A graphics rendering performance prediction method and device based on configuration space mapping

Abstract

The application discloses a kind of based on configuration space mapping's graphic rendering performance prediction method and device.The application first changes the configuration item of graphic rendering software, and collects configuration item-performance mapping, constructs dataset;Then the dataset is pretreated, and the configuration characteristics are encoded by Transformer, to generate high-dimensional configuration space mapping features;Subsequently, with the aid of ANFIS, the association rules of "similar configuration corresponding similar performance" are extracted from the mapping features, to obtain rule activation intensity;Then the mapping features and rule activation intensity are fused and input into multi-task neural network;Finally, MRE and MAE are used as loss function in the training process, and the parameters of Transformer and multi-task network are optimized jointly, to realize multi-index parallel prediction.The application realizes feature enhancement through configuration space mapping, combines rule reasoning and multi-task learning, and effectively improves the prediction accuracy of multiple performance indicators in complex graphic rendering scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer software engineering and performance testing technology, in particular to a graphics rendering performance prediction method and device based on configuration space mapping. BACKGROUND

[0002] With the rapid improvement of the hardware performance of mobile terminals (such as smart phones, tablet computers), the graphics rendering system based on the Android operating system (such as the Unreal Engine engine in games, AR / VR applications, etc.) has become the core carrier supporting high-quality mobile graphics applications. Such systems usually contain rich configurable parameters, covering multiple dimensions such as hardware environment (such as CPU core number, memory size), graphics rendering pipeline (such as resolution, shadow switch, anti-aliasing strategy), and scene content (such as texture quality, particle effect quantity), and different parameter combinations have a direct and significant impact on rendering quality, running smoothness, and resource consumption.

[0003] In actual applications, in order to balance the rendering effect and system performance, it is necessary to accurately evaluate the influence of configuration parameters on key performance indicators - these indicators include not only graphics processing core indicators (such as DrawCall call times, GPU utilization, and video memory read-write volume), but also system resource usage indicators (such as CPU usage, and memory occupancy average value). However, in a high-dimensional configuration space, the way of traversing all parameter combinations to obtain performance data through actual testing is extremely costly and inefficient, so establishing a mapping model between configuration parameters and performance indicators becomes a key means for performance prediction and optimization.

[0004] In the prior art, although machine learning methods have been tried to construct configuration-performance prediction models, there are still significant limitations in the mobile graphics rendering scene: first, most methods focus on single performance indicator prediction, which is difficult to meet the demand for multi-indicator parallel prediction, and in actual rendering systems, indicators such as GPU utilization, CPU load, and memory consumption often balance each other, and a single model is difficult to consider the prediction accuracy of multiple targets; second, the graphics rendering configuration parameters contain a large number of discrete (such as anti-aliasing strategy) and mixed features, and traditional modeling methods have insufficient expression ability for high-dimensional features, which easily leads to problems such as coarse modeling granularity and low sample utilization.

[0005] In view of the specific needs of mobile terminal graphics rendering systems, there is an urgent need for a modeling method that can efficiently process high-dimensional configuration features, support multi-performance indicator parallel prediction, and have good interpretability, in order to reduce the dependence on large-scale actual testing, improve the accuracy and generalization ability of performance prediction, and provide reliable technical support for configuration optimization of graphics rendering systems. SUMMARY

[0006] The present application aims to solve the problems in the prior art, including: it is difficult to model the mapping relationship between high-dimensional configuration parameters and multiple performance indicators, and a single model cannot take into account the prediction accuracy of multiple indicators; traditional methods have insufficient expression ability for discrete and mixed configuration characteristics, and have problems such as coarse modeling granularity and low sample utilization.

[0007] The present application provides a graphics rendering performance prediction method and device based on configuration space mapping, which realizes accurate prediction of multiple performance indicators of a graphics rendering system by constructing an end-to-end model that integrates feature enhancement, rule extraction and multi-task prediction, and provides a reliable basis for configuration parameter optimization.

[0008] In a first aspect, the present application provides a graphics rendering performance prediction method based on configuration space mapping, which realizes parallel accurate prediction of multiple performance indicators by systematically sampling configuration-performance data and constructing a multi-task prediction model that integrates a Transformer and an ANFIS, including the following steps:

[0009] Multi-dimensional configuration and original data collection: collect key configuration parameters and performance indicator data of a graphics rendering system, and construct an original dataset of configuration-performance relationships;

[0010] Data cleaning and preprocessing: clean and normalize the original dataset;

[0011] Performance model training: based on the normalized dataset, construct an end-to-end prediction model, output parallel prediction results of performance indicators, and the training steps include feature mapping, feature encoding, rule extraction, multi-task prediction network construction and joint optimization training.

[0012] In a second aspect, the present application provides a graphics rendering performance prediction device based on configuration space mapping, which comprises:

[0013] A data collection module for collecting key configuration parameters and performance indicator data of a graphics rendering system, and constructing an original dataset of configuration-performance relationships;

[0014] A data preprocessing module for cleaning and normalizing the original dataset to obtain a preprocessed dataset;

[0015] A model training module for constructing an end-to-end prediction model based on the normalized dataset, and outputting parallel prediction results of performance indicators, the model training module comprising:

[0016] Feature encoding unit: encode the sequence input through the Transformer encoder, extract the high-order nonlinear representation of the configuration feature, obtain the encoded feature, and enhance the feature expression ability;

[0017] Rule extraction unit: fuzzy partitioning of encoded features by ANFIS, determination of fuzzy subsets by mean shift clustering and construction of membership functions, calculation of rule activation strength to obtain a matrix, and improvement of model interpretability;

[0018] Multi-task prediction network construction unit: based on encoded features and rule activation strength, construct fusion features, input multi-task network for parallel prediction, the multi-task network includes shared layer and task layer;

[0019] Joint optimization training unit: using Adam optimizer, updating Transformer and multi-task network parameters through back propagation, taking task-adaptive weighted mean square error as loss function, and setting early stopping mechanism to minimize multi-index prediction error.

[0020] The beneficial effects of the present application are:

[0021] 1. Support for parallel prediction of multiple performance indicators: through the multi-task network structure, the parallel prediction of multiple performance indicators such as GPU utilization, CPU usage and memory consumption is realized, and the problem that a single model cannot balance multiple target accuracy is solved;

[0022] 2. Strong feature expression ability: introducing Transformer encoder to extract high-order nonlinear features of configuration parameters, effectively processing high-dimensional and mixed configuration features, improving modeling granularity and sample utilization;

[0023] 3. High interpretability: fuzzy rule extraction by ANFIS, quantifying the correlation between configuration features and performance indicators through membership functions and rule activation strength, solving the problem of poor interpretability of traditional clustering methods;

[0024] 4. High prediction accuracy and generalization ability: through the joint optimization strategy of Transformer and multi-task network, combined with early stopping mechanism and adaptive loss weighting, the multi-index prediction error is significantly reduced, and the generalization ability of the model in different configuration scenarios is improved;

[0025] 5. Strong engineering practicability: reduce the dependence on large-scale actual tests, and provide efficient technical support for configuration parameter optimization (such as balancing rendering quality and smoothness) of mobile terminal graphics rendering system through accurate performance prediction. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A flowchart of a graphics rendering performance prediction method based on configuration space mapping in the embodiments of the present application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings.

[0028] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide the public with a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0029] The method disclosed in one embodiment of this invention is applied to a graphics rendering system deployed on a mobile terminal (e.g., a smartphone, tablet). The graphics rendering system is based on the Android operating system, and its core component is a graphics rendering engine, which is packaged in an application package (APK) for execution. In this embodiment, Unreal Engine (UE) is used as an example of the rendering engine, but the scope of protection of this invention is not limited thereto. The graphics rendering system includes a series of configurable parameters, the adjustment of which directly affects the system's operating performance, resource consumption, and rendering quality. These parameters may specifically include, but are not limited to: graphics resolution, shadow quality, lighting accuracy, material complexity, anti-aliasing strategy, and the number of rendering threads. The key to this invention lies in the systematic sampling, modeling, and prediction of the mapping between configuration parameters and performance indicators.

[0030] For the sake of standardization of notation, let: the number of configuration items be N; the number of performance indicators be M; and the total number of samples be K.

[0031] like Figure 1 As shown, in a preferred embodiment of the present invention, a method for predicting graphics rendering performance based on configuration space mapping is provided. The following steps S1 to S3 describe data acquisition, preprocessing, and prediction models. The specific implementation steps are as follows:

[0032] S1: Raw Data Acquisition for Multi-Dimensional Configuration and Performance: In this step, in order to build a configuration-performance mapping model, it is necessary to systematically sample the key configuration parameters and performance indicators of the target graphics rendering system (in this embodiment, the Unreal Engine engine of the Android system is taken as the specific object) to construct the raw configuration-performance dataset D.

[0033] Furthermore, the training process in step S1 includes:

[0034] S11: Selection and collection of configuration parameter dimensions. To ensure the diversity and comprehensiveness of the data, the selected configuration parameter dimensions cover hardware, graphics rendering, and scene complexity, including:

[0035] Hardware-related configurations: to respond to different mobile terminal hardware environments, parameters such as CPU core number (e.g. 6 cores, 8 cores, 12 cores), memory size (e.g. 2GB, 4GB, 8GB);

[0036] Graphics rendering-related configurations: core parameters related to rendering pipeline, parameters such as whether to enable shadow, resolution setting (e.g. 720p, 1080p, 4K), anti-aliasing strategy (such as None, FXAA, TemporalAA and MSAA);

[0037] Material and scene complexity: scene content parameters that affect rendering load, parameters such as map quality level (e.g. 256, 512, 1024, 2048), number of particle effects (e.g. 0, 50, 200), etc.

[0038] S12: Selection and collection of performance indicators: to comprehensively evaluate the system state, the selected performance indicators cover two dimensions of graphics processing and system resources, including:

[0039] Graphics processing-related indicators: such as DrawCall method call times, GPU active time, GPU utilization, video memory write size, and video memory read size;

[0040] System resource usage indicators: such as average runtime memory usage and average CPU usage.

[0041] Among them, the following tools are used for performance data collection:

[0042] Snapdragon Profiler is used to obtain GPU indicators (such as GPU utilization, GPU active time, video memory write size, video memory read size, etc.);

[0043] The top command is used to obtain system-level indicators such as CPU usage and memory usage, and scripts are used to sample at regular intervals to calculate the average value;

[0044] UE Profiler is used to obtain DrawCall method call times and other runtime information.

[0045] S13: Structured representation of data set: after the above sampling process, each combination of configuration parameters and its corresponding performance indicator observation value constitutes a sample record. The original configuration-performance data set D can be formally represented as:

[0046]

[0047]

[0048] wherein represents the configuration item vector of the i-th sample, represents the corresponding performance indicator vector of the i-th sample; represents the specific value of the j-th configuration item of the i-th sample, represents the specific value of the j-th configuration item of the i-th sample, represents the observed value of the m-th performance indicator of the i-th sample. This data structure will serve as the input basis for the subsequent modeling process. S2: Data cleaning and preprocessing: In this embodiment, the original data set D collected in step S1 is read, cleaned and preprocessed to obtain the preprocessed data set .

[0049] Further, in the embodiments of the present application, the data cleaning process includes the following sub-steps:

[0050] (1) Missing value processing, deleting samples containing missing values (such as NaN, empty string) or incomplete records;

[0051] (2) Category variable encoding: for category configuration parameters (such as anti-aliasing algorithms), label encoding (LabelEncoding) technology is used to convert them into numerical type representation, so as to facilitate model processing;

[0052] (3) Outlier identification and elimination: IQR method is used to identify outliers and eliminate or truncate for numerical fields;

[0053] (4) De-duplication: remove identical sample records to ensure the uniqueness of each data.

[0054] In order to eliminate the dimensional differences between different performance indicators and configuration parameters (for example, the dimension of memory usage is MB, while the dimension of CPU usage is percentage), and accelerate the convergence process of model training, normalization processing needs to be performed on all numerical fields. In this embodiment, the Min-Max Normalization method is used to process all numerical fields, and all numerical features are scaled to the interval [0, 1]. The normalized data is denoted as .

[0055] S3: Training of performance prediction model. According to the embodiments of the present application, the training step S3 of the performance prediction model aims to train the performance prediction model based on the normalized data set .

[0056] ​The process of building an end-to-end prediction model includes feature mapping, encoding, rule extraction, multi-task network and joint optimization, with the goal of outputting parallel prediction results for M performance indicators.

[0057] Furthermore, the training process in step S3 includes:

[0058] S31: Feature Mapping: Adapting the cleaned original samples to the model input dimensions to generate a model input with a unified format. Let the normalized feature matrix be... (Each row is a sample) The preprocessed feature matrix will be projected through the input projection layer of the Transformer. ∈ (N is the number of configuration items) The input is mapped to the model dimension d to obtain .

[0059]

[0060] in and For trainable parameters, subsequently, to match the sequence input format of the Transformer encoder, a sequence of length 1 can be added to the feature dimension to reshape Z into shape [K, 1, d].

[0061] S32: Feature Encoding: The sequence input is encoded using a Transformer encoder to extract a high-order nonlinear representation of the configuration features, resulting in encoded features. To enhance feature representation capabilities, the specific steps are as follows:

[0062] (1) Initialization phase: During model initialization, freeze the Transformer gradient (e.g., using torch.no_grad()) and perform forward computation on all samples at once to obtain the initial encoding. As the input basis for ANFIS rule partitioning.

[0063] (2) Training phase: Unfreeze the Transformer parameters, and obtain the encoded features through the Transformer encoder during batch training:

[0064]

[0065] in These are the trainable parameters of the Transformer, which is responsible for extracting a high-order nonlinear representation of the configuration vector, thereby enhancing the feature representation capability.

[0066] S33: Rule extraction: fuzzy partitioning of the Transformer encoded features by ANFIS and generating rule items, determining fuzzy subsets by mean shift clustering and constructing membership functions, calculating rule activation strength to obtain matrix (T is the number of rules), to improve the interpretability of the model and the modeling ability of "similar features correspond to similar performance", the specific operation is as follows:

[0067] (1) Fuzzy partitioning: taking the encoded feature set } as input, using the built-in clustering algorithm of ANFIS (such as mean shift clustering) to determine the fuzzy subsets, and constructing membership functions for each subset, the default number of rules is denoted as T (for example, T=2 is an example, which can be increased or decreased as needed).

[0068] (2) Rule activation strength: calculate the combined membership of each rule t and sample i , and normalize to obtain the rule activation strength:

[0069]

[0070]

[0071]

[0072] Where matrix represents the rule activation of all samples, represents the combined membership of the t-th rule to the i-th sample, represents the membership of the t-th rule to the j-th dimension of the i-th sample, which is determined by the Gaussian membership function, and are obtained by the internal clustering method of ANFIS.

[0073] S34: Multi-task prediction network construction: based on encoded features and rule activation strength , construct fusion features (expand the encoded vector and weightedly splice all rule features), and input the multi-task prediction network for parallel prediction, the multi-task prediction network includes shared layers (extract common information of fusion features) and task layers (set independent prediction heads for each performance indicator), as follows:

[0074] Fusion feature construction includes:

[0075] 1. For each sample i, expand the encoded vector to include a constant bias term: ;

[0076] 2. For each rule t, use rule activation strength for weighting: ;

[0077] 3. Concatenate the weighted features of all rules to obtain a fusion feature vector .

[0078] In the embodiments of the present application, the multi-task prediction network structure is implemented by using a MultiTaskModel class and includes:

[0079] The shared layer includes: wherein W represents a weight matrix of the shared layer, and b represents a bias vector of the shared layer.

[0080] The task layer includes a linear head for the mth performance indicator .

[0081] S35: Joint optimization training: using an Adam optimizer, synchronously updating the parameters of the Transformer and the multi-task prediction network through back propagation, taking the task-adaptive weighted mean square error (MSE) as the loss function (the weight factor is dynamically adjusted based on the performance indicator standard deviation), and setting an early stopping mechanism (terminating training when the validation set loss change rate is less than 0.001), minimizing the multi-indicator performance prediction error, and obtaining a final performance prediction model. The specific process is as follows:

[0082] Loss function design: for the mth performance indicator, the mean square error is defined as , and the total loss is obtained by task-adaptive weighting: ; wherein w represents a weight factor, and , .

[0083] Parameter optimization strategy: the optimizer uses Adam, and the optimization parameter set includes the parameters of the Transformer and the parameters of the multi-task network.

[0084] wherein each training step performs:

[0085] 1. Extract batch data from the training set;

[0086] 2. Encode the features (made in S31, S32);

[0087] 3. Calculate the rule activation strength (made in S33);

[0088] 4. Construct the fusion feature and input the multi-task network to obtain the predicted value;

[0089] 5. Calculate the total loss and perform back propagation to update the parameters.

[0090] Early stopping mechanism: when the validation set loss change rate is less than 0.001, terminate training in advance and save the optimal model parameters. ​

[0091] In the embodiment of the present application, through the above steps, an end-to-end model capable of simultaneously predicting multiple performance indicators is finally obtained, wherein the Transformer is responsible for feature enhancement, the ANFIS provides rule prior, the multi-task network realizes parallel prediction, and the three are optimized in cooperation to improve the prediction accuracy.

[0092] Table 1 data description

[0093]

[0094] In the embodiment of the present application, Table 1 shows the detailed features of the present application.

[0095] In Table 1, the first column is the configuration item name contained in the system, the second column is the average value of the configuration item value, the third column is the standard deviation of the configuration item value, and the fourth, fifth and sixth columns are the quartiles of the configuration item value.

[0096] Further, in order to verify the technical effect of the technical scheme of the present application, two widely recognized evaluation indexes are selected: the average relative error and the average absolute error.

[0097] MRE: The average relative error (MRE) is used to reflect the proportion of the prediction error relative to the true value, and is not affected by the absolute value of the performance indicator. The calculation formula is as follows:

[0098]

[0099] wherein, represents the number of test sets in the performance prediction data, represents the actual performance value of the i-th data, represents the predicted performance value of the i-th data, is a minimum value, in this example, a negative sixth of the smallest non-zero absolute value or a tenth.

[0100] MAE: The average absolute error (MAE) is used to reflect the average deviation between the predicted value and the true value, and retains the original dimension of the performance indicator. The calculation formula is as follows:

[0101]

[0102] wherein, represents the number of test sets in the performance prediction data, represents the actual performance value of the i-th data, represents the predicted performance value of the i-th data.

[0103] In this embodiment, experimental research is selected to compare the present application with two baselines to prove the effectiveness of the method proposed in the present application. Their characteristics are as follows:

[0104] (1)Multi RSFIN: Use ANFIS network to obtain the relationship between features and prediction targets, and use MLPRegressor to realize prediction of multiple targets.

[0105] (2)RSFIN+neural network: On the basis of Multi RSFIN, replace MLPRegressor with a multi-task network to realize prediction of multiple targets.

[0106] In the embodiment of the application, the experimental results are shown in Table 2. The experimental results show that the method proposed in the application has achieved significant improvement in multi-target prediction.

[0107] Performance comparison of the method in Table 2 and the baseline method

[0108]

[0109] Based on the same concept of the above method embodiment, the embodiment of the application also proposes a graphics rendering performance prediction device based on configuration space mapping, characterized by comprising:

[0110] A data acquisition module is configured to select hardware configuration, graphics rendering configuration and material scene complexity configuration parameters of a target graphics rendering system, synchronously acquire graphics processing indicators and system resource indicators, and construct an original configuration-performance dataset D;

[0111] A data preprocessing module is configured to clean and normalize the original dataset to obtain a preprocessed dataset ;

[0112] A model training module is configured to construct an end-to-end prediction model based on the normalized dataset, output parallel prediction results of performance indicators, and the model training module comprises:

[0113] A feature mapping unit is configured to map input to model dimensions through a Transformer input projection layer, and reshape the input into a sequence input format to adapt to a Transformer encoder;

[0114] A feature encoding unit is configured to encode the sequence input through the Transformer encoder, extract high-order nonlinear representations of configuration features, obtain encoded features, and enhance feature expression capability;

[0115] A rule extraction unit is configured to perform fuzzy division on the encoded features by means of ANFIS, determine fuzzy subsets and construct membership functions through mean shift clustering, calculate rule activation strength to obtain a matrix, and improve model interpretability;

[0116] The multi-task prediction network construction unit: based on the coding features and the rule activation strength, the fusion features are constructed, and the multi-task network is input for parallel prediction, the multi-task network includes a shared layer and a task layer;

[0117] The joint optimization training unit: using the Adam optimizer, synchronously updating the parameters of the Transformer and the multi-task network through back propagation, taking the task adaptive weighted mean square error as the loss function, and setting the early stopping mechanism, the multi-index prediction error is minimized.

[0118] The above-described embodiments are only a preferred scheme of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical solutions obtained by equivalent replacement or equivalent transformation shall fall within the protection scope of the present application.

Claims

1. A method for predicting graphics rendering performance based on configuration space mapping, the method comprising: receiving a configuration space map; receiving a graphics workload; and predicting graphics rendering performance based on the configuration space map and the graphics workload. The method comprises the following steps: Multi-dimensional configuration and original data collection: collecting key configuration parameters and performance index data of a graphics rendering system, and constructing an original dataset of configuration-performance relationship; Data cleaning and preprocessing: cleaning and normalizing the original dataset; Performance model training: constructing an end-to-end prediction model based on the normalized dataset, outputting parallel prediction results of performance indexes, and the training steps include feature mapping, feature coding, rule extraction, multi-task prediction network construction and joint optimization training; The feature mapping: mapping the input to the model dimension through the Transformer input projection layer, and reshaping it into a sequence input format to adapt to the Transformer encoder; The feature coding: coding the sequence input through the Transformer encoder, extracting high-order nonlinear representation of configuration features, obtaining coded features, and enhancing feature expression ability; The rule extraction: fuzzy division of coded features by ANFIS, determination of fuzzy subsets by mean shift clustering and construction of membership functions, calculation of rule activation strength to obtain a matrix, and improvement of model interpretability; The calculation of rule activation strength specifically is: calculating the combined membership degree of each rule and sample, and normalizing to obtain the rule activation strength; The multi-task prediction network construction: based on the coded features and the rule activation strength, constructing a fusion feature, inputting the multi-task network for parallel prediction, and the multi-task network includes a shared layer and a task layer; Wherein the fusion feature is constructed by: For each sample, the coding vector is expanded to include a constant bias term; Each rule uses rule activation strength for weighting; Concatenate the weighted features of all rules to obtain a fusion feature vector; The joint optimization training: using the Adam optimizer, synchronously updating the parameters of the Transformer and the multi-task network through back propagation, taking the task-adaptive weighted mean square error as the loss function, and setting an early stopping mechanism to minimize the multi-index prediction error.

2. The method of claim 1, wherein, The configuration parameter dimension selection and collection, the selection of the configuration parameter dimension includes hardware-related configurations, graphics rendering-related configurations, and material and scene complexity.

3. The method of claim 1 or 2, wherein, The performance index is selected and collected, and the selection of the performance index includes: Graphics processing related indexes: including DrawCall method call times, GPU active time, GPU utilization, video memory write size and video memory read size; System resource usage indexes: including runtime memory usage average and CPU usage average; The performance index collection includes: Using Snapdragon Profiler to obtain GPU indexes; Using the top command to obtain CPU usage and memory usage system-level indexes, and calculating the average value by scripting timing sampling; Using UE Profiler to obtain DrawCall method call times.

4. The method of claim 1, wherein, The data cleaning includes: Missing value processing: deleting samples containing missing values or incomplete records; Category variable coding: using label coding for category type parameters to realize numerical conversion; Outlier identification and elimination: IQR method is used to identify outliers and eliminate or truncate for numerical fields; De-duplication: remove the same sample records.

5. The method of claim 4, wherein, Max-min normalization is performed on the numerical fields to eliminate dimension differences and accelerate model convergence.

6. The method of claim 1, wherein, The feature encoding includes: Initialization phase: freeze the Transformer gradient and forward calculate all samples at a time to get the initial encoding as the input basis for ANFIS rule division; Training phase: unfreeze the Transformer parameters, and get the encoding features through the Transformer encoder in batch training.

7. The method of claim 1, wherein, The graph rendering performance prediction uses MRE and MAE as loss functions to jointly optimize the Transformer and multi-task network parameters to realize multi-index parallel prediction.

8. A graphics rendering performance prediction device based on configuration space mapping, characterized in that, It includes: A data acquisition module for acquiring key configuration parameters and performance index data of a graphics rendering system and constructing an original dataset of configuration-performance relationship; A data preprocessing module for cleaning and normalizing the original dataset to obtain a preprocessed dataset; A model training module for constructing an end-to-end prediction model based on the normalized dataset and outputting parallel prediction results of performance indicators; The training steps include feature mapping, feature encoding, rule extraction, multi-task prediction network construction, and joint optimization training; The feature mapping: the input is mapped to the model dimension through the Transformer input projection layer, and reshaped into a sequence input format to adapt to the Transformer encoder; The feature encoding: the sequence input is encoded by the Transformer encoder to extract high-order nonlinear representation of configuration features, obtain encoding features, and enhance feature expression ability; The rule extraction: ANFIS is used to fuzzy divide the encoding features, determine the fuzzy subsets through mean shift clustering, construct the membership function, calculate the rule activation strength to obtain the matrix, and improve the model interpretability; The calculation of rule activation strength specifically is: calculating the combined membership of each rule and sample, and normalizing to obtain the rule activation strength; The multi-task prediction network construction: based on the encoding features and rule activation strength, construct the fusion features, input the multi-task network for parallel prediction, and the multi-task network includes shared layer and task layer; Wherein the construction of the fusion features includes: For each sample, extend the encoding vector to include a constant bias term; Each rule uses rule activation strength for weighting; Concatenate the weighted features of all rules to get the fusion feature vector; The joint optimization training: use Adam optimizer, update the Transformer and multi-task network parameters synchronously through back propagation, take the task adaptive weighted mean square error as the loss function, set the early stopping mechanism, and minimize the multi-index prediction error.

Images