Integrated circuit congestion prediction method and apparatus, and electronic device

Abstract

The invention relates to the technical field of integrated circuits, and discloses an integrated circuit congestion prediction method and device, and electronic equipment, and the method comprises the steps: extracting at least two congestion prediction feature maps of a to-be-detected integrated circuit; performing rasterization and normalization processing on each congestion prediction feature graph to obtain a corresponding two-dimensional pixel matrix graph; and inputting the two-dimensional pixel matrix graph into a congestion prediction model to obtain a congestion hotspot graph. The technical problems of low congestion model prediction precision, huge model and long time consumption in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated circuits, and in particular to an integrated circuit congestion prediction method and device and electronic equipment. BACKGROUND

[0002] As the core of modern electronic devices, the design of integrated circuits (ICs) highly depends on electronic design automation (EDA) tools. Compared with analog ICs, the market size of digital ICs is larger, and the dependence on EDA tools is more significant, especially in the logic synthesis and layout stages. The efficiency and accuracy of EDA tools become the key to the success of design. However, with the continuous evolution of process nodes and the continuous improvement of circuit complexity, the amount of layout data increases dramatically. Most modern design processes need to go through multiple design iterations through a full place and route (PnR) process, which is obviously expensive for large-scale designs. To address these challenges, machine learning (ML) based layout routability prediction methods have emerged, which use early physical design stage feature information to efficiently predict layout congestion, identify potential design problems in advance, optimize the design process, and reduce repeated iterations, thereby greatly improving design efficiency and having high application value.

[0003] The academic and industrial communities have extensively explored the application of machine learning techniques in congestion modeling to replace traditional time-consuming routing engines. Current research mainly falls into two categories: topology-based methods and geometry-based methods. Topology-based methods focus on the logical relationship of circuits and usually use graph neural networks (GNNs) to capture the complex interaction between units to improve prediction accuracy. Geometry-based methods focus on the spatial information of circuits and use handcrafted features to rasterize circuits, converting them into images, and then use CNNs or GANs to extract features from the circuits. Some research has also introduced Vision Transformers (ViTs) into congestion prediction. In the process of implementing the present application, the inventors have found at least the following problems in the prior art: CNN models have difficulty effectively extracting global features due to the local nature of convolution kernels, and even with the introduction of attention mechanisms, there are still limitations.

[0004] Secondly, graph neural networks (GNNs) have an over-smoothing problem, which limits their ability to model long-distance information. At the same time, chip designs are large in scale, with node numbers often reaching hundreds of thousands and edge numbers reaching millions, making the process of constructing a graph structure very time-consuming.

[0005] In addition, traditional ViT models have difficulty adapting to multi-scale scenarios due to their fixed token receptive fields, and their self-attention calculations have high complexity. SUMMARY

[0006] The embodiment of the present application aims to provide an integrated circuit congestion prediction method and device, and an electronic device, so as to solve the technical problems of low prediction accuracy, large model and long time consumption of the congestion model in the related art.

[0007] According to a first aspect of the embodiment of the present application, an integrated circuit congestion prediction method is provided, comprising: extracting at least two congestion prediction feature maps of the integrated circuit to be tested; performing rasterization and normalization processing on each of the above congestion prediction feature maps respectively to obtain a corresponding two-dimensional pixel matrix map; inputting the two-dimensional pixel matrix map into a congestion prediction model to obtain a congestion hotspot map; The congestion prediction model comprises an encoding module, a decoding module and a double upsampling module. The encoding module is configured to fuse at least two feature maps and generate a first feature matrix through unit segmentation and linear embedding processing, and then input the first feature matrix into a cascaded Swin Transformer module for hierarchical feature encoding, and each module outputs an encoded congestion feature matrix of the corresponding level. The decoding module is configured to receive the first feature matrix and the encoded congestion feature matrix of each level, first perform smoothing and feature upsampling processing on the congestion feature matrix output by the bottommost Swin Transformer module in turn, then perform element-by-element superposition with the congestion feature matrix output by the upper Swin Transformer module, and iteratively in this way until the first feature matrix is fused, and finally output a second feature matrix. The double upsampling module is configured to perform linear interpolation upsampling and pixel-level feature restoration on the second feature matrix to output a congestion hotspot distribution map in the horizontal and vertical directions.

[0008] According to a second aspect of the embodiment of the present application, an integrated circuit congestion prediction device is provided, comprising: The extraction module is configured to extract at least two congestion prediction feature maps of the integrated circuit to be tested. The processing module is configured to perform rasterization and normalization processing on each of the above congestion prediction feature maps respectively to obtain a corresponding two-dimensional pixel matrix map. The prediction module is configured to input the two-dimensional pixel matrix map into a congestion prediction model to obtain a congestion hotspot map. The congestion prediction model comprises an encoding module, a decoding module and a double upsampling module. The encoding module is configured to fuse at least two feature map features, and generate a first feature matrix through unit segmentation and linear embedding processing, and then input the first feature matrix into a cascaded Swin Transformer module to perform hierarchical feature encoding, and each module outputs an encoded congestion feature matrix of a corresponding level; The decoding module is configured to receive the first feature matrix and the encoded congestion feature matrices of each level, perform smoothing and feature upsampling processing on the congestion feature matrix output by the bottommost Swin Transformer module in sequence, and then perform element-by-element superposition with the congestion feature matrix output by the upper Swin Transformer module, and iteratively in sequence until the first feature matrix is fused, and finally output a second feature matrix. The double upsampling module is configured to perform linear interpolation upsampling and pixel-level feature recovery on the second feature matrix, and output congestion hotspot distribution maps in the horizontal direction and the vertical direction.

[0009] According to a third aspect of the embodiments of the present application, an electronic device is provided, including: one or more processors; a memory configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to the first aspect.

[0010] The technical solutions provided by the embodiments of the present application can include the following beneficial effects: As can be seen from the above embodiments, the Swin Transformer structure is adopted to introduce hierarchical window self-attention mechanism and shift window strategy, so that the model can efficiently capture local and global features at different scales, overcoming the problem of insufficient long-range dependency modeling ability caused by fixed convolution kernel in CNN, while avoiding the high computational complexity of traditional Transformer. Further, the expression ability of the model for image spatial hierarchical features is significantly improved, and higher accuracy is achieved in congestion prediction and other tasks.

[0011] The feature pyramid structure is adopted to fuse deep semantic information and shallow detail features through top-down path and horizontal connection, overcoming the problem of insufficient semantic information of shallow features and low spatial resolution of deep features in ordinary hierarchical structure, and significantly improving the detection ability of the model for multi-scale targets.

[0012] The complex Swin Transformer module of the decoding part is replaced by a lightweight feature pyramid (FPN) convolution structure, which overcomes the problem of high computational complexity and large parameter quantity of the Transformer decoder, while retaining the multi-scale feature fusion capability. On the premise of maintaining the prediction accuracy of the model, the model volume is greatly reduced, and the prediction speed is greatly improved.

[0013] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0014] The accompanying drawings, which are incorporated into the specification and constitute part of it, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0015] Figure 1 is a flowchart of an integrated circuit congestion prediction method according to an exemplary embodiment.

[0016] Figure 2 is a simple integrated circuit layout diagram composed of three networks according to an exemplary embodiment.

[0017] Figure 3 is a four-congestion-feature map of a prediction model according to an exemplary embodiment.

[0018] Figure 4 is a block diagram of a congestion prediction model according to an exemplary embodiment.

[0019] Figure 5 is a Swin Transformer module and Swin Transformer layer diagram according to an exemplary embodiment.

[0020] Figure 6 is a diagram of the double upsampling module according to an exemplary embodiment.

[0021] Figure 7 is a prediction result diagram of the model according to an exemplary embodiment.

[0022] Figure 8 is a feature pyramid structure diagram according to an exemplary embodiment.

[0023] Figure 9 is a structure diagram of an integrated circuit congestion prediction device according to an exemplary embodiment.

[0024] Figure 10 is a diagram of different numbers of feature map prediction indicators. DETAILED DESCRIPTION

[0025] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, like reference numerals refer to like elements, unless the context clearly dictates otherwise. The following description of exemplary embodiments is not intended to represent all embodiments in accordance with the present application. Rather, they are merely examples in accordance with some aspects of the present application as detailed in the appended claims.

[0026] The terminology used in the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the description of the application and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It also will be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0027] It is to be understood that although the terms first, second, third, etc. can be used herein to describe various information, but these information should not be limited to these terms. These terms are only used to distinguish one piece of information from another piece of information of the same type. For example, without departing from the scope of the present application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon determination" or "in response to determining".

[0028] Embodiment 1: Figure 1 is a flow chart of an integrated circuit congestion prediction method according to an exemplary embodiment, as shown in Figure 1 the method can include the following steps: S1: extracting at least two congestion prediction feature maps of the integrated circuit to be tested; Specifically, Figure 2 is a simple integrated circuit layout diagram composed of three networks according to an exemplary embodiment. In the layout planning stage, the macro cell distribution feature map is extracted; in the layout stage, the following three feature maps are extracted simultaneously: rectangular region line density distribution feature map, rectangular region pin line density distribution feature map and cell density distribution feature map.

[0029] This embodiment selects the four congestion prediction feature maps, i.e. the macro cell distribution feature map, the rectangular region line density distribution feature map, the rectangular region pin line density distribution feature map and the cell density distribution feature map.

[0030] S2: respectively rasterizing and normalizing each of the above congestion prediction feature maps to obtain the corresponding two-dimensional pixel matrix map; Specifically, the above four feature maps are rasterized, and the feature values of each feature map are represented by the pixel values of the corresponding pixel points; the four rasterized feature maps are converted into four two-dimensional pixel matrix maps with the same resolution, as shown in Figure 3 .

[0031] For the rectangular area line density distribution feature map, the feature value calculation formula is as follows: ; Wherein, represents the rectangular area line density feature value corresponding to the kth network at the (i, j) plane coordinate, wherein is the difference between the maximum coordinate and the minimum coordinate of all nodes in the network k in the horizontal direction, is the difference between the maximum coordinate and the minimum coordinate in the vertical direction; the network k is a unit connected in series from the chip input port to the output port; represents the ratio of the overlapping area of the network k and the (i, j) pixel block to the area of the pixel block; The specific calculation formula of the above-mentioned is as follows: ; ; ; In the formula, represents the area of the (i, j) pixel block, the overlapping area of the (i, j) pixel block and the kth network, , respectively represent the maximum value and the minimum value of the horizontal coordinate of the (i, j) pixel block, , respectively represent the maximum value and the minimum value of the vertical coordinate of the (i, j) pixel block; , , , respectively represent the maximum value and the minimum value of the horizontal direction of the kth network, and the maximum value and the minimum value of the vertical direction.

[0032] The above-mentioned effectively balances the non-overlapping of the edges of the network k and the edges of the pixel block, as shown in Figure X , effectively alleviates the network edge distortion, and improves the accuracy of congestion prediction.

[0033] For the rectangular area pin line density distribution feature map, the feature value calculation formula is as follows: wherein, represents a pin on net, represents This pin is in the pixel block (i, j). and The difference between the maximum value and the minimum value of the horizontal coordinate and the difference between the maximum value and the minimum value of the vertical coordinate of all pins on network k are respectively represented by and The calculation method is similar, respectively representing the difference between the maximum value and the minimum value of the horizontal coordinate and the difference between the maximum value and the minimum value of the vertical coordinate of all pins on network k.

[0034] For the macro unit distribution feature map, the feature value calculation formula is as follows: If there is a macro unit distribution in the corresponding pixel block, the weight of the pixel block is set to 1, otherwise, it is set to 0 For the unit density feature map, the feature value of the corresponding pixel block is equal to the number of units in the pixel block. The units include standard units of macro units.

[0035] S3: inputting the two-dimensional pixel matrix map into the congestion prediction model to obtain a congestion hotspot map; Specifically, referring to Figure 4 , the congestion prediction model comprises an encoding module, a decoding module and a double upsampling module; The encoding module is used for feature fusion of the unit distribution feature map, the rectangular area line density distribution feature map, the rectangular area pin line density distribution feature map and the unit density distribution feature map, and generates a first feature matrix through unit segmentation and linear embedding processing, and then inputs the first feature matrix into a plurality of Swin Transformer modules in cascade for hierarchical feature encoding. Each Swin Transformer module has a plurality of Swin Transformer layers connected in series, as shown in Figure 5 The encoding congestion feature matrix of each level is output by each module; The decoding module is used for receiving the first feature matrix and the encoding congestion feature matrix of each level. The congestion feature matrix output by the bottommost Swin Transformer module is sequentially smoothed and feature up-sampled, and then is element-wise superimposed with the congestion feature matrix output by the last Swin Transformer module. This is iterated upwards until the first feature matrix is fused, and finally a second feature matrix is output.

[0036] The double upsampling module is as shown in Figure 6As shown, for linear interpolation upsampling and pixel-level feature recovery of the second feature matrix, output horizontal and vertical congestion hotspot distribution maps. Specifically, the second feature matrix is input into a linear difference value module and a pixel-level feature recovery module respectively to obtain corresponding congestion features, which are directly spliced in the channel dimension, and the spliced features are output to the output layer, and the output layer outputs 2 channels corresponding to the horizontal and vertical congestion maps respectively We use the method of transfer learning to train the model, first pre-training on CircuitNetN28, to ensure the fairness of comparison and reflect the advancement of the model, then fine-tuning on ISPD2015 dataset, as shown in Table 1, which performs significantly better than RouteNet and GAN, score improves by 72.2% and 19.3%, compared with Lay-Net, the model improves by 11.0%, 10.8% and 51.6% in SSIM, NRMS and score respectively. Further, the congestion of horizontal and output directions is combined, and the prediction result of the model is shown in Figure 7 As shown in Table 2, our model reduces the NRMS by 20.3% compared with GAN, and the model parameter amount is only one fifth of GAN. We judge the unit area with a normalized congestion value greater than 0.125 as a hotspot area. In the binary classification task, the ROC index improves by 0.8% / 7% / 2.7%, and the PSNR is improved by more than 0.4dB compared with GAN.

[0037] Table 1 Table 2 The congestion prediction model adopts a feature pyramid, as shown in Figure 8 As shown, the encoding module includes a linear embedding submodule and a plurality of Swin Transformer modules connected in series, and the decoding module includes a plurality of decoding submodules connected in series, each decoding submodule includes a smoothing processing layer and a superimposed upsampling module, and the feature information output by the linear embedding submodule and the plurality of Swin Transformer modules is cross-layer connected into the decoding submodule to recover the congestion feature.

[0038] In an embodiment, the smoothing processing is performed by a nn.conv convolution function to change the number of channels of the congestion feature matrix information obtained by the next layer, which is the same as the cross-connected congestion feature matrix information.

[0039] In an embodiment, the encoding module and / or the decoding module further comprises a gradient clipping layer, which is arranged at any layer connection of the network structure, for optimizing the model training process; the decoding module further introduces a mixed convolution layer, which is divided into multiple subgroups, and each subgroup uses a convolution kernel of different scales to extract features.

[0040] As can be seen from the above embodiments, the application introduces a hierarchical window self-attention mechanism and a shift window strategy using a Swin Transformer structure, so that the model can efficiently capture local and global features at different scales, overcoming the problem of insufficient long-range dependency modeling ability caused by fixed convolution kernels in CNN, while avoiding the high computational complexity of traditional Transformer. Furthermore, the expression ability of the model for image spatial hierarchical features is significantly improved, and higher accuracy is achieved in congestion prediction and other tasks.

[0041] The feature pyramid structure is adopted to fuse deep semantic information and shallow detail features through a top-down path and a horizontal connection, overcoming the problem of insufficient semantic information of shallow features and low spatial resolution of deep features in ordinary hierarchical structures, and significantly improving the detection ability of the model for multi-scale targets.

[0042] The complex Swin Transformer module in the decoding part is replaced by a lightweight feature pyramid (FPN) convolution structure, overcoming the problem of high computational complexity and large parameter quantity of the Transformer decoder, while retaining the multi-scale feature fusion capability. On the premise of maintaining the prediction accuracy of the model, the model volume is greatly reduced, and the prediction speed is greatly improved.

[0043] Corresponding to the foregoing embodiments of the integrated circuit congestion prediction method, the application also provides embodiments of an integrated circuit congestion prediction device.

[0044] Figure 9 is a block diagram of an integrated circuit congestion prediction device according to an exemplary embodiment. Referring to Figure 9 , the device comprises: an extraction module 1 for extracting at least two congestion prediction feature maps of the integrated circuit to be tested; a processing module 2 for performing rasterization and normalization processing on each of the above congestion prediction feature maps to obtain a corresponding two-dimensional pixel matrix map; a prediction module 3 for inputting the two-dimensional pixel matrix map into a congestion prediction model to obtain a congestion hotspot map; wherein the congestion prediction model comprises an encoding module, a decoding module and a double upsampling module; The encoding module is configured to fuse at least two feature map features, and generate a first feature matrix through unit segmentation and linear embedding processing, and then input the first feature matrix into a cascaded Swin Transformer module to perform hierarchical feature encoding, and each module outputs an encoded congestion feature matrix of a corresponding level; The decoding module is configured to receive the first feature matrix and the encoded congestion feature matrices of the levels, perform smoothing and feature upsampling processing on the congestion feature matrix output by the bottommost Swin Transformer module in sequence, and then perform element-by-element superposition on the congestion feature matrix output by the Swin Transformer module of the upper layer, and iteratively perform the above operations in sequence until the first feature matrix is fused, and finally output a second feature matrix. The double upsampling module is configured to perform linear interpolation upsampling and pixel-level feature recovery on the second feature matrix, and output congestion hotspot distribution maps in horizontal and vertical directions.

[0045] As to the device in the above embodiment, the specific manner in which the various modules perform operations has been described in detail in the embodiment of the method, and will not be described in detail here.

[0046] For the device embodiment, since it basically corresponds to the method embodiment, the relevant part is described in the method embodiment. The device embodiments described above are only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present application according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0047] Correspondingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the integrated circuit congestion prediction method as described above.

[0048] Correspondingly, the present application also provides a computer readable storage medium having computer instructions stored thereon, which are executed by a processor to implement the integrated circuit congestion prediction method as described above.

[0049] Embodiment 2: In this embodiment, three features, i.e. macro unit distribution feature map, rectangular region line density distribution feature map and rectangular region pin line density distribution feature map, are selected as inputs of the model, and the corresponding feature map acquisition method and processing method are the same as those in embodiment 1.

[0050] In addition, the number of input channels of the congestion prediction model is adjusted to 3, and the input features become HxWx3, each channel corresponding to a feature map.

[0051] Similarly, the combination forms of the remaining three features are not described here.

[0052] Embodiment 3: In this embodiment, the rectangular region line density feature and the pin region line density feature map are selected as the input of the model, and the number of input channels of the congestion prediction model is adjusted to 2, and the input features become HxWx2, each channel corresponding to a feature map.

[0053] Similarly, the combination forms of the remaining two features are not described here.

[0054] Figure 10 is the index prediction map predicted by different numbers of feature maps, where (a) is the structural similarity index (SSIM), (b) is the mean square error index (NRMS), and (c) is the score index. The three bar charts in the figure correspond to the input of four feature maps, the input of three feature maps, and the input of two feature maps in embodiments 1, 2, and 3.

[0055] As shown in embodiments 2 and 3, using three or two features as the input of the model can achieve a certain prediction effect, but the score index is reduced by 37.4% and 5.1% respectively compared with the prediction model of embodiment 1 inputting four features, and the SSIM and NRMS are different degrees of deterioration.

[0056] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given are exemplary only and the true scope and spirit of the application is indicated by the claims. The true scope and spirit of the application is indicated by the claims.

[0057] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the claims that follow.

Claims

1. A method for predicting congestion of an integrated circuit, characterized in that: include: extracting at least two congestion prediction feature maps of the integrated circuit under test; Rasterize and normalize each of the above congestion prediction feature maps to obtain the corresponding two-dimensional pixel matrix map; Inputting the two-dimensional pixel matrix image into a congestion prediction model to obtain a congestion hotspot map; The congestion prediction model includes: an encoding module, a decoding module and a double upsampling module; The encoding module is used to fuse at least two feature map features and generate a first feature matrix through unit segmentation and linear embedding processing, and then input the first feature matrix into the cascaded Swin Transformer modules for hierarchical feature encoding, and each module outputs a coded congestion feature matrix of the corresponding level; The decoding module is used to receive the first feature matrix and the coding congestion feature matrices of each layer, first smooth and upsample the congestion feature matrix output by the bottom Swin Transformer module, and then superimpose it element by element with the congestion feature matrix output by the upper Swin Transformer module, and iterate in this way until the first feature matrix is ​​fused, and finally output the second feature matrix; The dual upsampling module is used to perform linear interpolation upsampling and pixel-level feature restoration on the second feature matrix, and output congestion hotspot distribution maps in the horizontal and vertical directions.

2. The method according to claim 1, characterized in that The congestion prediction feature map is selected from two or more of a macro cell distribution feature map, a rectangular area line density distribution feature map, a rectangular area pin line density distribution feature map, and a cell density distribution feature map.

3. The method according to claim 2, characterized in that The macro cell distribution characteristic diagram, the rectangular area line density distribution characteristic diagram, the rectangular area pin line density distribution characteristic diagram and the cell density distribution characteristic diagram are obtained by: In the layout planning stage, a macro cell distribution feature map is extracted; During the layout stage, the following three feature maps are extracted simultaneously: a rectangular area line density distribution feature map, a rectangular area pin line density distribution feature map, and a cell density distribution feature map.

4. The method according to claim 1, wherein Each of the above congestion prediction feature maps is rasterized and normalized to obtain the corresponding two-dimensional pixel matrix map, including: Each of the above congestion prediction feature maps is rasterized, and the feature value of each feature map is represented by the pixel value of its corresponding pixel point; the rasterized feature map is converted into a two-dimensional pixel matrix map with the same resolution.

5. The method according to claim 2, characterized in that For the rectangular area line density distribution feature map, the eigenvalue calculation formula is as follows: ； in, Represents the line density eigenvalue of the rectangular area corresponding to the k-th network on the plane coordinate (i, j), where is the difference between the maximum and minimum coordinates of all nodes in network k in the horizontal direction, is the difference between the maximum coordinate and the minimum coordinate in the vertical direction; the network k is composed of all units from the chip input port to the output port connected in series; Represents the ratio of the overlapping area between network k and the (i, j)th pixel block to the area of ​​the pixel block; described The specific calculation formula is as follows: ； ； ； In the formula represents the area of ​​the (i, j)th pixel block, The overlapping area of ​​the (i, j)th pixel block and the kth network, , Respectively represent the maximum and minimum values ​​of the horizontal coordinates of the (i, j)th pixel block, , Respectively represent the maximum and minimum values ​​of the vertical coordinates of the (i, j)th pixel block; , , , It represents the maximum and minimum values ​​in the horizontal direction and the maximum and minimum values ​​in the vertical direction of the k-th network respectively.

6. The method according to claim 1, characterized in that The congestion prediction model adopts a feature pyramid. The encoding module includes a linear embedding submodule and several Swin Transformer modules connected in series. The decoding module includes several decoding submodules connected in series. Each decoding submodule includes a smoothing layer and a stacked upsampling module. The feature information output by the linear embedding submodule and the several Swin Transformer modules is bridged layer by layer to the decoding submodule to restore the congestion characteristics.

7. The method according to claim 1, characterized in that The smoothing process is performed through a convolution function of nn.conv, which is used to change the number of channels of the next layer to obtain the congestion feature matrix information, which is the same as the congestion feature matrix information obtained by the crossover.

8. The method according to claim 1, characterized in that The encoding module and / or decoding module also includes a gradient clipping layer, which is set at any inter-layer connection of the network structure to optimize the model training process; the decoding module also introduces a hybrid convolutional layer, whose channels are divided into multiple subgroups, and each subgroup uses a convolution kernel of a different scale for feature extraction.

9. An integrated circuit congestion prediction device, characterized in that: include: an extraction module, configured to extract at least two congestion prediction feature maps of the integrated circuit under test; A processing module is used to perform rasterization and normalization processing on each of the above congestion prediction feature maps to obtain a corresponding two-dimensional pixel matrix map; A prediction module, configured to input the two-dimensional pixel matrix image into a congestion prediction model to obtain a congestion hotspot map; The congestion prediction model includes: an encoding module, a decoding module and a double upsampling module; The encoding module is used to fuse at least two feature map features and generate a first feature matrix through unit segmentation and linear embedding processing, and then input the first feature matrix into the cascaded Swin Transformer modules for hierarchical feature encoding, and each module outputs a coded congestion feature matrix of the corresponding level; The decoding module is used to receive the first feature matrix and the coding congestion feature matrices of each layer, first smooth and upsample the congestion feature matrix output by the bottom Swin Transformer module, and then superimpose it element by element with the congestion feature matrix output by the upper Swin Transformer module, and iterate in this way until the first feature matrix is ​​fused, and finally output the second feature matrix; The dual upsampling module is used to perform linear interpolation upsampling and pixel-level feature restoration on the second feature matrix, and output congestion hotspot distribution maps in the horizontal and vertical directions.

10. An electronic device, characterized in that: include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 8.

Images