ICCAD'19, such as Figure 7 As shown in Table 2 below:

[0123]

[0124] The results show that:

[0125] a. On SSIM and NRMS, the present invention achieves an average improvement of approximately 3% and 12%, respectively, indicating that the congestion graph prediction is closer to the actual APR results;

[0126] b. In terms of Precision and F1, the present invention improves by an average of about 0.1% and 19%, respectively, and is significantly better than existing models in network-level overflow prediction, which helps to reduce invalid MR1 scheduling and missed detection of critical overflow networks.

[0127] The final experimental data for the ablation experiment are shown in Table 3 below:

[0128]

[0129] (2) In the ablation experiments, the present invention verified the importance of the following key designs:

[0130] a. When removing geometric features carrying L-shaped wiring information, NRMS increases sharply from 0.50 to 0.91, and F1 decreases from 0.67 to 0.19, indicating that the geometric congestion features after LPR are crucial for three-dimensional congestion prediction.

[0131] b. When the L-shaped congestion feature is replaced with the traditional RUDY feature, NRMS drops to 0.63 and F1 rises to 0.55, but it is still not as good as the full configuration, proving that the geometric features based on the real LPR cabling can better reflect the actual congestion distribution in high-density scenarios.

[0132] c. Remove LPR / APR Net When the Order feature is applied, both NRMS and F1 deteriorate, verifying the importance of wiring order as a topology "location encoding" in the model;

[0133] d. If LayNet's graph structure G is used directly h,Lay-Net Due to its E CC The number of edges results in a quadratic complexity, leading to significant memory overflows (OOM) during training, thus demonstrating the E-factor of this invention from the opposite perspective. CC Pruning strategies are necessary and effective for deploying multimodal models in large-scale, high-density designs.

[0134] (3) The experimental results of the global router performance test are shown in Table 4 below:

[0135]

[0136] Experiments show that on the public dataset of the ICCAD'19 competition, the High-Density Cell Global Routing (HDGR) of this invention achieves an overall runtime reduction of approximately 16.4% and a spillover reduction of approximately 12.5%, while maintaining or even slightly improving trace length and via metrics on most benchmarks. Trace length (WL) and via count (#Via): Overall, it is roughly the same as or slightly improved compared to CUGR2. In some ultra-high density scenarios (such as i8t10), due to the extremely high device density leading to large demand spillover networks after APR, maze routing will transfer demand from heavily overflowing GCcells to adjacent areas, thus introducing more local overflow, but without affecting the overall trend.

[0137] Fine-grained analysis shows that the acceleration of HDGR mainly comes from Post The remaining overflow network number after processing is "OFNet" Post "Significantly smaller than the original overflow network number after APR" OFNet APR This means that a large number of difficult networks have already completed high-precision cabling in the MR1 stage, and MR2 only needs to handle the remaining small number of overflow networks, which significantly reduces the number of repeated MR calls.

[0138] The above merely describes the preferred embodiments of the present application, and it should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present application, and these will not affect the effect of the present application and the practicality of the patent.

Claims

1. A deep learning-driven high-density standard cell global router, characterized in that, include: The preprocessing module is used to extract layout geometric features and netlist topology features from the intermediate results of the high-density standard cell global router, and to construct a heterogeneous hypergraph based on the layout geometric features and netlist topology features. The model inference module is used to input the heterogeneous hypergraph into the multimodal congestion prediction model for prediction, and obtain the predicted 3D congestion graph and the predicted overflow network. The parallel routing scheduling module is used to perform advance routing in the first maze routing stage in parallel scheduling based on the predicted three-dimensional congestion map and the predicted overflow network during the enhanced mode routing stage. The post-processing module is used to filter and replace the advance routing results based on the actual routing results of the enhanced mode routing and the advance routing results of the first maze routing stage after the enhanced mode routing stage is completed, obtain the updated three-dimensional congestion map and overflow network, and submit the updated three-dimensional congestion map and overflow network to the second maze routing stage for final routing. The routing result output module is used to output the final wiring result; The specific process of using the multimodal congestion prediction model to predict heterogeneous hypergraphs is as follows: Obtain the geometric feature tensor of the heterogeneous hypergraph; Block embedding and position encoding are performed on the geometric feature tensor to obtain serialized token features; The geometric feature tensor is input into a convolutional neural network to construct a multi-scale pyramid feature; The serialized token features, multi-scale pyramid features, and heterogeneous hypergraph are input into multiple sequentially connected three-feature fusion modules for fusion processing to obtain fused graph node features and updated pyramid features. The updated pyramid features are input into the decoder to perform upsampling and fusion of features at each layer to obtain the predicted three-dimensional congestion map. The fused graph node features are input into a fully connected layer to obtain the overflow probability of each network, and binary labels are obtained through Sigmoid or thresholding operations as the predicted overflow network.

2. The deep learning-driven high-density standard cell global router according to claim 1, characterized in that, The three-feature fusion module uses the following method to fuse serialized token features, multi-scale pyramid features, and heterogeneous hypergraphs: The heterogeneous hypergraph is input into a relational graph convolutional network to model the topology and relationships between networks in the heterogeneous hypergraph and output updated graph node features. After the multi-scale pyramid features are enhanced by the multi-receptive field feature pyramid module, the enhanced multi-scale pyramid features and the updated graph node features are then input into the CGI module for reconstruction. The reconstructed multi-scale pyramid features are jointly modeled with the serialized token features to obtain the updated pyramid features.

3. The deep learning-driven high-density standard cell global router according to claim 2, characterized in that, The CGI module reconstructs multi-scale pyramid features using the following method: The multi-scale pyramid features are upsampled and fused to the original resolution of the map by transposed convolution. Then, the multi-scale pyramid features processed by transposed convolution are mapped to the graph node space, so that the mapped multi-scale pyramid features processed by transposed convolution are aligned and fused with the updated graph node features. Finally, the multi-scale pyramid features are reconstructed through several convolutional layers and multiple skip connections.

4. A deep learning-driven high-density standard cell global router according to claim 2, characterized in that, Relational graph convolutional networks model the topology and relationships between networks in heterogeneous hypergraphs, achieved through the following expression: ， In the formula, Let i be the graph node feature of node i at layer l. This represents the updated graph node features. ={r GC-GC ,r Net-GC , r GC-Net } represents a set of relations. Let r be the set of neighboring nodes. , All are trainable parameters, c i,r σ is the normalization coefficient, and σ is the nonlinear activation function.

5. A deep learning-driven high-density standard cell global router according to claim 1, characterized in that, In the post-processing module, the specific method for obtaining the final wiring is as follows: Step 1: After the enhanced mode routing phase is completed, obtain the actual routing result, traverse all the routing meshes in the mesh graph of the actual routing result, and count the set of real overflow networks based on the actual routing result. The second step is to replace the path of each network in the real overflow network set with the result produced in the early wiring result of the first maze wiring stage. The third step is to recalculate the demand and overflow status of each GEdge in the replaced mesh graph to obtain the updated 3D congestion graph and overflow network. Step 4: Submit the updated 3D congestion graph and overflow network to the second maze wiring stage, perform maze wiring, and obtain the final wiring.

Images