Embedded tablet touch optimization method and system based on heterogeneous computing

Abstract

The invention relates to the technical field of computers, and discloses an embedded tablet touch optimization method and system based on heterogeneous computation.The embedded tablet touch optimization method based on heterogeneous computation.The method comprises the steps that information entropy values of different touch operations are calculated, a touch path Markov model is established, and the information entropy values of the different touch operations are calculated; identifying importance and touch state transition probabilities of different touch operations; calculating a user intention duration probability, and selecting a corresponding touch state storage strategy according to the user intention duration probability; distributing the touch processing data into heterogeneous memories of different levels, and optimizing a cache replacement strategy; intelligently distributing a touch control processing task to a heterogeneous computing unit; a differentiated dormancy strategy and a rapid wakeup mechanism of the heterogeneous computing unit are realized; according to the method, the touch response performance is improved, the interaction experience coherence is enhanced, the system resource utilization efficiency is optimized, and the method is suitable for embedded tablet equipment with limited computing resources, changeable application scenes and complex user use modes.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically, to an embedded tablet touch optimization method and system based on heterogeneous computing. Background Art

[0002] With the wide application of embedded tablet devices, users' requirements for touch response performance, interaction fluency, and battery life are increasing day by day, but there are many deficiencies in the existing technologies.

[0003] Existing embedded tablet devices face the following technical dilemmas in touch processing: The computing and memory resources of the device are limited and cannot meet the requirements of touch accuracy and response speed simultaneously; The touch behavior characteristics vary significantly in different application scenarios, and the fixed resource allocation strategy cannot adapt to this diversity; When the user briefly leaves and then quickly returns to use, the frequent complete sleep-wake cycles result in a fragmented touch experience; The heterogeneous computing environment fails to perform intelligent resource scheduling according to the characteristics of touch tasks.

[0004] Existing technologies generally adopt static cache management strategies and fixed computing resource allocation methods, without considering the importance differences between touch operations, the persistence of user intentions, and the conversion characteristics of application scenarios. Traditional technologies usually use a unified cache replacement strategy, which cannot distinguish between frequently occurring but simple operations and occasionally occurring but critical complex operations; When the user switches applications, due to the lack of a predictive cache preheating mechanism, the cache hit rate often drops significantly, and the touch response latency increases; Existing technologies usually adopt a complete sleep strategy when the user temporarily leaves the device, ignoring the persistence of user intentions, resulting in the need to reload all resources when the user returns, and the experience is not coherent. Summary of the Invention

[0005] The present invention provides an embedded tablet touch optimization method and system based on heterogeneous computing, which solves the technical problems in the related technologies that cannot meet the requirements of touch accuracy and response speed simultaneously, and the heterogeneous computing environment fails to perform intelligent resource scheduling according to the characteristics of touch tasks.

[0006] The present invention provides an embedded tablet touch optimization method based on heterogeneous computing, including the following steps: Calculate the information entropy values of different touch operations, establish a touch path Markov model, and identify the importance of different touch operations and the touch state transition probabilities; Based on the touch operation entropy values and the state transition probabilities of the Markov model, combined with the operation type, application type, time pattern, and environmental factors before the user's touch interruption, calculate the user intention persistence probability, and accordingly select the corresponding touch state saving strategy; Cache allocation is performed based on the entropy value of touch operations and the predicted path probability, and touch processing data is allocated to heterogeneous memories at different levels, and the cache replacement policy is optimized; According to the entropy value of touch operations, the intention persistence probability, the cache allocation result, and the characteristics of the computing task, the touch processing tasks are intelligently allocated to heterogeneous computing units; Based on the user intention persistence probability, the computing task allocation status, and the environmental perception data, a differential sleep strategy and a fast wake-up mechanism for heterogeneous computing units are implemented.

[0007] In a preferred embodiment, the steps of calculating the information entropy value of different touch operations and establishing a touch path Markov model include: Obtain the original touch data stream from the touch sensor module; For each type of touch operation, collect and analyze its historical state distribution, and calculate its information entropy value; Build a touch path Markov model for different application types and calculate the conditional transition probability; Adopt an incremental learning method to update the entropy value and Markov model parameters regularly.

[0008] In a preferred embodiment, the steps of calculating the user intention persistence probability include: Extract features related to user intention persistence from multiple sources to form a feature vector; Use a fusion model to calculate the intention persistence probability; Divide the intention persistence probability into multiple intervals corresponding to different persistent intention intensity levels; Select the optimal state saving strategy according to the intention persistence probability and the touch state complexity.

[0009] In a preferred embodiment, the heterogeneous memory includes: L1 layer: including GPU shared memory and CPU L1 cache; L2 layer: including CPU L2 cache and part of GPU local memory; L3 layer: including main memory.

[0010] In a preferred embodiment, the cache allocation based on the entropy value of touch operations and the predicted path probability includes: Allocate the data related to high-entropy value operations to the L1 layer for storage; Allocate the data related to medium-entropy value operations to the L2 layer for storage; Allocate the data related to low-entropy value operations to the L3 layer for storage; When the cache space is insufficient and data needs to be replaced, replace the cache item with the lowest retention value according to the retention value first.

[0011] In a preferred embodiment, the step of intelligently allocating touch processing tasks to heterogeneous computing units includes: Analyze the characteristics of the touch processing tasks and calculate the load feature vector; Model the performance characteristics of the heterogeneous computing resources in the device; Based on the next operation probability distribution predicted by the Markov model, perform predictive task allocation and scheduling; Real-time monitor the load conditions of each computing unit, and when detecting load imbalance, perform dynamic adjustment.

[0012] In a preferred embodiment, the steps of implementing the differentiated sleep strategy and fast wake-up mechanism for heterogeneous computing units include: Based on the continuous probability of user intention, formulate a hierarchical sleep strategy for heterogeneous computing units; Build a lightweight low-power environment perception network to continuously monitor the signs of user return; Based on the output of the environment perception network, implement a condition-triggered hierarchical wake-up mechanism; Analyze the historical wake-up data to achieve intelligent wake-up pipeline optimization.

[0013] In a preferred embodiment, the hierarchical sleep strategy includes: Shallow sleep: Keep the core data structures and states in memory and reduce the clock frequency; Medium sleep: Write the core states to fast storage, turn off most functional units, and only retain the monitoring function; Deep sleep: Write the complete states to non-volatile storage and turn off all units except the lowest-power monitoring module.

[0014] In a preferred embodiment, the touch operations in the touch operation entropy value, intention continuous probability, cache allocation result, and computing task characteristics include: fine line drawing, area selection, eraser operation, layer switching, color selection, canvas panning and zooming.

[0015] In a preferred embodiment, an embedded tablet touch optimization system based on heterogeneous computing is used to execute an embedded tablet touch optimization method based on heterogeneous computing, including: A touch operation entropy calculation module, which is used to calculate the information entropy values of different touch operations, establish a touch path Markov model, and identify the importance of different touch operations and the touch state transition probability; An intention persistence evaluation module, which is used to calculate the continuous probability of user intention based on the touch operation entropy value and the state transition probability of the Markov model, combined with the operation type, application type, time pattern, and environmental factors before the user touch interruption, and select the corresponding touch state saving strategy accordingly; The heterogeneous memory allocation module is used to allocate touch processing data to heterogeneous memories at different levels based on the touch operation entropy value, state transition probability, and user intention persistence probability, and optimize the cache replacement policy; The task scheduling module is used to intelligently allocate touch processing tasks to heterogeneous computing units according to the touch operation entropy value, intention persistence probability, cache allocation result, and computing task characteristics; The sleep-wake control module is used to implement the differential sleep strategy and fast wake-up mechanism of heterogeneous computing units based on the user intention persistence probability, computing task allocation status, and environmental perception data.

[0016] The beneficial effects of the present invention are as follows: The touch response performance is improved, the response time fluctuation of key touch operations is reduced, the average response time is decreased, and the accuracy and smoothness of user interaction are enhanced; The utilization efficiency of system resources is improved, the memory usage efficiency is increased, and the hierarchical sleep mechanism based on intention persistence evaluation reduces the system energy consumption, effectively extending the device usage time; The coherence of the interaction experience is enhanced, the touch response delay in the application switching scenario is reduced, and the recovery response time when the user returns to use is shortened, improving the coherence of the usage experience; The adaptability of the system to different users and application scenarios is significantly improved. The model parameter adaptive update mechanism enables the system to automatically adapt to the touch habits of different users and the requirements of application scenarios, and the overall touch interaction satisfaction score is increased; The dynamic balance between resource consumption and response performance is achieved. Through the intelligent task allocation and hierarchical sleep strategy of heterogeneous computing units, the system can flexibly adjust resource allocation according to the current scenario and user needs, reducing unnecessary energy consumption while maintaining high responsiveness. Description of the Drawings

[0017] Figure 1 is a flowchart of an embedded tablet touch optimization method based on heterogeneous computing of the present invention. Detailed Embodiments

[0018] Now, the subject matter described herein will be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the scope of protection of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0019] At least one embodiment of the present invention discloses an embedded tablet touch optimization method based on heterogeneous computing, asFigure 1 As shown, it includes the following steps: Step 1: Calculate the information entropy values of different touch operations, establish a Markov model of the touch path, and identify the importance of different touch operations and the touch state transition probabilities; Specifically, it includes the following steps: Step 1.1: Touch operation entropy calculation; Use the information entropy theory to quantify the importance of different touch operations (such as clicks, swipes, zooms, etc.). The system first obtains the original touch data stream from the touch sensor module and then performs the following operations: For each type of touch operation , collect and analyze its historical state distribution. The state can include characteristics such as the duration of the operation, the number of contacts involved, and the touch pressure. The state set is represented as . Among them, represents the set of all possible states, which contains different states; , , respectively represent the , , th states, represents the total number of states.

[0020] Calculate the probability of each state appearing under each touch operation , and then apply the entropy calculation formula: ; Among them, represents the information entropy of the touch operation , which is used to quantify the uncertainty and complexity of the touch operation; represents the probability of the state appearing under the operation , reflecting the distribution of various states under a specific touch operation; represents the summation over all possible states ; represents the logarithmic function, which is used to convert the probability into information content; indicates that the entropy value is non - negative. The higher the entropy value, the greater the uncertainty of the touch operation, and more computing resources need to be allocated for processing.

[0021] Through this calculation, a high entropy value indicates that the operation has high uncertainty and information content, usually corresponding to key but infrequent operations; a low entropy value indicates that the operation has low uncertainty, usually corresponding to simple but frequent operations.

[0022] Step 1.2, Construction of the Touch Path Markov Model; This process constructs a touch path Markov model for different application types to predict the possible sequences of user touch operations. The specific steps are as follows: For each application type (such as games, document editing, image processing, etc.), collect the sequence of touch states of the user in this application , where , , respectively represent the , , th touch states executed by the user in this application in chronological order; represents the total number of collected touch states.

[0023] Based on the collected sequence, calculate the conditional transition probability , which represents the probability that in application , the current state transfers to the next state .

[0024] Construct the state transition matrix , where the matrix element represents the probability of transferring from state to state in application .

[0025] Finally, each application type corresponds to a Markov model, which can be represented as a triple , where is the set of states, is the transition probability matrix, is the initial state distribution.

[0026] In this application, the touch path Markov model has the following structure and characteristics: State representation method: Each touch state consists of multiple feature dimensions, including touch type (such as click, long press, swipe, multi-finger operation, etc.), relative screen position area (divide the screen into grids), touch pressure range, gesture duration, etc.

[0027] Calculation of transition probability: The system calculates the transition probability based on a large amount of user interaction data. For user groups with similar operation patterns, the data merging method can be used to improve the reliability of the model.

[0028] The calculation of the transition probability is based on the observation sequence within a sliding time window: ; wherein, represents the conditional transition probability that in the application , when the current state is , the next state is ; represents the number of times of observing the transition from state to state in the application ; represents the total number of times of transitioning from state to any state in the application , that is, the cumulative number of all transition events with state as the starting state; represents any possible state in the state space.

[0029] Application scenario adaptability: The system maintains independent Markov models for different types of applications. For example, in a drawing application, a precise touch operation is usually followed by a tool selection operation; while in a reading application, a page sliding operation is often followed by a reading pause state. Such application-related models can more accurately reflect the user behavior patterns in specific scenarios.

[0030] Model order selection: According to the actual touch interaction characteristics, the system can flexibly select a first-order Markov model (only considering the current state) or a higher-order Markov model (considering multiple previous states). In practical applications, for different application scenarios, the system can dynamically adjust the model order to balance the prediction accuracy and computational complexity.

[0031] Step 1.3, Adaptive update of model parameters; To ensure the model accuracy, the system uses an incremental learning method to update the model parameters regularly: For the calculation of touch operation entropy, the system adopts a sliding window method to update the operation state probability distribution : ; wherein, is the smoothing factor, with a value range of 0 - 1, used to control the weight ratio of new and old data; is the probability of the historical calculation of state under the condition of operation ; is the probability distribution of the newly observed state under the condition of operation ; is the updated probability of state under the condition of operation when When it is close to 1, the model tends to retain historical data; when it is close to 0, the model tends to adopt newly observed data.

[0032] For the Markov model, the system updates the transition probability matrix through a similar method to ensure that the model can adapt to changes in user touch behavior.

[0033] The output result of this step is: The entropy value mapping table for each touch operation , where represents a specific touch operation type (such as click, slide, zoom, etc.); represents the information entropy value corresponding to the touch operation, quantifying the uncertainty and information content of the operation; this mapping table establishes the correspondence between the touch operation type and its importance (represented by the entropy value).

[0034] The touch path Markov model for each application type , where represents a specific application type; represents the state set, containing all possible touch states; represents the transition probability matrix, describing the probability of transitioning from one touch state to another; represents the initial state distribution, describing the probability distribution of each touch state when the user starts interacting in the application.

[0035] These data structures will be used as the input for subsequent steps in touch resource optimization decisions, helping the system allocate resources based on the importance of operations and user behavior patterns.

[0036] Step 2: Based on the entropy value of touch operations and the state transition probability of the Markov model, combined with the operation type, application type, time pattern, and environmental factors before the user touch interruption, calculate the user intention persistence probability, and accordingly select the corresponding touch state saving strategy; Specifically, it includes the following steps: Step 2.1: Intention persistence feature extraction; This process extracts features related to user intention persistence from multiple sources: Extract features from the operation type before touch interruption Extract features: The system records attributes such as the type, duration, and completion degree of the user's last executed touch operation to form a feature vector . For example, an unfinished fine drawing operation has a higher probability of persistent intention than a completed page scrolling operation.

[0037] Extract from the application type Feature extraction: Different application types have different usage continuity characteristics. The system classifies applications (such as creative, browsing, gaming, etc.) and extracts application status information to form a feature vector. For example, the unsaved content in a document editing application indicates that the user may return soon.

[0038] Feature extraction from the time pattern Feature extraction: Analyze the time pattern of the user's device usage, including the typical session duration, the frequency and duration of short interruptions, etc., to form a feature vector. The system learns the user's interruption return pattern from historical data.

[0039] Feature extraction from environmental factors Feature extraction: Collect environmental context information, such as location, environmental noise level, the status of surrounding devices, etc., to form a feature vector. For example, a short departure at the user's office location is more likely to result in a quick return than a departure in a public place.

[0040] Step 2.2, Calculation of the intention persistence probability; Based on the extracted features, calculate the user's intention persistence probability: The system uses a fusion model to calculate the intention persistence probability : ; Where, represents the user's intention persistence probability, that is, the possibility that the user returns and continues the previous operation after a touch interruption; represents the intention persistence evaluation function, which is used to comprehensively process various features and output the final probability value; represents the feature vector extracted from the touch operation type, including information such as the operation type, duration, and completion degree; represents the feature vector extracted from the application type, including application classification, status information, etc.; represents the feature vector extracted from the time pattern, including time-related features such as the user's usage pattern and interruption frequency; represents the feature vector extracted from environmental factors, including environmental context information such as location and noise level; represents the sigmoid activation function, which maps the output value to the interval (0, 1) so that the result can be interpreted as a probability value; 、 、 、 respectively represent the weight coefficients of the feature vectors extracted from the touch operation type, application type, time pattern, and environmental factors; is the bias term, which is used to adjust the overall prediction baseline of the model. The weight coefficients are optimized by machine learning methods to minimize the intent prediction error.

[0041] The system will divide the value of into multiple intervals, corresponding to different levels of persistent intent intensity: highly likely to return (0.8 - 1.0), moderately likely to return (0.5 - 0.8), lowly likely to return (0.2 - 0.5), and likely not to return (0 - 0.2).

[0042] According to the embodiments of the present application, the intent persistence evaluation model is implemented using a multi-layer perceptron (MLP) structure, including the following layers: Input layer: Receives four types of feature vectors , , and . After each feature vector is normalized, they are concatenated to form a unified input feature vector; Hidden layer: Contains 2 to 3 fully connected layers, each followed by a ReLU activation function and a Dropout regularization layer (dropout rate is 0.3) to improve the generalization ability of the model; Output layer: A single neuron, using the sigmoid activation function, outputs the probability score of the user's return .

[0043] The training of this model adopts a supervised learning method. The training data includes feature vectors and labels (0 or 1) indicating whether the user returns within a predetermined time window. The model uses binary cross-entropy as the loss function and updates the parameters through the Adam optimizer. The model is verified on different user groups and device types to ensure its generalization ability.

[0044] In practical applications, the model can dynamically adjust its complexity according to the device's computing power. For example, a simplified version of the model is used on low-end devices, or a more complex network structure is enabled on high-end devices.

[0045] Step 2.3, Touch state saving policy selection; Based on the calculated intent persistence probability and the touch state complexity , the system selects the optimal state saving policy: The system evaluates the complexity of the current touch state , considering factors such as the number of active touch sessions, the size of touch context data, the number of associated resources, etc.

[0046] Then, the system selects the optimal state saving policy through a decision function : ; Among them, is the utility function, considering the state recovery performance, storage overhead, and energy consumption impact; , , respectively represent the complete state saving, partial critical state saving, and metadata index saving strategies; represents the touch state saving strategy; represents the optimal state saving strategy; represents the user intention persistence probability, reflecting the possibility that the user returns and continues the previous operation; represents the complexity of the current touch state, including factors such as the number of active touch sessions and the size of touch context data; represents selecting the strategy that makes the utility function achieve the maximum value . The purpose of this formula is to select the touch state saving strategy that can maximize the system utility under the constraints of user intention persistence and state complexity.

[0047] For the scenario with a high intention persistence probability ( ), usually select the strategy to save the complete touch state and context; For the scenario with a medium intention persistence probability (0.5 - 0.8), usually select the strategy to save only the critical touch state; For the scenario with a low intention persistence probability ( ), usually select the strategy to save only the metadata index of the state.

[0048] The output result of this step is: The user intention persistence probability ; The selected touch state saving strategy and the corresponding state data.

[0049] These information will be used in the subsequent heterogeneous computing resource scheduling and state recovery processes.

[0050] Step 3: Based on the touch operation entropy value and the predicted path probability, perform cache allocation, allocate touch processing data to heterogeneous memories at different levels, and optimize the cache replacement strategy; Specifically, it includes the following steps: Step 3.1: Heterogeneous memory hierarchy configuration; The system hierarchically configures the heterogeneous memory resources of the embedded tablet device: Level L1: It includes the GPU shared memory and the CPU L1 cache, which has the lowest access latency (<3 ns), but the smallest capacity; Level L2: It includes the CPU L2 cache and part of the GPU local memory, which has a medium access latency (10 to 20 ns) and a moderate capacity; Level L3: It includes the main memory, which has a higher access latency (50 to 100 ns) and the largest capacity.

[0051] The system determines the capacity allocation ratio of each level of memory according to the specific configuration of the device , to meet , and the initial value can be set to , and it can be dynamically adjusted according to the actual usage situation later.

[0052] Step 3.2, Touch operation priority classification; Based on the touch operation entropy value , the system classifies touch operations into three categories: High-priority (high-entropy) operations: Operations with entropy values in the top 20%, indicating important but infrequent operations, such as precise positioning, text selection, etc.; Medium-priority (medium-entropy) operations: Operations with entropy values in the middle 60%, such as regular clicks, short-distance drags, etc.; Low-priority (low-entropy) operations: Operations with entropy values in the bottom 20%, indicating simple and frequent operations, such as page scrolling, etc.

[0053] In addition, the system also considers the operation path probability predicted by the Markov model , and for the operations related to the next state with a high predicted probability, appropriately increase their priority.

[0054] Step 3.3, Hierarchical cache resource allocation; According to the operation priority, the system performs hierarchical allocation of cache resources: For high-priority operations, the system preloads their related data and instructions into the L1-level memory, and the allocation strategy is: ; Among them, represents the set of operations allocated to the L1-level memory related data set; represents a single data item; represents the operation all related data sets; represents the operation entropy value, which is used to measure the complexity and importance of the operation; represents the entropy value threshold of high-priority operations. When the operation entropy value exceeds this threshold, the related data is allocated to the L1-level memory.

[0055] For medium-priority operations, the system allocates their related data and instructions to the L2 memory layer: ; Among them, represents the operations allocated to the L2 memory layer related data set; represents the operation a single data item; represents the operation all related data sets; represents the operation entropy value, used to measure the complexity and importance of the operation; represents the lower limit of the entropy value threshold for low priority; represents the upper limit of the entropy value threshold for high priority. This formula means that when the entropy value of the operation is between the low threshold and the high threshold, its related data will be allocated to the L2 memory layer.

[0056] For low-priority operations, the system mainly retains their related data and instructions in the L3 memory layer: ; Among them, represents the operation related data set, that is, the set of all data items that need to be accessed during the execution of the touch operation ; represents the operation a single data item; represents the operation entropy value, used to quantify the complexity and importance of the touch operation; represents the entropy value threshold for low priority. Operations with an entropy value lower than this threshold are regarded as low-priority operations; represents the operation allocated to the L3 memory layer (main memory) related data set.

[0057] In some embodiments, the system can dynamically adjust the entropy value threshold according to the specific configuration and application scenario of the device. For example, for low-end devices with limited memory resources, the value can be increased to reduce the amount of data allocated to the L1 cache; while for high-end devices with sufficient resources, the threshold can be lowered to allow more data to enter the L1 cache and further improve the response speed.

[0058] The system can also consider the temporal locality and spatial locality of operations. For operations that are temporally or spatially close to the current high-entropy operation, even if their own entropy value is not high, their priorities can be appropriately increased. For example, after an accurate drawing operation (high entropy), a color selection operation (medium entropy) usually follows, and the relevant data can be pre-loaded into a higher-level cache together.

[0059] Step 3.4, cache replacement policy optimization; The system optimizes the traditional LRU or FIFO cache replacement policy and introduces an entropy-weighted cache replacement algorithm: When the cache space is insufficient and data needs to be replaced, the system calculates the retention value of each cache entry : ; Among them, represents the comprehensive retention value of the data item and is used to determine the cache replacement priority; is the retention value of the data item calculated by the traditional cache replacement algorithm, reflecting the temporal locality characteristics of the data; is the operation associated with the data item and represents the entropy value of the operation, indicating the complexity and importance of the operation; represents the touch operation associated with the data item ; is the weight coefficient that balances the two, with a value range of 0 - 1, used to adjust the proportion of the traditional cache policy and the influence of the entropy value, and the larger the value, the more the operation entropy value factor is emphasized.

[0060] The system preferentially replaces the cache entry with the lowest retention value to ensure that data related to high-entropy operations has a higher probability of being retained in the cache.

[0061] In practical applications, the specific implementation of this cache replacement policy can be carried out in the following way: The system maintains a priority queue sorted by the retention value . When a cache entry needs to be replaced, the item with the lowest retention value is directly taken out from the head of the queue for replacement. To improve efficiency, this priority queue is implemented using a binary heap, ensuring that the time complexity of insertion and deletion operations is , where is the number of cache entries.

[0062] Optionally, the system can dynamically adjust the weight coefficient according to different application scenarios. For example, in a game application, The value is more considered with the factor of operation entropy value; while in the document editing application, it can reduce the value to more consider the factor of temporal locality.

[0063] In some embodiments, the system can also consider the comprehensive factors of the access frequency and the recent access time of the cached data to implement an improved entropy-weighted ARC (Adaptive Replacement Cache) policy. This policy maintains two cache lists: one based on the access frequency and the other based on the access time, and dynamically adjusts the sizes of the two lists to adapt to different access patterns.

[0064] The output result of this step is: the data allocation scheme of the cached data in each level of memory , and the optimized cache replacement policy parameters. Among them: : represents the data set allocated to the L1-level memory (the fastest cache level), usually including the data related to high-priority operations; : represents the data set allocated to the L2-level memory (the medium-speed cache level), usually including the data related to medium-priority operations; : represents the data set allocated to the L3-level memory (the slower cache level), usually including the data related to low-priority operations; Cache replacement policy parameters: including the weight coefficient (used to balance the proportion of the traditional cache policy and the influence of the entropy value), the entropy value threshold and (used to distinguish high, medium, and low-priority operations) These configurations will directly affect the response performance of the touch operation and the utilization efficiency of system resources. By reasonably allocating cache resources, it ensures that high-priority operations can respond quickly and at the same time optimizes the overall system performance.

[0065] Step 4, according to the touch operation entropy value, the intention persistence probability, the cache allocation result, and the characteristics of the computing task, intelligently allocate the touch processing task to the heterogeneous computing unit; Specifically, it includes the following steps: Step 4.1, task characteristic analysis and computing load estimation; The system first analyzes the characteristics of different task types in the touch processing process and classifies the tasks into three categories: Computation-intensive tasks: such as gesture recognition, trajectory smoothing, intention prediction, etc., which require high computing power; Memory access-intensive tasks: such as historical data query, status recovery, etc., which have high requirements for memory bandwidth; Hybrid tasks: such as touch event distribution, coordinate transformation, etc., which require both calculation and memory access simultaneously.

[0066] For each task , the system estimates its computational load eigenvector: ; Among them, represents the computational load eigenvector of task ; represents the computational complexity, which is used to quantify the amount of computational resources required by the task and the complexity of computational operations; represents the memory requirement, which is used to indicate the size of the memory space occupied during the execution of the task; represents the bandwidth requirement, which is used to measure the requirements of the task for memory access speed and data transmission ability during execution; represents the sensitivity to execution latency, which is used to characterize the strictness of the task's requirement for response time. The higher the value, the more timely response the task requires.

[0067] Step 4.2, Heterogeneous computing resource characteristic modeling; The system models the performance characteristics of heterogeneous computing resources in the device: CPU core: Usually has high single-thread performance and complex instruction processing ability, suitable for executing tasks with complex control flow; GPU shader core: Has large-scale parallel processing ability, suitable for executing compute-intensive tasks with high data parallelism; DSP processor: Optimized for specific algorithms (such as filtering, signal processing), with high energy efficiency ratio; Dedicated accelerator: Optimized for specific fields (such as deep learning), with obvious performance and energy efficiency advantages.

[0068] The system represents the performance characteristics of each computing resource as: ; Among them, represents the performance characteristic vector of computing resource ; represents the computing power of computing resource , such as the number of floating-point operations or instructions that can be executed per second; represents the memory capacity of computing resource , that is, the available storage space size; represents the memory bandwidth of computing resource , that is, the amount of data that can be transmitted per unit time; represents the energy efficiency ratio of computing resource , that is, the amount of computation that can be completed per unit energy consumption.

[0069] Step 4.3, Predictive Task Allocation and Scheduling; Next operation probability distribution predicted based on the Markov model , the system performs predictive task allocation and scheduling: For the next operation whose prediction probability exceeds the threshold , the system prepares the required computing resources and data in advance; The system calculates the fitness score of the task on different computing resources : ; where represents the fitness score of task on computing resource ; represents the summation over all considered factors ; is the weight coefficient of each evaluation factor, used to adjust the importance of different factors; is a function for evaluating the matching degree of task load and resource characteristics, used to calculate the matching degree of task characteristics and resource characteristics in the th dimension; represents the computational load feature vector of task , including computational complexity , memory requirement , bandwidth requirement and latency sensitivity ; represents the performance characteristic vector of computing resource , including computing power , memory capacity , memory bandwidth and energy efficiency ratio .

[0070] The system considers the operation entropy value and the intention persistence probability , and constructs a task priority metric: ; where represents the priority metric of task ; represents the entropy value of operation , reflecting the complexity and information content of the operation; represents the user intention persistence probability, characterizing the possibility that the user's current intention remains unchanged; represents the sensitivity of task to execution latency, and a higher value indicates that the task is more sensitive to latency; , , represent the weight coefficients of the operation entropy value, the user intention persistence probability, and the task delay sensitivity respectively, satisfying to ensure weight normalization.

[0071] Based on the fitness score and task priority, the system uses an improved heterogeneous scheduling algorithm for task allocation: ; where represents the computing resources to which task is allocated; represents the resource index that selects the maximum value of the objective function ; represents the fitness score between task and computing resources ; represents the priority index of task ; is the priority influence factor, and the larger the value, the greater the influence of the priority on the allocation decision. This formula combines the fitness score with the priority factor to determine the computing resources most suitable for executing task .

[0072] In a specific application scenario, such as a multi-finger gesture recognition operation, the workflow of this predictive task allocation and scheduling algorithm is as follows: When the user starts a multi-finger touch operation, the system detects the initial contact points, triggers the Markov model prediction. According to the current application context and the characteristics of the initial contact points, the system predicts that the most likely next operation is "two-finger zoom", and the prediction probability is 0.85 (exceeding the preset threshold of 0.7); The system immediately analyzes the processing tasks required for the "two-finger zoom" operation, including sub-tasks such as contact point trajectory smoothing, relative distance calculation, and real-time update of the zoom ratio, and estimates the computational load feature vectors of each sub-task; For the trajectory smoothing task, the system calculates the fitness scores of this task on different computing resources. Since the trajectory smoothing task has the characteristics of data parallelism and is computationally intensive, the fitness evaluation shows that this task has the highest execution efficiency on the GPU; At the same time, the system considers the priority factor of this task. Since "two-finger zoom" is a high-entropy operation (an important operation that the user is concerned about), and the current user intention persistence probability is high, the system assigns a high priority to this task; Finally, the system decides to allocate the trajectory smoothing task to the GPU for execution, preloads the relevant image transformation parameters into the GPU shared memory, and at the same time allocates the zoom ratio update task to the CPU for execution, so as to achieve parallel processing and reduce the response delay.

[0073] In the case of resource competition, the system can dynamically adjust the priority influence factor value. For example, when the device is in a high-load state, the value can be increased to enhance the influence of the priority on the allocation decision and ensure that critical operations obtain sufficient resources; when the device load is low, the value can be decreased to give more consideration to the inherent matching degree between tasks and resources.

[0074] In some other embodiments, the system can also implement a context-aware scheduling strategy to dynamically adjust the weight coefficients , and according to the user's current operation context. For example, in scenarios where the user is performing fine drawing and other operations that require high-precision response, the value is increased to pay more attention to the operation entropy value; in the usage pattern where the user frequently leaves and returns, the value is increased to pay more attention to the factor of intention persistence.

[0075] Step 4.4, dynamic resource allocation and load balancing; The system monitors the load conditions of each computing unit in real time. When load imbalance is detected, dynamic adjustment is performed: Calculate the load level of each computing unit: ; Among them, represents the load level of computing unit , represents the total computing requirements of all tasks assigned to computing unit , represents the task set currently assigned to computing unit , represents the computing requirement of task , represents the upper limit of the computing capacity of computing unit . The formula calculates the ratio of the current load of the computing unit to its maximum computing capacity, which is used to evaluate the usage of computing resources.

[0076] When the load of a certain unit exceeds the threshold or is lower than the threshold , the system re-evaluates the task allocation: ; Among them, represents the minimum migration loss value from computing unit to computing unit ; represents the task to be migrated; Represents the set of tasks currently assigned to the computing unit ; Represents finding the task with the minimum migration loss among all the assigned tasks in ; Represents migrating the task from the resource to . This loss may include factors such as migration overhead and changes in execution efficiency. This formula is used to find the most suitable task for migration during the load balancing process to minimize the performance impact of migration.

[0077] When the task with the minimum migration loss is found and the loss is less than the threshold, the system performs task migration to achieve load balancing.

[0078] In some embodiments, the system can adopt a predictive load balancing strategy. Based on historical load patterns and current touch operation trends, it adjusts task allocation in advance to avoid load imbalance. For example, when the system predicts that it is about to enter an image processing intensive stage, it can transfer some non-critical tasks from the GPU to the CPU in advance to reserve resources for the upcoming computationally intensive tasks.

[0079] The system can also dynamically adjust the load balancing strategy according to the temperature and battery status of the device. For example, when the device temperature is too high, the system can appropriately lower the threshold and more actively disperse the tasks of high-load units; when the battery power is low, the system can preferentially allocate tasks to the computing units with high energy efficiency ratios to extend the device's battery life.

[0080] The output result of this step is: The task allocation scheme for current and predicted touch operations ; The load distribution of computing resources ; The set of data that needs to be pre-loaded into the cache. These configurations will directly affect the system's response speed to touch operations and resource utilization efficiency.

[0081] Step 5: Based on the continuous probability of user intent, the computing task allocation status, and the environmental perception data, implement a differentiated sleep strategy and a fast wake-up mechanism for heterogeneous computing units; Specifically, it includes the following steps: Step 5.1, the hierarchical sleep strategy for computing units; Based on the continuous probability of user intent , the system formulates a hierarchical sleep strategy for heterogeneous computing units: For different computing units , the system determines an appropriate sleep level according to their importance and energy consumption characteristics in touch processing : Shallow Sleep (S1 level): Keep the core data structure and status in memory, reduce the clock frequency, turn off unnecessary functional units, and the wake-up latency < 10 ms; Medium Sleep (S2 level): Write the core status to fast storage, turn off most functional units, only retain the monitoring function, and the wake-up latency is 20 - 50 ms; Deep Sleep (S3 level): Write the complete status to non-volatile storage, turn off all units except the lowest power consumption monitoring module, and the wake-up latency > 100 ms.

[0082] The system allocates the sleep level according to the continuous probability of the intention : ; Among them, represents the sleep level of the computing unit , represents the continuous probability of the user intention, represents the processing priority of the computing unit , , respectively represent the high-priority threshold and the low-priority threshold. S1, S2, and S3 respectively represent three different sleep levels: shallow sleep, medium sleep, and deep sleep. This formula allocates appropriate sleep levels for different computing units according to the combined conditions of the continuous probability of the user intention and the priority of the computing unit to balance energy consumption savings and wake-up response speed.

[0083] Step 5.2, construction of the low-power environment perception network; The system constructs a lightweight low-power environment perception network to continuously monitor the signs of the user's return: Sensor layer: Use multi-modal sensors (such as proximity sensors, ambient light sensors, acceleration sensors, etc.) to construct a perception network, and each sensor operates in a low-frequency sampling working mode; Feature extraction layer: Perform lightweight feature extraction on the sensor data: ; Among them, represents the feature vector extracted from the sensor , represents the feature extraction function applied to the sensor , represents the sampling value of the sensor at the current moment , represents the sampling value of the sensor at the previous moment , Represents a sensor Before Sampling value at a time unit, Represents the size of the historical data window considered during feature extraction. This formula describes how to extract meaningful features from the time series data of the sensor for subsequent user behavior analysis and prediction.

[0084] Fusion decision layer: Fuses the features of multiple sensors and calculates the user return possibility score: ; Wherein, Represents the feature fusion function, which can be a weighted sum, decision tree or lightweight neural network; , , Respectively represent the features extracted from the , , th sensor; Represents the number of sensors; Represents the user return possibility score calculated by the system. The higher this score, the greater the possibility that the user will return to the device. Based on this, the system decides whether to trigger the wake-up operation of the computing unit.

[0085] According to the embodiments of the present application, the low-power environment perception network has the following specific implementation details: Perceptor configuration and sampling strategy: This network adopts a multi-layer perceptron structure and integrates multiple sensors, specifically including: Proximity sensor: Detects the change in the distance between the user and the device, adopts a dynamic threshold trigger mechanism, reduces the sampling frequency in the user's leaving state, and increases the sampling frequency when a proximity event is detected; Ambient light sensor: Monitors the change in ambient light, adopts a change amplitude trigger sampling mechanism, and only records data when the light intensity change exceeds a preset threshold to reduce power consumption; Microphone array: Listens to the ambient sound characteristics in an extremely low-power mode, mainly identifies the sound characteristics that the user may return, such as footsteps, door opening and closing sounds, etc., and adopts an intermittent sampling method triggered by a sound energy threshold; Acceleration sensor: Detects the movement and vibration of the device, adopts an event-triggered sparse sampling strategy, and only activates more accurate attitude analysis when significant movement is detected; Low-power Bluetooth (Bluetooth Low Energy, BLE) signal: Detects the change in the signal strength of the user's wearable device to evaluate the relative position relationship between the user and the device.

[0086] Feature extraction method: For different sensor data streams, the system adopts a feature extraction algorithm with low computational complexity but significant effects: Time-domain features: Calculate statistical features (such as mean, variance, peak value, etc.) within a sliding window to capture the short-term change trend of sensor data; Frequency-domain features: For periodic sensor data (such as sound, vibration), use the Simplified Fast Fourier Transform (Fast Fourier Transform, FFT) to extract spectral features and identify specific behavior patterns; Rate-of-change features: Calculate the numerical change rate between adjacent sampling points to detect sudden change events, such as user behavior indicators like sudden light changes and sudden motion changes; Duration features: Analyze the duration of signal changes to distinguish between short-term interferences and actual user behaviors.

[0087] Fusion decision-making model: The system adopts a hierarchical lightweight decision-making model for multi-sensor information fusion: First layer: Each sensor's independent evaluation model converts the original data into a behavior possibility score, using a simplified logistic regression or threshold judgment model; Second layer: Feature-level fusion combines the feature vectors of each sensor for comprehensive evaluation, using a Support Vector Machine (Support Vector Machine, SVM) or a lightweight random forest algorithm; Third layer: Decision-level fusion combines the output results and confidence levels of each sub-model and obtains the final judgment through a weighted voting mechanism.

[0088] This multi-level fusion method can provide reliable user behavior prediction while maintaining low power consumption.

[0089] Step 5.3, Conditional-triggered hierarchical wake-up mechanism; The system implements a conditional-triggered hierarchical wake-up mechanism based on the output of the environmental perception network: Construct a wake-up condition judgment function: ; Among them, represents the wake-up decision function, with an output of 0 or 1, determining whether the system needs to wake up; represents the data set from each sensor; represents the fusion decision-making model, which is used to process sensor data and output the user return possibility score; is an indicator function, taking a value of 1 when the condition within the square brackets is satisfied, otherwise 0; is the wake-up threshold, dynamically adjusted according to the current state of the device, and used to control the sensitivity of system wake-up.

[0090] When the system performs wake-up operations in the following order: First, wake up the high-priority computing units in the light sleep state (S1 level) and restore their working states; Then, based on the touch state data saved in Step 2 , selectively restore the cache states of the critical touch processing paths; According to the possible operation sequences predicted by the Markov model , preload relevant resources; among them, , , respectively represent the possible operations of the sensor at , time, represents the size of the historical data window considered during feature extraction. This sequence contains a series of consecutive operations that the system predicts the user is most likely to perform in the short term. By preloading the resources required for these operations, the loading delay during actual operations can be reduced; According to actual needs, wake up other computing units in the medium (S2 level) and deep sleep (S3 level) states in order of priority.

[0091] Step 5.4, Intelligent Wake-up Pipeline Optimization; To further reduce the wake-up delay, the system implements intelligent wake-up pipeline optimization: Analyze historical wake-up data to identify the critical paths and bottlenecks in the wake-up process: ; Among them, represents the total time consumption of the entire wake-up process, , , respectively represent the time consumption of the , , th steps in the wake-up process, represents the total number of steps in the wake-up process. This formula shows that the time consumption of the entire wake-up process depends on the step with the longest time consumption among all steps, reflecting the critical path principle in the system wake-up process.

[0092] For different continuous probabilities of intentions and touch state complexities , the system constructs different wake-up pipeline templates .

[0093] For frequently occurring wake-up scenarios, the system pre-compiles optimized wake-up instruction sequences to reduce the decision-making overhead during runtime.

[0094] Implement a parallel wake-up strategy to allow multiple non-dependent wake-up steps to be executed simultaneously, further reducing the wake-up delay.

[0095] The output result of this step is: Sleep level allocation for each heterogeneous computing unit ; Configuration parameters of the environment perception network; Wake-up condition judgment function and hierarchical wake-up strategy; Optimized wake-up pipeline template set. These mechanisms together ensure the optimal balance between low energy consumption and high responsiveness of the system.

[0096] Application example of this embodiment: This embodiment can be applied to a variety of embedded tablet devices. Taking a digital drawing tablet as an example below, the implementation process and effects of the touch optimization method based on heterogeneous computing in practical applications are introduced.

[0097] The professional digital drawing tablet application scenario has obvious heterogeneous requirements: precise drawing operations require low-latency and high-precision touch responses; brush style switching and color mixing operations need to frequently access the resource library; the user's working state is frequently interrupted (such as referring to materials, taking breaks, etc.) and then quickly returns to continue working. Traditional drawing tablets face problems such as inconsistent responses, low resource utilization efficiency, and slow user state recovery in these scenarios.

[0098] Implementation process example: Touch operation entropy calculation: The system collected the usage data of 28 professional illustrators on the digital drawing tablet and analyzed the entropy values of various touch operations. In the implementation, the system divided 20 common touch operations into three categories: High-entropy operations (entropy value > 2.5): Precise operations such as fine line drawing, precise area selection, and detail eraser; Medium-entropy operations (1.0 ≤ entropy value ≤ 2.5): Tool operations such as color selection, brush replacement, and layer operations; Low-entropy operations (entropy value < 1.0): Common operations such as canvas panning, zooming, and rough coloring.

[0099] Example of touch operation entropy value calculation results: Fine line drawing (3.42), area selection (2.87), eraser (2.63), layer switching (1.85), color selection (1.53), canvas panning (0.82), zooming (0.76).

[0100] Construction of touch path Markov model: The system built a specific Markov model for the drawing application to capture the user's drawing behavior pattern. Taking the operation sequence after creating a new layer as an example, the system learned the following typical transition probabilities: New layer creation → Brush selection (0.72) → Color mixing (0.65) → Fine drawing (0.81).

[0101] Hierarchical Cache Resource Allocation: Based on the operation entropy value, the system allocates data to three levels of caches: In the GPU shared memory and the CPU L1 cache (with a total capacity of 32KB), the system preferentially allocates data related to high-entropy operations such as fine drawing and selection operations, such as the current brush parameters, selection coordinates, pressure sensitivity mapping table, etc.; In the CPU L2 cache (512KB), the system allocates data for medium-entropy operations such as color adjustment and common tools; In the main memory (8GB), all resources are saved and mainly accessed for low-entropy operations.

[0102] Predictive Task Allocation: According to the prediction results of the Markov model, the system pre-allocates computing resources. For example, when it is recognized that the user has completed a line drawing and it is predicted that the next step is likely to be a coloring operation, the system pre-loads the color mixing related algorithms into the GPU and prepares the layer composition task on the CPU at the same time.

[0103] Intention Persistence Evaluation: The system analyzes the user interruption behavior pattern and finds that when professional users refer to external materials, they will return to continue working within 3 minutes in 87% of the cases. Based on this feature, the system keeps the core drawing state in the fast cache when the user is away and makes the drawing processing unit stay in a shallow sleep state.

[0104] Verification of Technical Effects: Tables 1 and 2 show the effect comparison of this method on a digital drawing tablet: Table 1: Data on the improvement of drawing response performance;

[0105] Table 2: Data on the improvement of resource utilization efficiency;

[0106] Through the above actual application tests, it can be seen that this embodiment significantly improves the response performance of key touch operations in the digital drawing tablet scenario. In particular, the response time of fine drawing operations is shortened by nearly half, and at the same time, the fluctuation of the response time is greatly reduced, providing a more stable drawing experience. The system resource utilization efficiency is also significantly improved, the GPU and CPU resources are utilized more evenly and efficiently, and the memory cache hit rate is increased by 44%. These improvements together support a smoother drawing experience, enabling the drawing tablet to better meet the high-precision and low-latency requirements of professional users.

[0107] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An embedded tablet touch optimization method based on heterogeneous computing, characterized in that It includes the following steps: Calculate the information entropy values of different touch operations, establish a Markov model for the touch path, and identify the importance of different touch operations and the touch state transition probabilities; Based on the touch operation entropy values and the state transition probabilities of the Markov model, combined with the operation type, application type, time pattern, and environmental factors before the user's touch interruption, calculate the user intention persistence probability, and accordingly select the corresponding touch state saving strategy; Perform cache allocation based on the touch operation entropy values and the predicted path probabilities, allocate the touch processing data to heterogeneous memories at different levels, and optimize the cache replacement strategy; Intelligently allocate the touch processing tasks to heterogeneous computing units according to the touch operation entropy values, intention persistence probability, cache allocation results, and computing task characteristics; Based on the user intention persistence probability, computing task allocation status, and environmental perception data, implement a differential sleep strategy and a fast wake-up mechanism for heterogeneous computing units.

2. The optimized method for embedded tablet touch control based on heterogeneous computing according to claim 1, wherein The steps of calculating the information entropy values of different touch operations and establishing a Markov model for the touch path include: Obtain the original touch data stream from the touch sensor module; For each touch operation type, collect and analyze its historical state distribution, and calculate its information entropy value; Construct a Markov model for the touch path for different application types, and calculate the conditional transition probabilities; Adopt an incremental learning method to regularly update the entropy values and Markov model parameters.

3. An embedded tablet touch optimization method based on heterogeneous computing according to claim 1, characterized in that, The steps of calculating the user intention persistence probability include: Extract features related to user intention persistence from multiple sources to form a feature vector; Use a fusion model to calculate the intention persistence probability; Divide the intention persistence probability into multiple intervals corresponding to different persistent intention intensity levels; Select the optimal state saving strategy according to the intention persistence probability and the touch state complexity.

4. An embedded tablet touch optimization method based on heterogeneous computing according to claim 1, characterized in that, The heterogeneous memory includes: L1 layer: including GPU shared memory and CPU L1 cache; L2 layer: including CPU L2 cache and part of the GPU local memory; L3 layer: including the main memory.

5. A method for optimizing embedded tablet touch control based on heterogeneous computing according to claim 1, characterized in that, The cache allocation based on the touch operation entropy values and the predicted path probabilities includes: Allocate the data related to high entropy value operations to the L1 layer for storage; Allocate the data related to medium entropy value operations to the L2 layer for storage; Allocate the data related to low entropy value operations to the L3 layer for storage; When the cache space is insufficient and data needs to be replaced, replace the cache item with the lowest remaining value according to the remaining value priority.

6. A method for optimizing embedded tablet touch control based on heterogeneous computing according to claim 1, characterized in that The steps of intelligently allocating the touch processing tasks to heterogeneous computing units include: Analyze the characteristics of the touch processing tasks and calculate the load feature vector; Model the performance characteristics of the heterogeneous computing resources in the device; Based on the probability distribution of the next operation predicted by the Markov model, perform predictive task allocation and scheduling; Monitor the load conditions of each computing unit in real time, and perform dynamic adjustment when detecting load imbalance.

7. An embedded tablet touch optimization method based on heterogeneous computing according to claim 1, characterized in that The steps of implementing a differential sleep strategy and a fast wake-up mechanism for heterogeneous computing units include: Based on the user intention persistence probability, formulate a hierarchical sleep strategy for heterogeneous computing units; Construct a lightweight low-power environmental perception network to continuously monitor the signs of the user's return; Based on the output of the environmental perception network, implement a condition-triggered hierarchical wake-up mechanism; Analyze the historical wake-up data to optimize the intelligent wake-up pipeline.

8. An embedded tablet touch optimization method based on heterogeneous computing according to claim 7, characterized in that, The hierarchical sleep strategy includes: Shallow sleep: Keep the core data structures and states in memory and reduce the clock frequency; Medium sleep: Write the core states to fast storage, turn off most functional units, and only retain the monitoring function; Deep sleep: Write the complete states to non-volatile storage and turn off all units except the lowest-power monitoring module.

9. An embedded tablet touch optimization method based on heterogeneous computing according to claim 1, characterized in that, The touch operations in the touch operation entropy value, intention persistence probability, cache allocation result, and computing task characteristics include: fine line drawing, area selection, eraser operation, layer switching, color selection, canvas panning and zooming.

10. An embedded tablet touch optimization system based on heterogeneous computing, for performing a method for optimizing embedded tablet touch based on heterogeneous computing according to any one of claims 1-9, characterized in that, It includes: A touch operation entropy calculation module, which is used to calculate the information entropy values of different touch operations, establish a Markov model of the touch path, and identify the importance of different touch operations and the touch state transition probability; An intention persistence evaluation module, which is used to calculate the user intention persistence probability based on the touch operation entropy value and the state transition probability of the Markov model, combined with the operation type, application type, time pattern, and environmental factors before the user touch interruption, and select the corresponding touch state saving strategy accordingly; A heterogeneous memory allocation module, which is used to allocate touch processing data to different levels of heterogeneous memory based on the touch operation entropy value, state transition probability, and user intention persistence probability, and optimize the cache replacement strategy; A task scheduling module, which is used to intelligently allocate touch processing tasks to heterogeneous computing units according to the touch operation entropy value, intention persistence probability, cache allocation result, and computing task characteristics; A sleep wake-up control module, which is used to implement a differentiated sleep strategy and a fast wake-up mechanism for heterogeneous computing units based on the user intention persistence probability, computing task allocation status, and environmental perception data.

Images