Process angle prediction method and system in static timing analysis based on ensemble learning

Abstract

The application discloses a process angle prediction method and system in static timing analysis based on ensemble learning, which combines multilayer perceptron and graph neural network for process angle prediction. Firstly, the MLP model is used for preliminary prediction, and then the prediction result of the MLP is input into the GNN model to optimize the prediction result in the form of graph structure data. The application also uses a generative adversarial network to generate synthetic data, thereby expanding the training set and improving the prediction ability of the model. The method of the application can significantly improve the prediction accuracy and is more efficient in running speed than the traditional method. The method not only reduces the calculation overhead of the traditional STA method, but also has good adaptability and high efficiency, and is particularly suitable for process angle prediction tasks of large-scale circuits. The system architecture of the application includes a data acquisition module, an MLP module, a GNN module and a GAN data generation module, and can realize efficient and accurate process angle prediction in industrial applications.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated learning and static timing analysis, and particularly relates to a process corner prediction method combining a multilayer perceptron (MLP) and a graph neural network (GNN) and a system thereof. BACKGROUND

[0002] Static timing analysis is used to evaluate the timing performance of integrated circuits under different environmental and process conditions. However, as the size of the circuit increases, the traditional STA method needs to perform simulation calculations repeatedly, resulting in high time cost. In addition, since different process conditions can have a significant impact on path delay, STA needs to be performed under multiple process corners, which further increases the computational overhead.

[0003] At present, most of the existing researches are based on traditional algorithms for path delay prediction. However, these methods perform poorly when dealing with complex and non-Gaussian distributed process variations. Therefore, how to use modern machine learning methods to improve the accuracy of process corner prediction and reduce the computation time has become a research hotspot. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application discloses a process corner prediction method in static timing analysis based on integrated learning and a system thereof, which effectively improves the prediction accuracy of unobserved process corners by combining MLP and GNN, to solve the problems raised in the background.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a process corner prediction method in static timing analysis based on integrated learning, which adopts an integrated learning framework of multilayer perceptron (MLP) and graph neural network (GNN) for process corner prediction, effectively improving the prediction accuracy of unobserved process corners, including the following steps:

[0006] S1, obtain the required timing data through simulation and preprocess, construct a timing matrix according to the circuit path, and generate a data set for training;

[0007] S2, then, the MLP model obtains a preliminary path delay prediction result through training; then, the preliminary prediction result is passed to the GNN model for further optimization, specifically based on the topological structure of the current predicted path, the graph structure of the graph neural network is constructed, and the node feature vector and the edge feature vector are constructed, the fan-in and fan-out number of the logic cell, the cell category, and the rising and falling of the input signal are taken as the node features, and the timing increment of the feature end corner is taken as the edge feature, and finally the timing result of the corresponding circuit path under the to-be-predicted end corner is output.

[0008] S3, generating new data samples according to the training data through the GAN network, expanding the training set, and additionally generating 20% of data for the current data set for training to further improve the prediction capability;

[0009] S4, finally, the prediction result is fed back to the static timing analysis (STA) tool, and the user obtains the optimized path delay prediction result and adjusts if necessary.

[0010] Preferably, in step S1, the data set used includes timing path delay data under different process angles extracted from the PrimeTime tool, and a timing matrix is constructed according to the circuit path; the correlation of all end angles is analyzed by the Pearson correlation coefficient, and the end angle with the highest correlation is selected as the seed end angle in the feature set.

[0011] Preferably, in step S2, the input of the MLP network is the path delay data extracted from the static timing analysis (STA) tool; the seed end angle is used to predict other end angle information, the end angle with the lowest prediction accuracy is selected to join the feature set, and the feature end angle is iterated out in turn, the number of feature end angles is determined through the relationship between the prediction accuracy and the time spent in prediction, and the prediction result is used as the global information of the GNN network as the input.

[0012] Preferably, in step S2, the GNN model models the non-linear relationship between the circuit paths through the graph structure data, including node features, edge features and global features, so as to capture the complex relationship between the paths and optimize the path delay prediction result, wherein:

[0013] 1) node features, including fan-in and fan-out number of logic cells, cell category and input signal rise and fall time;

[0014] 2) edge features, including timing increment of feature end angle;

[0015] 3) global features, provided by the preliminary prediction result of the MLP model, used to enhance the input information of the GNN model.

[0016] Preferably, in step S3, the generator part of the GAN network generates new samples according to the existing path delay data, and the discriminator is responsible for verifying the authenticity of the synthesized data; in the training process, the data generated by the GAN network is mixed with the original data for the MLP and GNN to use together.

[0017] The application also provides a process angle prediction system in a static timing analysis based on integrated learning, and the system architecture comprises:

[0018] A data collection module is configured to obtain path delay data under different process corners from a static timing analysis tool, construct a timing matrix, and perform preprocessing;

[0019] An MLP module is configured to preliminarily predict path delay, train an MLP model using collected data, and generate a preliminary prediction result.

[0020] A GNN module is configured to optimize the prediction result of the MLP, take the prediction result of the MLP as input, and further optimize the prediction result through a graph neural network.

[0021] A GAN data generation module is configured to generate synthetic data based on existing data and expand a training set.

[0022] Preferably, the prediction system further comprises a feedback module configured to feed back the prediction result to a static timing analysis platform and optimize path delay calculation according to the prediction result.

[0023] Preferably, the GNN module is configured to process topological information of a circuit and capture complex dependency relationships between paths.

[0024] Compared with the prior art, the present application has the following beneficial effects:

[0025] 1. The present application uses STA software to perform static timing analysis on a reference circuit, extracts path delay data according to different process corners, generates synthetic data through a generative adversarial network, thereby expanding the training set and improving the prediction ability of the model, and improves the robustness and accuracy of the model.

[0026] 2. The present application uses MLP to preliminarily predict process corners, obtains preliminary path delay values for each process corner, and further optimizes the preliminary prediction result of the MLP in combination with a GNN model, thereby capturing complex relationships between paths through graph structure learning, and being more efficient in running speed than traditional methods. The present application significantly improves the accuracy of path delay prediction in static timing analysis, not only reduces the time consumption of traditional STA methods, but also improves the prediction efficiency on the basis of high accuracy, has good adaptability and high efficiency, and is particularly suitable for process corner prediction tasks of large-scale circuits and can realize efficient and accurate process corner prediction in industrial applications.

[0027] 3. The present application is based on a pytorch platform, realizes timing prediction under multiple end corners in an offline network environment with a relatively simple neural network structure, has a relatively fast prediction speed and a relatively high prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, but do not constitute a limitation of the present application.

[0029] In the drawings:

[0030] Figure 1 is a schematic diagram of the end angle screening process which is a feature of the present application;

[0031] Figure 2 is a structural diagram of the combination of MLP and GNN model in the present application;

[0032] Figure 3 is a schematic diagram of the graph structure and feature vector construction of the GNN network part in the present application;

[0033] Figure 4 is a schematic diagram of the three-part information convergence process of node features, edge features and global features of the GNN network in the present application;

[0034] Figure 5 is a comparison diagram of generated data and real data of GAN in the present application;

[0035] Figure 6 is a residual diagram of the prediction results of the model after training under five different end angles for multiple paths;

[0036] Figure 7 is a comparison diagram of the prediction time and the simulation time of the commercial tool under the same end angle and different circuits after the model is trained in the present application. DETAILED DESCRIPTION

[0037] The preferred embodiments of the present application will be described below with reference to the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0038] Embodiment 1: Implementation of process angle prediction method in static timing analysis based on ensemble learning, specifically including the following steps:

[0039] 1. Data collection:

[0040] The method of the present application first collects timing path delay data under different process angles from a static timing analysis (STA) tool. Taking the commonly used PrimeTime software as an example, STA analysis is performed on multiple benchmark circuits (such as ISCAS'89, ITC'99, etc.).

[0041] In this process, the tool will extract path delay data according to different process angles (for example: tt0p85v25c, ss0p95v125v, etc.) to construct a timing matrix.

[0042] At the same time, as Figure 1The process shown involves obtaining a seed angle through correlation coefficient analysis, using this angle as a feature, and then using a neural network to predict other angles. The angle with the lowest prediction accuracy is added to the feature set, and this process is repeated to iterate and generate a certain number of feature angles for predicting the remaining angles.

[0043] 2. Preliminary MLP prediction:

[0044] like Figure 2 As shown, an MLP model is used for preliminary path delay prediction. The input to the MLP network is path delay data extracted from the STA tool. This dataset is preprocessed and normalized to a uniform scale to improve prediction accuracy. After training, the MLP model is able to make preliminary path delay predictions based on the input data.

[0045] 3. GNN optimizes prediction:

[0046] The prediction results of the MLP are used as input, and a graph neural network (GNN) is used for further optimization. The GNN model can capture the complex relationships between paths by processing the topological information of the circuit path, and optimize the predicted path delay through the information propagation mechanism.

[0047] like Figure 3 As shown, nodes in the GNN model represent various devices in the circuit, and edges represent the connections between these devices. In addition, nodes contain information such as the type of logic gate, the rise / fall delay of the input signal, and the fan-in / fan-out number of the unit. Edge nodes also contain the timing increment information of the corresponding connection at all feature endpoints. By learning the graph structure, GNN can more accurately adjust the prediction results and capture the relationship between path delay and circuit structure.

[0048] Figure 4 This demonstrates the mechanism of information transfer and aggregation in GNN networks across global feature vectors, node feature vectors, and edge feature vectors.

[0049] 4. Data Generation and Augmentation:

[0050] To further improve prediction accuracy, this invention employs GAN to generate synthetic data. By generating new data samples using GAN, the training set is expanded, making the model more robust when dealing with unseen process corners.

[0051] In GAN, the generator part generates new samples based on existing path delay data, while the discriminator is responsible for verifying the authenticity of these synthetic data. During training, the data generated by GAN is mixed with the original data and used by both MLP and GNN.

[0052] Figure 5The (a) tt0p85v125c, (b) ff1p16v125c, (c) ff1p16v25c, (d) ff0p95v125c, (e) ss0p95v125c show the trend between generated data and real data under different process corner conditions, the generated data can conform to the distribution of the real data, the use of generated data can reduce the time of simulation construction training set, and the model can converge faster when training.

[0053] 5. Experiment and verification:

[0054] The benchmark circuits used in the experiments of the embodiment include a plurality of circuit designs, such as ISCAS'89 and ITC'99, on which the model is trained and tested. The final results show that by integrating the framework of MLP and GNN, the predicted path delay has high accuracy, and the prediction time is significantly reduced compared with the traditional method.

[0055] Figure 6 The residual distribution of the trained model under different end angles and different circuits is shown, Figure 7 The comparison of the time taken by the predicted data obtained by the trained model and the timing data obtained by the simulation of the traditional commercial tool under the same end angle and the same circuit is shown.

[0056] Embodiment 2: Implementation of a process corner prediction system in static timing analysis based on ensemble learning, specifically comprising:

[0057] 1. System architecture:

[0058] The embodiment provides a process corner prediction system in static timing analysis based on ensemble learning. The system includes a data acquisition module, an MLP module, a GNN module, and a GAN data generation module.

[0059] Data acquisition module: responsible for extracting path delay data under different process corners from the STA tool, constructing a timing matrix and performing preprocessing.

[0060] MLP module: responsible for preliminary path delay prediction, training the MLP model using the collected data, and generating preliminary prediction results.

[0061] GNN module: takes the prediction results of MLP as input and further optimizes the prediction results through graph neural network. This module can process the topological information of the circuit and capture the complex dependency relationship between paths.

[0062] GAN data generation module: generates synthetic data based on existing data, enhances the training set, and improves the robustness of the model.

[0063] 2. Workflow:

[0064] When the system is working, first, the required timing data is acquired through the data acquisition module and preprocessed to generate a data set for training.

[0065] Next, the MLP module obtains preliminary path delay prediction results through training. Then, these preliminary prediction results are passed to the GNN module for more accurate optimization.

[0066] Finally, the GAN module generates additional data samples based on the training data to expand the training set and further improve the prediction capability.

[0067] Finally, the prediction results are fed back to the STA tool, and the user can obtain the optimized path delay prediction results and make adjustments if necessary.

[0068] Embodiment 3: Performance test of the system

[0069] In this embodiment, a plurality of benchmark circuits such as ISCAS'89, ITC'99, etc. are selected for testing under different process corners. The experimental setup includes the following steps:

[0070] Extract path delay data with the STA tool;

[0071] Train the data using MLP and GNN respectively and compare them;

[0072] Generate additional data samples using GAN to enhance the training set;

[0073] Finally, compare the optimized prediction results with the traditional method to verify the advantages of the method of the present application in prediction accuracy and computational efficiency.

[0074] Through Figure 6 , Figure 7 The test results show that the error of predicting path delay using the method of the present application is significantly lower than that of the traditional method, and the prediction time is significantly reduced under the same hardware conditions.

[0075] Finally, it should be noted that the above description is only a preferred example of the present application and does not limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A process angle prediction method in static timing analysis based on ensemble learning, using an ensemble learning framework of multilayer perceptron and graph neural network for process angle prediction, characterized in that, The method comprises the following steps: S1, obtaining required timing data through simulation and preprocessing to generate a data set for training; S2, then, the MLP model obtains a preliminary path delay prediction result through training; then, the preliminary prediction result is transmitted to the GNN model for further optimization; S3, the GNN network generates new data samples according to the training data and expands the training set; S4, finally, the prediction result is fed back to the static timing analysis tool, and the user obtains an optimized path delay prediction result; In step S1, the data set used includes timing path delay data under different process angles extracted from the PrimeTime tool, and a timing matrix is constructed according to the circuit path; the correlation degrees of all end angles are analyzed through correlation coefficient analysis, and the end angle with the highest correlation degree is selected as the seed end angle in the feature set; In step S2, the input of the MLP network is the path delay data extracted from the static timing analysis tool; the seed end angle is used to predict the information of other end angles, the end angle with the lowest prediction accuracy is selected to join the feature set, and the feature end angle is iterated out in turn, and the prediction result is used as the global information of the GNN network as the input; In step S2, the GNN model models the non-linear relationship between the circuit paths through the graph structure data, including node features, edge features and global features, so as to capture the complex relationship between the paths and optimize the path delay prediction result, wherein: 1) the node features include the fan-in and fan-out number of the logic unit, the unit category and the rise and fall time of the input signal; 2) the edge features include the timing increment of the feature end angle; 3) the global features are provided by the preliminary prediction result of the MLP model, and are used to enhance the input information of the GNN model.

2. The process corner prediction method in integrated learning based static timing analysis according to claim 1, characterized in that: In step S3, the generator part of the GAN network generates new samples according to the existing path delay data, and the discriminator is responsible for verifying the authenticity of the synthesized data; in the training process, the data generated by the GAN network is mixed with the original data for the MLP and GNN to use together.

3. A process corner prediction system in integrated learning based static timing analysis, applied to the method of claim 1, characterized in that, The system comprises: a data acquisition module for acquiring path delay data under different process angles from a static timing analysis tool, constructing a timing matrix and preprocessing; an MLP module for preliminary prediction of path delay, training the MLP model using the collected data to generate a preliminary prediction result; a GNN module for optimizing the prediction result of the MLP, taking the prediction result of the MLP as the input and further optimizing the prediction result through the graph neural network; a GAN data generation module for generating synthetic data based on existing data and expanding the training set.

4. The process corner prediction system in integrated learning based static timing analysis according to claim 3, wherein: The prediction system further comprises a feedback module for feeding back the prediction result to a static timing analysis platform and optimizing the path delay calculation according to the prediction result.

5. The process corner prediction system in integrated learning based static timing analysis according to claim 3, wherein: The GNN module is used to process the topological information of the circuit and capture the complex dependence relationship between the paths.

Images