Microarchitecture design space exploration method based on gaussian mixture regression
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2023-01-04
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies in processor design suffer from problems such as increased design complexity, process approaching physical limits, long verification and simulation times, and unbalanced multi-objective optimization, making it difficult to quickly and effectively explore the processor microarchitecture design space.
We adopt a microarchitecture design space exploration method based on Gaussian mixture regression, combined with Pareto front for multi-objective optimization, using Bayesian optimization as a framework, and combining conjugate gradient method and Gaussian mixture regression model for eigenvector estimation, thereby reducing the number of simulations and validations and improving model fitting accuracy.
It achieves high model fitting accuracy while shortening algorithm execution speed, improves the efficiency and accuracy of processor microarchitecture design, reduces the number of simulations and verifications, and simplifies the multi-objective optimization process.
Smart Images
Figure CN115983129B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of processor design, and in particular to a microarchitecture design space exploration method based on Gaussian mixture regression. Background Technology
[0002] In recent years, the open-source instruction set RISC-V has gradually become a focus of CPU design. Compared to other instruction sets, RISC-V is characterized by its simplicity, making it ideal for developing small, fast, and low-power processors. Since RISC-V was open-sourced in 2010, the industry has developed a complete toolchain for it, allowing developers to create fully functional processors. However, with the development of chips, the processor industry is facing significant challenges. First, processors are becoming increasingly complex, with a growing number of transistors, posing significant design challenges for designers. Second, with advancements in manufacturing processes, chip production is approaching its physical limits, making it difficult to improve chip performance simply by increasing transistor density. Therefore, the industry needs to focus its research on other aspects of chips, such as power consumption and area. Third, chip verification and simulation account for the largest portion of the chip development cycle. To address these issues, the industry needs faster and more convenient methods for processor development.
[0003] Existing solutions include: using statistical methods and surrogate models to approximate the relationship between processor microarchitecture and processor performance; for example, reference [1] Guo Qi, Chen Tianshi, and Chen Yunji, “Multi-core design space exploration technology based on model tree,” Journal of Computer-Aided Design and Graphics 24, no. 06 (2012): 710–20. proposes a model tree-based architecture modeling method, which belongs to the white-box model and can interpret the model, but has poor robustness; reference [2] Dandan Li et al., “Efficient Design Space Exploration via StatisticalSampling and AdaBoost Learning,” in Proceedings of the 53rd Annual Design Automation Conference (DAC '16: The 53rd Annual Design Automation Conference2016, Austin Texas: ACM, 2016), 1–6, proposes a method based on ANN As a design space exploration method for improving weak learners, this method has a prominent effect on approximating the relationship between target value and input, but it lacks the balance between multi-objective optimization and still requires a large number of simulations and verifications; Reference [3] Qi Guo et al., "Robust Design Space Modeling", ACM Transactions on Design Automation of Electronic Systems 20, no. 2 (March 2, 2015): 1–22, uses a stacking algorithm to use SVM, ANN and M5 model tree as weak learners, and uses M5 model tree to integrate the results of weak learners, and proposes a robust design space exploration algorithm, but this method requires a long training time; Reference [4] Bai C, Sun Q, Zhai J, et al. BOOM-Explorer: RISC-V BOOM Microarchitecture Design SpaceExploration Framework[C] / / 2021 IEEE / ACM International Conference On ComputerAided Design (ICCAD). 0.A related Bayesian optimization method was used as a framework to reduce the number of processor microarchitecture simulations and verifications, and multi-objective optimization functionality was implemented. However, the authors did not provide more specific implementation methods, and this method can only select the optimal microarchitecture from the generated microarchitecture dataset, but cannot estimate microarchitectures outside the dataset. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a microarchitecture design space exploration method based on Gaussian mixture regression. This method combines the Pareto front to perform multi-objective optimization of the design goal; uses Bayesian optimization as a framework to reduce the number of simulations and verifications required in the exploration process, shortening the algorithm execution speed while maintaining high model fitting accuracy; and uses the conjugate gradient method to estimate the feature vector of the optimal microarchitecture.
[0005] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0006] Microarchitecture design space exploration methods based on Gaussian mixture regression include:
[0007] S1. Extract the characteristic attributes of the microarchitecture, set the constraint rules of the microarchitecture, and determine the design space D to be explored;
[0008] S2. Randomly sample several feature points X from the design space, simulate and verify these feature points X to obtain the corresponding PPA value Y, and remove the sampled feature points from the design space to obtain the remaining design space. ;
[0009] S3. Find the Pareto set in Y. and the corresponding Pareto solution set Find Pareto hypervolume;
[0010] S4. Input X and Y into a Gaussian mixture regression model for model training to obtain model G;
[0011] S5, in the design space Multiple feature points are randomly sampled to form a set X*, which is then fed into the model G for prediction, resulting in the expected set Y* of the predicted values and the set Σ* of the covariance matrix of the target values.
[0012] S6. Input X*, Σ*, Y*, and Y into the CEIPV acquisition function. Using each eigenvector in X* and each covariance matrix in Σ* as initial inputs, use the conjugate gradient method to find the eigenvector that maximizes the CEIPV function. ;
[0013] S7, Judgment Does it conform to the constraints of microarchitecture? If it does, proceed directly to the next step; otherwise, search for relevant elements in the design space. The nearest feature point is used as the true Then proceed to the next step;
[0014] S8, to Simulation and verification were performed to obtain... Actual PPA value and put From design space Remove from the middle;
[0015] S9, will join in and X, join in and ;
[0016] S10, in Find the Pareto set and solve for the corresponding Pareto hypervolume. If the Pareto hypervolume is larger than the original Pareto hypervolume, proceed directly to the next step; otherwise, proceed to the next step. , From the set and Remove it from the list and proceed to the next step;
[0017] S11, Cycle Step S This continues until the loop termination condition is met;
[0018] S12, Final Output This is the Pareto solution set of the processor microarchitecture. This is the corresponding PPA value.
[0019] Furthermore, after extracting the feature attributes of the microarchitecture in step S1, the labeling method is used to replace the actual feature values of the microarchitecture.
[0020] Furthermore, in step S3, we find... When calculating the Pareto hypervolume, two reference points are first artificially selected. and And the selected reference point and The enclosed hypervolume contains the target values of all feature points in the design space.
[0021] Furthermore, in step S4, the Gaussian mixture regression model is a generative model, which first solves for the joint probability density of the input and output, and then uses Bayes' theorem to solve for the conditional probability distribution of the output given the input, specifically including:
[0022] Given the input of the training set Output Input to be predicted ,but Corresponding output The probability distribution is shown in equation (1):
[0023]
[0024] in This represents the value estimated by the i-th Gaussian distribution. The mean, its number of rows and They have the same dimension, and have
[0025]
[0026] in Let be the input mean vector and output mean vector of the i-th Gaussian distribution after training, respectively. The number of rows equals dimensionality The number of rows equals dimensionality;
[0027] The estimated value for the i-th Gaussian distribution The variance has
[0028]
[0029] in Let be the covariance among the inputs of the i-th Gaussian distribution. Let be the covariance between the output and input of the i-th Gaussian distribution. Let be the covariance between the outputs of the i-th Gaussian distribution;
[0030] Finally obtained The estimated value of the output y and the corresponding covariance matrix They are respectively:
[0031]
[0032]
[0033] in
[0034]
[0035] The mean and covariance of the i-th Gaussian distribution mentioned in formulas (2) and (3) are the parameters that need to be trained; the EM method is used to iteratively optimize these parameters, and the initial number of clusters of the Gaussian mixture regression model is set to be equal to the number of categories of a certain feature in the processor microarchitecture; finally, the Bayesian information criterion is used to evaluate the number of clusters in the data to reduce the Gaussian components.
[0036] Furthermore, for the acquisition function CEIPV in step S6, assuming that the target values are correlated, the expected increment of the Pareto hypervolume formed by the target values is calculated.
[0037] Furthermore, in step S7, the Euclidean Distance algorithm is used to find [a match / match] in the design space. The nearest feature point is used as the true .
[0038] Furthermore, in step S8, during simulation and verification, the eigenvalue vector is converted into the actual microarchitecture, and then the microarchitecture is verified and simulated.
[0039] Compared with existing technologies, the principles and advantages of this solution are as follows:
[0040] 1. Using Bayesian optimization methods for incomplete supervised learning can reduce the cost of dataset labeling and accelerate model training by leveraging prior probabilities.
[0041] 2. The Gaussian mixture regression model used is the fastest among mixture models and has unbiased properties. It has achieved excellent results in robot path optimization. The Gaussian mixture regression model can simultaneously calculate the predicted mean of the target value and the covariance between multiple targets. Compared with other models, it can perform better multi-objective optimization.
[0042] 3. The Bayesian information optimization criterion is used to determine the final number of Gaussian components, realizing the function of dynamically optimizing the number of Gaussian components, so that the model can better approximate the true function.
[0043] 4. The acquisition function CEIPV used combines target values, providing the observation point with the highest uncertainty while considering the balance between targets. Traditional methods for calculating the next observation point involve using different acquisition functions for different targets and then employing strategies to select the final next observation point, failing to simultaneously consider the balance between targets and model uncertainty prediction. This results in the inability to select the input at the point of maximum model uncertainty, and the target balancing method is more complex. Therefore, the method used in this proposal is simpler and more effective than traditional methods. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a flowchart illustrating the principle of the microarchitecture design space exploration method based on Gaussian mixture regression of this invention.
[0046] Figure 2 A schematic diagram of the Pareto hypervolume;
[0047] Figure 3 This is a schematic diagram of the Pareto hypervolume increment. Detailed Implementation
[0048] The present invention will be further described below with reference to specific embodiments:
[0049] like Figure 1 As shown in this embodiment, the microarchitecture design space exploration method based on Gaussian mixture regression includes the following steps:
[0050] S1. Extract the characteristic attributes of the microarchitecture, set the constraint rules of the microarchitecture, and determine the design space D to be explored;
[0051] In this step, all feature points within the design space D must conform to the architectural constraints of the microarchitecture of the processor to be designed. To standardize the units of input features, this embodiment uses a labeling method to replace the feature values of the actual processor microarchitecture. For example, if the number of entries in the fetch buffer register of the processor microarchitecture is {8, 16, 24, 32, 35, 40}, then the actual feature values are replaced by the six elements {1, 2, 3, 4, 5, 6} in sequence.
[0052] S2. Randomly sample several feature points X from the design space, simulate and verify these feature points X to obtain the corresponding PPA value Y, and remove the sampled feature points from the design space to obtain the remaining design space. ;
[0053] S3. Find the Pareto set in Y. and the corresponding Pareto solution set Find Pareto hypervolume;
[0054] In this step, determining the Pareto hypervolume requires first manually selecting two reference points. and The hypervolume enclosed by the selected reference points must contain the target values of all feature points in the design space. As the reference point for the minimum value, This serves as the reference point for the maximum value. Figure 2 Using a Pareto hypervolume in two-dimensional space as a schematic diagram, this paper illustrates the method for setting the reference point of the Pareto hypervolume. The region in the upper right corner is the Pareto hypervolume, and the point... The reference points are set by humans. The three circled points are the Pareto sets in the data, and the other three points are the dominated points in the data.
[0055] S4. Input X and Y into a Gaussian mixture regression model for model training to obtain model G;
[0056] In this step, the Gaussian mixture regression model is a generative model, meaning that it first solves for the joint probability density of the input and output, and then uses Bayes' theorem to solve for the conditional probability distribution of the output given the input. Specifically, this includes:
[0057] Given the input of the training set Output Input to be predicted ,but Corresponding output The probability distribution is shown in equation (1):
[0058]
[0059] in This represents the value estimated by the i-th Gaussian distribution. The mean, its number of rows and They have the same dimension, and have
[0060]
[0061] in Let be the input mean vector and output mean vector of the i-th Gaussian distribution after training, respectively. The number of rows equals dimensionality The number of rows equals dimensionality;
[0062] The estimated value for the i-th Gaussian distribution The variance has
[0063]
[0064] in Let be the covariance among the inputs of the i-th Gaussian distribution. Let be the covariance between the output and input of the i-th Gaussian distribution. Let be the covariance between the outputs of the i-th Gaussian distribution;
[0065] Finally obtained The estimated value of the output y and the corresponding covariance matrix They are respectively:
[0066]
[0067]
[0068] in
[0069]
[0070] The mean and covariance of the i-th Gaussian distribution mentioned in formulas (2) and (3) are the parameters that need to be trained. The EM method is used to iteratively optimize these parameters, and the initial number of clusters in the Gaussian mixture regression model is set to be equal to the number of categories of a certain feature in the processor microarchitecture. For example, in the design space, if the decoder bit width of the microarchitecture, which is one of the features, is 5, then the initial number of components in the Gaussian mixture regression model is also 5. Finally, the Bayesian Information Criterion (BIC) is used to evaluate the number of clusters in the data to reduce the number of Gaussian components.
[0071] S5, in the design space Multiple feature points are randomly sampled to form a set X*, which is then fed into the model G for prediction, resulting in the expected set Y* of the predicted values and the set Σ* of the covariance matrix of the target values.
[0072] S6. Input X*, Σ*, Y*, and Y into the CEIPV acquisition function. Using each eigenvector in X* and each covariance matrix in Σ* as initial inputs, use the conjugate gradient method to find the eigenvector that maximizes the CEIPV function. ;
[0073] The acquisition function CEIPV in this step is based on the assumption that the target values are correlated, and calculates the expected increment of the Pareto hypervolume formed by the target values. The increment of the Pareto hypervolume is as follows: Figure 3 As shown.
[0074] S7, Judgment If the microarchitecture constraints are met, proceed to the next step; otherwise, use the Euclidean Distance algorithm to find a match in the design space. The nearest feature point is used as the true Then proceed to the next step;
[0075] S8, to Simulation and verification were performed to obtain... Actual PPA value and put From design space The eigenvectors are removed from the model; during simulation and verification, the eigenvectors are converted into the actual microarchitecture, and then the microarchitecture is verified and simulated.
[0076] S9, will join in and X, join in and ;
[0077] S10, in Find the Pareto set and solve for the corresponding Pareto hypervolume. If the Pareto hypervolume is larger than the original Pareto hypervolume, proceed directly to the next step; otherwise, proceed to the next step. , From the set and Remove it from the list and proceed to the next step;
[0078] S11, Cycle Step S This continues until the loop termination condition is met;
[0079] S12, Final Output This is the Pareto solution set of the processor microarchitecture. This is the corresponding PPA value.
[0080] In this embodiment, the power consumption, performance, and area of the processor chip are referred to as PPA.
[0081] The following example, using the BOOM (Berkeley Out-of-Order Machine) architecture for space exploration, will be used to explain the above design space exploration method:
[0082] A1. Based on the BOOM architecture, the feature values shown in Table 1 can be extracted:
[0083] Table 1. Eigenvalues of BOOM Microarchitecture
[0084]
[0085] Eigenvalue engineering is performed on Table 1: Based on the microarchitecture constraints in Table 3, the feature attributes IntPhyRegister, STQEntry, and FpIssueWidth can be removed to reduce computational cost. Table 2 is then obtained.
[0086] Table 2 Modified BOOM microarchitecture feature values
[0087]
[0088] A2. Based on Table 2, randomly select 12 eigenvectors that meet the constraints of the BOOM microarchitecture in Table 3 according to certain rules. The rule is as follows: based on the number of categories in DecodeWidth, two feature vectors are randomly selected from each category;
[0089] Table 3 BOOM Microarchitectural Constraints
[0090]
[0091] A3, will The corresponding microarchitecture is placed in a virtual prototyping platform to obtain the simulated PPA value. The simulated feature vectors are removed from the design space;
[0092] A4. Find the Pareto set in Y. and the corresponding set of eigenvectors Find Pareto hypervolume ;
[0093] A5. Input X and Y into a Gaussian mixture regression model for training to obtain model G. The initial number of clusters K in the Gaussian mixture regression model is set according to DecodeWidth in Table 2. The training method uses EM, and BIC is used to optimize the number of Gaussian components.
[0094] A6. Based on Table 2, randomly generate 100 feature vectors X* that conform to Table 3, and put them into model G for prediction to obtain the expected set of predicted values Y* and the set of covariance matrices of the target values Σ*;
[0095] A7. Input X*, Σ*, Y*, and Y into the CEIPV function. Using each eigenvector in X* and each covariance matrix in Σ* as initial inputs, use the conjugate gradient method to find the eigenvector that maximizes CEIPV. ;
[0096] A8. Judgment Does it meet the microarchitecture constraints in Table 3? If it does, proceed directly to step A9; otherwise, modify it. New feature vectors are generated from certain features in the table. Then, feature vectors that do not conform to Table 3 are removed. Finally, features that are far from the generated new feature vectors are searched. The most recent feature vector as the true Finally, execute step A9;
[0097] A9, Yes Simulation and verification were performed to obtain... Actual PPA value and put From design space Remove from the middle;
[0098] A10, will join in and X, join in and ;
[0099] A11, in Find the Pareto set and solve for the corresponding Pareto hypervolume. If the Pareto hypervolume is larger than the original Pareto hypervolume... If the value is large, proceed directly to step 11; otherwise, proceed to step 2. , From the set and Remove it from the list and then proceed to step A11;
[0100] A12, Loop Steps Set the loop termination condition to "the number of solutions in the Pareto solution set reaches 20" or "the maximum number of loops reaches 50".
[0101] A13, Output and These serve as the Pareto solution set and the corresponding set of simulated PPA values for the processor microarchitecture, respectively.
[0102] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.
Claims
1. A microarchitecture design space exploration method based on Gaussian mixture regression, characterized in that, Includes the following steps: S1. Extract the characteristic attributes of the microarchitecture, set the constraint rules of the microarchitecture, and determine the design space D to be explored; S2. Randomly sample several feature points X from the design space, simulate and verify these feature points X to obtain the corresponding PPA value Y, and remove the sampled feature points from the design space to obtain the remaining design space. ; S3. Find the Pareto set in Y. and the corresponding Pareto solution set Find Pareto hypervolume; S4. Input X and Y into a Gaussian mixture regression model for model training to obtain model G; S5, in the design space Multiple feature points are randomly sampled to form a set X*, which is then fed into the model G for prediction, resulting in the expected set Y* of the predicted values and the set Σ* of the covariance matrix of the target values. S6. Input X*, Σ*, Y*, and Y into the CEIPV acquisition function. Using each eigenvector in X* and each covariance matrix in Σ* as initial inputs, use the conjugate gradient method to find the eigenvector that maximizes the CEIPV function. ; S7, Judgment Does it conform to the constraints of microarchitecture? If it does, proceed directly to the next step; otherwise, search for relevant elements in the design space. The nearest feature point is used as the true Then proceed to the next step; S8, to Simulation and verification were performed to obtain... Actual PPA value and put From design space Remove from the middle; S9, will join in and X, join in and ; S10, in Find the Pareto set and solve for the corresponding Pareto hypervolume. If the Pareto hypervolume is larger than the original Pareto hypervolume, proceed directly to the next step; otherwise, proceed to the next step. , From the set and Remove it from the list and proceed to the next step; S11, Cyclic Steps This continues until the loop termination condition is met; S12, Final Output This is the Pareto solution set of the processor microarchitecture. This is the corresponding PPA value.
2. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, After extracting the feature attributes of the microarchitecture in step S1, the labeling method is used to replace the actual feature values of the microarchitecture.
3. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, In step S3, find When calculating the Pareto hypervolume, two reference points are first artificially selected. and And the selected reference point and The enclosed hypervolume contains the target values of all feature points in the design space.
4. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, In step S4, the Gaussian mixture regression model is a generative model, which first solves for the joint probability density of the input and output, and then uses Bayes' theorem to solve for the conditional probability distribution of the output given the input. Specifically, it includes: Given the input of the training set Output Input to be predicted ,but Corresponding output The probability distribution is shown in equation (1): ； in This represents the value estimated by the i-th Gaussian distribution. The mean, its number of rows and Since they have the same dimension, we have: ； in Let be the input mean vector and output mean vector of the i-th Gaussian distribution after training, respectively. The number of rows equals dimensionality The number of rows equals dimensionality; The estimated value for the i-th Gaussian distribution The variance is: ； in Let be the covariance among the inputs of the i-th Gaussian distribution. Let be the covariance between the output and input of the i-th Gaussian distribution. Let be the covariance between the outputs of the i-th Gaussian distribution; Finally obtained The estimated value of the output y and the corresponding covariance matrix They are respectively: ； ； in ； The mean and covariance of the i-th Gaussian distribution mentioned in formulas (2) and (3) are the parameters that need to be trained; the EM method is used to iteratively optimize these parameters, and the initial number of clusters of the Gaussian mixture regression model is set to be equal to the number of categories of a certain feature in the processor microarchitecture; finally, the Bayesian information criterion is used to evaluate the number of clusters in the data to reduce the Gaussian components.
5. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, For the acquisition function CEIPV in step S6, assuming that the target values are correlated, the expected increment of the Pareto hypervolume formed by the target values is calculated.
6. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, In step S7, the Euclidean Distance algorithm is used to find the corresponding [object] in the design space. The nearest feature point is used as the true .
7. The microarchitecture design space exploration method based on Gaussian mixture regression according to claim 1, characterized in that, In step S8, during simulation and verification, the eigenvalue vector is converted into the actual microarchitecture, and then the microarchitecture is verified and simulated.