Flexible display real-time picture quality rendering control system based on edge computing

Abstract

The application relates to the field of edge computing, and particularly discloses a flexible display real-time picture quality rendering control system based on edge computing. An edge computing node collects curvature sensing data of each area of a flexible screen in real time; data mining software performs association rule mining on a historical curvature change sequence and a local pixel rendering calculation amount, and constructs a curvature-computing power demand prediction model; the edge computing node predicts the rendering computing power demand of each display sub-area of a next frame according to the model, dynamically reconstructs an access bandwidth allocation queue and a rendering task scheduling sequence of frame buffer data, preferentially routes rendering data flow of a high-computing power demand sub-area to an idle computing power cluster on the edge side, and writes rendering data flow of a low-computing power demand sub-area into the frame buffer according to a normal time sequence. The application overcomes picture tearing and delay caused by local computing power supply-demand mismatch during deformation of the flexible screen, reduces resource consumption of the edge node, improves the timeliness and continuity of rendering data writing into the frame buffer, and guarantees stable picture quality under complex deformation of the flexible screen.

Description

Technical Field

[0001] This invention relates to the field of edge computing, and specifically discloses a real-time image quality rendering control system for flexible displays based on edge computing. Background Technology

[0002] Flexible display technology achieves diverse display forms through bendable and foldable screen substrates. In the image rendering control process of flexible displays, a fixed rendering power allocation strategy and a linear frame buffer data writing mechanism are typically employed. After receiving the image data stream to be displayed, a conventional rendering system divides the image data evenly into multiple data blocks according to the screen's physical resolution and preset rendering priority. These blocks are then allocated to the various computing cores of the central processing unit (CPU) or graphics processing unit (GPU) for geometric transformation and color calculation. The calculated pixel data is sequentially written to the corresponding physical address space of the frame buffer according to a fixed linear address mapping relationship. The display controller then reads the data from the frame buffer according to a fixed refresh sequence and outputs it to the screen pixel drive circuit. This rendering data flow control method, based on a static topology and fixed computing power scheduling, can maintain normal image output when the screen is flat, ensuring synchronized data writing and reading sequences and even distribution of the rendering load across the screen. The above process constitutes the basic framework for rendering data processing in conventional states for existing flexible displays.

[0003] When a flexible display undergoes physical deformation, local areas of the screen experience curvature changes, leading to non-linear alterations in the pixel spatial distribution density. This results in a sharp increase in the complexity of texture mapping and lighting calculations, causing a sudden shift in local rendering computational demands. Current technologies still rely on fixed computational allocation and data writing strategies to address such deformation scenarios. Lacking a mechanism to perceive the correlation between the physical deformation state of different screen regions and the amount of rendering computation, the edge side cannot anticipate the impending sudden shift in computational demands in local areas. When the rendering computational demand of a folded or curved area exceeds its fixed allocated computational capacity, the rendering data stream processing in that area experiences delays, and the frame buffer data writing timing lags. Simultaneously, the rendering data streams from undeformed areas continue to be written to the frame buffer and output for display according to their original fixed timing. The time difference between the arrival of rendering data streams from different regions at the frame buffer disrupts the synchronicity of frame data writing. The display controller reads composite frame data spanning different time points, visually manifesting as screen tearing.

[0004] Based on the specific implementation methods of the existing technologies, the core technical problem of the existing real-time image quality rendering control system for flexible displays lies in the fact that when the flexible screen deforms, causing a sudden change in local rendering computing power demand, the lack of a prediction and dynamic scheduling mechanism based on the correlation between physical deformation and computing power demand leads to a mismatch between local computing power supply and demand due to fixed computing power allocation and linear data writing strategies. Rendering data stream writing in high-computing-power-demand areas lags behind low-computing-power-demand areas, and frame buffer data writing timing becomes out of sync, resulting in screen tearing and rendering delays. The existing rendering control system only treats the flexible screen as a static display carrier for data processing, failing to use physical deformation characteristics as the basis for rendering data stream scheduling and frame buffer writing control. The system cannot adaptively reconstruct the rendering task scheduling sequence and the frame buffer data access bandwidth allocation queue before or during deformation, leading to congestion in local rendering data processing and disordered global screen output timing. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time image quality rendering control system for flexible displays based on edge computing, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The edge computing-based real-time image quality rendering control system for flexible displays includes:

[0008] Edge computing nodes collect real-time curvature sensing data from various areas of the flexible screen;

[0009] Data mining software, deployed on the edge computing node, performs association rule mining on the historical curvature change sequence and local pixel rendering computation volume to construct a curvature-computing power demand prediction model.

[0010] The edge computing node processes the curvature sensing data according to the curvature-computing power demand prediction model to predict the rendering computing power demand of each display sub-region in the next frame.

[0011] The edge computing node dynamically reconstructs the frame buffer data access bandwidth allocation queue and the rendering task scheduling sequence based on the rendering computing power requirements;

[0012] The edge computing node prioritizes routing the rendering data stream of high-computing-power-demand sub-regions to the idle computing power cluster on the edge side, and writes the rendering data stream of low-computing-power-demand sub-regions into the frame buffer according to the normal timing.

[0013] Preferably, the process of the edge computing node acquiring curvature sensing data of each region of the flexible screen in real time includes extracting the spatial gradient features and time series change rate of the curvature sensing data, and fusing the spatial gradient features and the time series change rate into a multidimensional deformation feature vector.

[0014] The data mining software performs deep association rule mining based on the multidimensional deformation feature vector and the historical local pixel rendering computation, and generates the curvature-computing power demand prediction model in the multidimensional feature space. The edge computing node inputs the multidimensional deformation feature vector obtained in real time into the curvature-computing power demand prediction model and outputs the predicted computing power demand value of each display sub-region in the next frame.

[0015] Preferably, the process of the data mining software performing association rule mining on the historical curvature change sequence and the local pixel rendering computation includes calculating the time offset and amplitude ratio coefficient between the extreme points of curvature change rate and the peak points of rendering computation in the historical curvature change sequence, and storing the time offset and the amplitude ratio coefficient as an association weight coefficient matrix in the curvature-computing power demand prediction model.

[0016] The edge computing node, based on the extreme point of the rate of curvature change before the current moment and combined with the associated weight coefficient matrix, calculates the position and magnitude of the peak point of the rendering computation amount in a specific frame interval after the current moment, and forms the rendering computational power requirement of each display sub-region in the next frame.

[0017] Preferably, the process of the edge computing node dynamically reconstructing the access bandwidth allocation queue of frame buffer data according to the rendering computing power requirements includes sorting the rendering data stream writing priority in descending order according to the rendering computing power requirements prediction value of each display sub-region, and generating a dynamic priority list.

[0018] The display sub-regions ranked first in the dynamic priority list are allocated continuous frame buffer write time slots and high-bandwidth transmission channels, while the display sub-regions ranked later are allocated discrete frame buffer write time slots and low-bandwidth transmission channels. The duty cycle of the continuous frame buffer write time slots and the discrete frame buffer write time slots is adjusted according to the real-time update of the dynamic priority list.

[0019] Preferably, the process of the edge computing node dynamically reconstructing the rendering task scheduling sequence according to the rendering computing power requirement includes extracting the image data quantity corresponding to each display sub-region in the rendering data stream, and calculating the rendering load density of each display sub-region in combination with the rendering computing power requirement prediction value.

[0020] The rendering tasks corresponding to the display sub-regions whose rendering load density exceeds a preset density threshold are migrated from the regular execution queue of the rendering task scheduling sequence to the accelerated execution queue. The edge computing node prioritizes routing the data streams corresponding to the rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side.

[0021] Preferably, when the time offset is less than a preset time threshold, it is determined that the flexible screen has undergone abrupt deformation;

[0022] The edge computing node triggers the interruption and reconstruction mechanism of the rendering pipeline, freezes the rendering tasks that have not been executed in the rendering task scheduling sequence at the current moment, and truncates the rendering data stream that is being written to the frame buffer.

[0023] The edge computing node re-collects curvature sensing data after the sharp deformation, recalculates the rendering computing power requirement prediction value for the next frame based on the curvature-computing power requirement prediction model, overwrites the current frame buffer data access bandwidth allocation queue and rendering task scheduling sequence with the recalculated rendering computing power requirement prediction value, and restores the writing of the truncated rendering data stream.

[0024] Preferably, the process by which the edge computing node prioritizes routing the data stream corresponding to the rendering task in the accelerated execution queue to the idle computing power cluster on the edge side includes monitoring the real-time load rate and cache occupancy rate of each computing core in the idle computing power cluster on the edge side, and selecting computing cores whose real-time load rate and cache occupancy rate are both lower than the corresponding preset threshold to form an available computing power pool.

[0025] Based on the predicted computing power requirements of each rendering task in the accelerated execution queue, the rendering tasks in the accelerated execution queue are split into multiple rendering subtasks, the multiple rendering subtasks are distributed to the computing cores in the available computing power pool for parallel execution, and the rendering result data streams output by each computing core are aggregated and written into the frame buffer.

[0026] Preferably, the process of the edge computing node transmitting rendering data streams through the high-bandwidth transmission channel and the low-bandwidth transmission channel includes performing lossless compression encoding on the rendering data streams entering the high-bandwidth transmission channel and inserting a high-priority transmission identifier and a first target physical address segment into the header of the compressed data packet;

[0027] Lossy compression encoding is performed on the rendering data stream entering the low-bandwidth transmission channel, discarding detail components in the rendering data stream that are below the visual resolution threshold, and inserting a low-priority transmission identifier and a second target physical address segment into the header of the compressed data packet.

[0028] The frame buffer controller arbitrates the execution order of writing corresponding compressed data packets into the frame buffer based on the high-priority transmission identifier and the low-priority transmission identifier.

[0029] Preferably, the cloud server is deployed with cloud-integrated application runtime support platform software;

[0030] The edge computing node uploads the mapping relationship log between the curvature sensing data and the rendering computing power requirement to the cloud-based integrated application runtime support platform software;

[0031] The cloud-integrated application operation support platform software aggregates mapping relationship logs uploaded by multiple edge computing nodes, performs periodic iterative training on the global curvature-computing power demand baseline model, and distributes the updated global curvature-computing power demand baseline model to the edge computing nodes.

[0032] The edge computing node replaces the excess weighting parameters in the curvature-computing demand prediction model with the updated global curvature-computing demand baseline model.

[0033] Preferably, the cloud server is also deployed with cloud computing software. When the flexible screen is in a static state and undergoes slow deformation, the edge computing node only extracts the contour coordinates of the region with a rate of change lower than the slow deformation threshold from the curvature sensing data, and uploads the contour coordinates of the region to the cloud computing software.

[0034] The cloud computing software calculates the pixel offset caused by slow deformation based on the region contour coordinates, generates a local coordinate remapping table, and sends it to the edge computing node.

[0035] The edge computing node reads the pixel data of the original still image from the frame buffer according to the local coordinate remapping table, performs a coordinate translation operation on the pixel data, and then writes it back to the corresponding physical address of the frame buffer.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. This invention constructs a curvature-computing power demand prediction model by deploying data mining software on edge computing nodes to mine association rules between historical curvature change sequences and local pixel rendering computation. Based on this model, it predicts the rendering computation power demand of each display sub-region in the next frame, dynamically reconstructs the bandwidth allocation queue for frame buffer data and the rendering task scheduling sequence, and prioritizes routing the rendering data stream of high-computing-power-demand sub-regions to idle computing power clusters on the edge side, while writing the rendering data stream of low-computing-power-demand sub-regions into the frame buffer according to the normal time sequence. This overcomes the screen tearing and delay caused by the mismatch between local rendering computation power supply and demand when flexible screens deform. The spatial gradient features and time series change rate of the curvature sensing data are extracted and fused into a multi-dimensional deformation feature vector. The time offset and amplitude ratio coefficient between the extreme points of curvature change rate and the peak points of rendering computation in the historical curvature change sequence are calculated as the correlation weight coefficient matrix, making the computation power demand prediction have both temporal and amplitude predictability and determinism. A dynamic priority list is generated based on the predicted computing power demand in descending order. Continuous frame cache write slots and high-bandwidth transmission channels are allocated to the top-ranked display sub-regions. Combined with rendering load density, rendering tasks exceeding a threshold are migrated to an accelerated execution queue. A rendering pipeline interruption and reconstruction mechanism is triggered when the time offset is less than a preset time threshold, indicating a drastic change. Unexecuted tasks are frozen, and the currently written data stream is truncated for recalculation, ensuring the timeliness of data writing in high-load areas. Monitoring the real-time load rate and cache occupancy of the computing cluster filters available computing power pools. Tasks in the accelerated execution queue are split and executed in parallel, reducing the ineffective consumption of overall computing resources at edge nodes and improving the continuity and synchronization of real-time rendering data stream writing to the frame cache.

[0038] 2. This invention performs lossless compression encoding on rendering data streams entering high-bandwidth transmission channels and inserts high-priority transmission identifiers and a first target physical address segment. For rendering data streams entering low-bandwidth transmission channels, it performs lossy compression encoding, discarding detail components below the visual resolution threshold and inserting low-priority transmission identifiers and a second target physical address segment. The frame buffer controller arbitrates the writing order based on the priority identifiers, ensuring priority writing of data in high-computing-power-demand areas and streamlined transmission of data in low-computing-power-demand areas under bandwidth-constrained conditions, reducing data redundancy in the transmission link. Edge computing nodes upload the mapping relationship log between curvature sensing data and rendering computing power requirements to a cloud server deploying the cloud-integrated application runtime support platform software. The cloud aggregates logs from multiple edge nodes to periodically iterate and train the global curvature-computing-power-demand baseline model and distributes updates, enabling the edge-side prediction model to continuously adapt to the deformation and aging characteristics of different flexible screen hardware and complex scene changes, maintaining long-term stability of prediction accuracy. When the flexible screen is in a static state and undergoes slow deformation, the edge computing node only extracts the contour coordinates of the area with a change rate lower than the slow deformation threshold and uploads them to the cloud server where cloud computing software is deployed. The cloud computing generates a local coordinate remapping table based on the pixel offset caused by the slow deformation and sends it to the edge side. The edge side reads the original pixel data from the frame buffer according to the table, performs coordinate translation operation, and then writes it back to the corresponding physical address. This avoids the full re-rendering of the static image under slow deformation and reduces the overall power consumption and data processing load of the system in low dynamic scenarios. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall operation of the edge computing-based flexible display screen image rendering control system of the present invention.

[0040] Figure 2 This is a flowchart of the construction of multidimensional deformation feature vectors and model input processing for the flexible screen of the present invention;

[0041] Figure 3 This is a flowchart illustrating the historical extreme value correlation weight construction and rendering peak computing power estimation process of this invention;

[0042] Figure 4 This is a flowchart illustrating the dynamic allocation and reconstruction process of frame buffer access bandwidth and time slot resources according to the present invention.

[0043] Figure 5 This is a flowchart of the emergency reconstruction process for the flexible screen rapid deformation determination and rendering pipeline of the present invention;

[0044] Figure 6 This is a flowchart of the cloud-based model iterative optimization and static image gradual deformation lightweighting process of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please refer to Figure 1 This embodiment provides a real-time image quality rendering control system for flexible displays based on edge computing. The system uses edge computing nodes to perceive the physical deformation state of the flexible screen in real time, establishes a correlation model between deformation characteristics and rendering computing power requirements, and achieves dynamic scheduling of rendering tasks and adaptive allocation of frame buffer bandwidth. This solves the problems of image tearing and rendering delay caused by local computing power supply-demand mismatch when the flexible screen deforms.

[0047] In one embodiment, a real-time image quality rendering control system for a flexible display screen based on edge computing includes an edge computing node, data mining software, a frame buffer controller, an edge-side idle computing power cluster, and the flexible display screen itself. The display area of ​​the flexible display screen itself is divided into multiple independent display sub-regions, each with at least one curvature sensor. The curvature sensors are connected to the edge computing node via a serial bus to collect curvature sensing data from each display sub-region in real time. The data mining software is deployed in the non-volatile storage medium of the edge computing node and is loaded into the random access memory of the edge computing node during system initialization. The edge computing node is connected to the edge-side idle computing power cluster via a high-speed interconnect bus. The edge-side idle computing power cluster consists of multiple independent graphics processing units, each containing multiple parallel computing cores. The frame buffer controller is integrated inside the edge computing node and connected to the system bus and frame buffer memory of the edge computing node, responsible for controlling the timing and address mapping of the rendering data stream written to the frame buffer.

[0048] In this embodiment, the overall workflow of the system is as follows: Edge computing nodes read the analog voltage signals output by each curvature sensor in real time via a serial bus, converting the analog voltage signals into digital curvature values ​​to form curvature sensing data sequences for each display sub-region. Data mining software performs association rule mining on the historical curvature change sequences accumulated during system operation and the corresponding historical local pixel rendering computation volume, constructing a curvature-computing power demand prediction model. The curvature-computing power demand prediction model is stored in the random access memory of the edge computing nodes in the form of a multi-dimensional array, containing the mapping relationship of rendering computing power demand for each display sub-region under different deformation states. The edge computing nodes input the real-time collected curvature sensing data into the curvature-computing power demand prediction model and output the predicted rendering computing power demand values ​​for each display sub-region in the next frame. Based on the predicted rendering computing power demand values ​​for each display sub-region, the edge computing nodes dynamically reconstruct the frame buffer data access bandwidth allocation queue and rendering task scheduling sequence. For display sub-regions where the predicted rendering computing power requirement exceeds a first preset threshold, the edge computing node prioritizes routing its corresponding rendering data stream to an idle computing cluster on the edge side for parallel computation. For display sub-regions where the predicted rendering computing power requirement is lower than or equal to the first preset threshold, the edge computing node writes its corresponding rendering data stream into the frame buffer according to the normal timing sequence. The frame buffer controller writes the rendering result data into the corresponding physical address space of the frame buffer based on the reconstructed access bandwidth allocation queue. The display controller reads complete frame data from the frame buffer according to a fixed refresh timing sequence and outputs it to the pixel driving circuit of the flexible display screen for display.

[0049] In this embodiment, the display area of ​​the flexible display screen is divided into an M-row, N-column matrix of sub-regions, where M and N are both positive integers greater than or equal to 2. Each sub-region has the same physical size and a unique identifier. The curvature sensor is a piezoresistive curvature sensor, whose output voltage is linearly correlated with the curvature value of its region. Edge computing nodes read the output voltage of each curvature sensor at a fixed sampling frequency, no less than the refresh rate of the flexible display screen. The edge computing nodes perform analog-to-digital conversion on the read analog voltage signals, with a conversion accuracy of 12 bits, resulting in a digital value ranging from 0 to 4095. The edge computing nodes convert the digital voltage value into the corresponding curvature value using a preset voltage-curvature conversion formula, with the curvature value measured in radians per meter.

[0050] In this embodiment, the data mining software uses the Apriori algorithm for association rule mining. The software first preprocesses the historical curvature change sequence and historical local pixel rendering computation, including data cleaning, data normalization, and data discretization. Data cleaning removes outliers and missing values ​​from the historical data; outliers are identified and removed using the 3σ criterion, and missing values ​​are filled using linear interpolation. Data normalization maps curvature values ​​and rendering computation to the [0,1] interval, eliminating the influence of different dimensions on the association rule mining results. Data discretization divides continuous curvature values ​​and rendering computation into multiple discrete intervals, each corresponding to a discrete value. The software scans the preprocessed historical dataset and generates all frequent itemsets that satisfy the minimum support threshold. Frequent itemsets represent combinations of simultaneously occurring curvature value intervals and rendering computation intervals. Based on these frequent itemsets, the software generates all association rules that satisfy the minimum confidence threshold. The association rule takes the form "when the curvature value is in interval A, the probability that the rendering computation is in interval B is C". The data mining software integrates all generated association rules into a curvature-computing power demand prediction model. The model's input is the curvature value of each display sub-region at the current moment, and the output is the predicted rendering computing power demand value of each display sub-region in the next frame.

[0051] In this embodiment, the process by which the edge computing node dynamically reconstructs the frame buffer data access bandwidth allocation queue based on the predicted rendering computing power demand is as follows: The edge computing node sorts each display sub-region from high to low according to the predicted rendering computing power demand, generating a dynamic priority list. Each element in the dynamic priority list contains the region identifier of the display sub-region and the corresponding predicted rendering computing power demand. Based on the dynamic priority list, the edge computing node divides the total write bandwidth of the frame buffer into multiple bandwidth segments, allocating larger bandwidth segments to the earlier-ranked display sub-regions and smaller bandwidth segments to the later-ranked display sub-regions. Simultaneously, the edge computing node divides the frame buffer write time slots into continuous time slots and discrete time slots, allocating continuous write time slots to the earlier-ranked display sub-regions and discrete write time slots to the later-ranked display sub-regions. The length of continuous write time slots is greater than the length of discrete write time slots, and the number of continuous and discrete write time slots is dynamically adjusted based on the real-time updates of the dynamic priority list.

[0052] In this embodiment, the process by which the edge computing node dynamically reconstructs the rendering task scheduling sequence based on the predicted rendering computing power demand is as follows: The edge computing node divides the image frame data to be rendered into multiple sub-image data blocks according to the division method of display sub-regions. Each sub-image data block corresponds to a display sub-region and contains the geometric information, texture information, and lighting information of all pixels within that display sub-region. The edge computing node generates a corresponding rendering task for each sub-image data block. The rendering task includes a task identifier, a corresponding display sub-region identifier, the storage address of the sub-image data block, and rendering parameters. The edge computing node initializes all rendering tasks as tasks in the regular execution queue and schedules them for execution in a preset order. Based on the predicted rendering computing power demand of each display sub-region, the edge computing node migrates the rendering tasks corresponding to display sub-regions with predicted rendering computing power demand values ​​higher than a second preset threshold from the regular execution queue to the accelerated execution queue. Rendering tasks in the accelerated execution queue have a higher scheduling priority and are preferentially allocated to the computing cores of the edge computing node or the idle computing power clusters on the edge side for execution.

[0053] In this embodiment, the process by which the edge computing node prioritizes routing the rendering data stream of high-computing-power-demand sub-regions to the idle computing power cluster on the edge side is as follows: The edge computing node monitors the operating status of each graphics processing unit in the idle computing power cluster on the edge side in real time, including the real-time load rate of each computing core and the occupancy rate of each level of cache. The edge computing node selects computing cores with a real-time load rate lower than the load rate threshold and a cache occupancy rate lower than the cache occupancy rate threshold to form an available computing power pool. The edge computing node splits each rendering task in the accelerated execution queue into multiple independent rendering subtasks, each rendering subtask corresponding to a pixel block within the display sub-region. The edge computing node distributes the rendering subtasks to the computing cores in the available computing power pool for parallel execution. After each computing core completes its rendering subtask, it returns the rendering result data to the edge computing node. The edge computing node aggregates the result data of all rendering subtasks to form a complete rendering result data stream for the display sub-region and writes it into the corresponding physical address space of the frame buffer.

[0054] This embodiment utilizes data mining software deployed on edge computing nodes to construct a curvature-computing power demand prediction model, enabling advance prediction of rendering computing power requirements for different regions of the flexible screen. By dynamically reconstructing the frame buffer data access bandwidth allocation queue and rendering task scheduling sequence, the rendering data streams of high-computing-power-demand sub-regions are preferentially routed to idle computing power clusters on the edge side, achieving adaptive allocation of rendering computing power. This embodiment solves the problem of screen tearing and rendering delay caused by local rendering computing power supply-demand mismatch during flexible screen deformation, ensuring the continuity and synchronization of screen output.

[0055] In one embodiment, reference Figure 2The process involves edge computing nodes acquiring real-time curvature sensing data from various regions of a flexible screen. This includes extracting spatial gradient features and time-series rates of change from the curvature sensing data, and fusing these features into a multi-dimensional deformation feature vector. Data mining software then performs deep association rule mining based on the multi-dimensional deformation feature vector and historical local pixel rendering computational costs, generating a curvature-computing power demand prediction model in a multi-dimensional feature space. The edge computing nodes input the real-time acquired multi-dimensional deformation feature vector into the curvature-computing power demand prediction model, outputting the predicted computing power demand values ​​for each display sub-region in the next frame.

[0056] In this embodiment, the process of extracting the spatial gradient features of curvature sensing data by the edge computing node is as follows. For any display sub-region within the display area of ​​the flexible display screen... Where i is the row index and j is the column index, the edge computing node obtains its four adjacent display sub-regions. , , , The curvature value. Edge computing nodes calculate and display sub-regions. Curvature gradient values ​​in the horizontal and vertical directions. The horizontal curvature gradient value is used to display the sub-region. curvature value and display sub-region The difference in curvature values ​​is divided by the horizontal distance between the two display sub-regions. The vertical curvature gradient value is the display sub-region. curvature value and display sub-region The difference in curvature values ​​is divided by the vertical distance between the two display sub-regions. The edge computing node combines the horizontal and vertical curvature gradient values ​​into a spatial gradient feature vector.

[0057] In this embodiment, the process of extracting the time-series rate of change of curvature sensing data from the edge computing node is as follows. For any display sub-region... The edge computing node obtains the curvature value at the current time t. and the moment before curvature value ,in The sampling interval for curvature sensing data is defined as follows: The edge computing node calculates the curvature time-varying rate of change for the display sub-region, which is the difference between the curvature value at the current moment and the curvature value at the previous moment, divided by the sampling interval. The edge computing node combines the curvature time-varying rates of change for all display sub-regions into a time-series rate of change vector.

[0058] In this embodiment, the process by which the edge computing node fuses spatial gradient features and time-series rate of change into a multidimensional deformation feature vector is as follows: The edge computing node combines the curvature value, horizontal curvature gradient value, vertical curvature gradient value, and curvature time rate of change of each display sub-region into a four-dimensional feature vector. For a flexible display screen composed of M rows and N columns of display sub-regions, the edge computing node generates a... A multidimensional deformation feature vector. Each element in the multidimensional deformation feature vector corresponds to a deformation feature component of a display sub-region.

[0059] In this embodiment, the data mining software performs deep association rule mining based on multidimensional deformation feature vectors and historical local pixel rendering computation costs as follows: First, the data mining software constructs a historical dataset. Each sample in the historical dataset contains a multidimensional deformation feature vector and the corresponding rendering computation cost vector for each display sub-region. The data mining software uses the FP-growth algorithm for association rule mining. The FP-growth algorithm compresses and stores frequent itemset information by constructing a frequent pattern tree, avoiding the overhead of scanning the dataset multiple times. The data mining software uses each feature component in the multidimensional deformation feature vector and its corresponding rendering computation cost as an item, scanning the historical dataset to construct the FP tree. The data mining software mines all frequent patterns from the FP tree that satisfy the minimum support and minimum confidence thresholds. Frequent patterns represent the association relationship between multiple deformation feature components and rendering computation costs. The data mining software generates association rules based on the frequent patterns and converts the association rules into a curvature-computing power demand prediction model in a multidimensional feature space.

[0060] In this embodiment, the curvature-computing power demand prediction model adopts a multilayer perceptron structure, including an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the dimension of the multidimensional deformation feature vector, i.e. The hidden layer contains two fully connected layers, each with 256 neurons, and uses the ReLU activation function. The number of neurons in the output layer is equal to the number of displayed sub-regions. The activation function of the output layer is a linear function. The input to the curvature-computing power demand prediction model is a real-time multidimensional deformation feature vector, and the output is the predicted rendering computing power demand value for each display sub-region in the next frame.

[0061] In this embodiment, the process by which the edge computing node inputs the real-time acquired multidimensional deformation feature vector into the curvature-computing power demand prediction model is as follows: The edge computing node normalizes the real-time multidimensional deformation feature vector, mapping each feature component to the [0,1] interval. The normalization process uses the min-max normalization method, and for each feature component... Normalized eigencomponents The calculation formula is:

[0062]

[0063] in, This is the minimum value of the feature component in the historical dataset. This represents the maximum value of the feature component in the historical dataset. Edge computing nodes input the normalized multidimensional deformation feature vector into the input layer of the curvature-computing power demand prediction model. The model maps the input features to a high-dimensional feature space through a nonlinear transformation in the hidden layer. The output layer calculates the predicted rendering computing power demand for each display sub-region in the next frame based on the output of the hidden layer. Edge computing nodes perform inverse normalization on the output of the output layer, mapping the predicted values ​​back to the original rendering computing power range.

[0064] This embodiment extracts the spatial gradient features and time-series rate of change from curvature sensing data and fuses them into a multi-dimensional deformation feature vector, comprehensively characterizing the physical deformation state of the flexible screen. By constructing a curvature-computing demand prediction model based on the multi-dimensional deformation feature vector, the accuracy of computing demand prediction is improved. This embodiment enables edge computing nodes to more accurately predict the rendering computing demand of each display sub-region, providing a reliable basis for subsequent dynamic bandwidth allocation and task scheduling.

[0065] In one embodiment, reference Figure 3 The data mining software performs association rule mining on historical curvature change sequences and local pixel rendering computation. This includes calculating the time offset and amplitude ratio coefficient between the extreme points of curvature change rate and the peak points of rendering computation in the historical curvature change sequence, and storing the time offset and amplitude ratio coefficient as an association weight coefficient matrix in the curvature-computing power demand prediction model. Edge computing nodes, based on the extreme points of curvature change rate before the current moment and combined with the association weight coefficient matrix, estimate the position and amplitude of the peak points of rendering computation at a specific frame interval after the current moment, forming the rendering computation demand for each display sub-region in the next frame.

[0066] In this embodiment, the process by which data mining software identifies extreme points of the rate of change of curvature in a historical curvature change sequence is as follows. For any historical rate of change of curvature sequence of a display sub-region, the data mining software uses a sliding window method to traverse the entire sequence. The length of the sliding window is 2k+1, where k is a positive integer. For each point in the sequence, the data mining software compares the rate of change of curvature of that point with the rate of change of curvature of the k points before and after it. If the rate of change of curvature of that point is greater than the rate of change of curvature of the k points before and after it, then that point is a local maximum; if the rate of change of curvature of that point is less than the rate of change of curvature of the k points before and after it, then that point is a local minimum. Local maxima and local minima are collectively referred to as extreme points of the rate of change of curvature. The data mining software records the occurrence time and magnitude of each extreme point of the rate of change of curvature.

[0067] In this embodiment, the process by which the data mining software identifies historical local pixel rendering computation peaks is as follows. For any given display sub-region's historical rendering computation sequence, the data mining software uses the same sliding window method as for identifying extreme points of curvature change rate to identify local maxima in the rendering computation sequence, i.e., rendering computation peaks. The data mining software records the occurrence time and amplitude of each rendering computation peak.

[0068] In this embodiment, the process by which the data mining software calculates the time offset and amplitude ratio coefficient between the extreme points of the rate of curvature change and the peak points of rendering computation is as follows: For each extreme point of the rate of curvature change, the data mining software searches for the corresponding peak point of rendering computation within a preset time window. The length of the preset time window is T, where T is a real number greater than 0. If a unique peak point of rendering computation exists within the preset time window, then the extreme point of the rate of curvature change and the peak point of rendering computation form a pair of associated extreme points. The data mining software calculates the time offset of the pair of associated extreme points. Time offset The peak time of rendering computation is calculated by subtracting the peak time of curvature rate of change from the peak time. Data mining software calculates the amplitude scaling factor *k* for associated extreme point pairs, which is the amplitude of the peak rendering computation point divided by the amplitude of the extreme point of curvature rate of change. The data mining software performs statistical analysis on the time offset and amplitude scaling factor of all associated extreme point pairs to calculate the average time offset. The ratio of the average amplitude .

[0069] In this embodiment, the data mining software stores the time offset and amplitude scaling factor as a correlation weight coefficient matrix in the curvature-computing power demand prediction model. The dimension of the correlation weight coefficient matrix W is... Where M is the number of rows in the displayed sub-region, and N is the number of columns in the displayed sub-region. Each element in the correlation weight coefficient matrix W Corresponding display sub-area Average time offset ,element Corresponding display sub-area Average amplitude ratio coefficient The correlation weight coefficient matrix W is loaded into the random access memory of the edge computing node during the initialization phase of the curvature-computing power demand prediction model.

[0070] In this embodiment, the process by which the edge computing node calculates the peak position and magnitude of the rendering computation cost for a specific frame interval after the current moment, based on the extreme points of the rate of curvature change before the current moment and in conjunction with the associated weight coefficient matrix, is as follows: The edge computing node monitors the rate of curvature change sequence of each display sub-region in real time and identifies all extreme points of the rate of curvature change that occurred before the current moment t. For each extreme point of the rate of curvature change, the edge computing node obtains its occurrence time. With amplitude The edge computing node reads the average time offset corresponding to the display sub-region from the correlation weight coefficient matrix W. The ratio of the average amplitude The estimated time of peak rendering computation for edge computing nodes. Compared with the expected amplitude The calculation formula is:

[0071]

[0072]

[0073] Edge computing node determines estimated occurrence time Is it within the next frame time window after the current time t? The start time of the next frame time window is t, and the end time is... ,in For the frame period of the flexible display screen. If the expected occurrence time... If the peak of the rendering computation occurs within the next frame's time window, it will appear in the next frame. Edge computing nodes will then calculate the expected amplitude. This serves as a predicted value for the rendering computational power requirement of the next frame in this display sub-region. If the expected occurrence time... If the display sub-region is not within the next frame time window, the predicted rendering computational demand for the next frame will be based on the historical average rendering computational load.

[0074] In this embodiment, the edge computing node performs fusion processing on the peak rendering computation points corresponding to multiple extreme points of curvature change rate. If multiple expected peak rendering computation points exist within the next frame time window of the same display sub-region, the edge computing node takes the maximum value among these expected amplitudes as the predicted rendering computational demand value for the next frame of that display sub-region. If no expected peak rendering computation points exist within the next frame time window of the same display sub-region, the edge computing node calculates the predicted rendering computational demand value for the next frame of that display sub-region based on the static mapping relationship between the curvature value at the current moment and the curvature-computing demand prediction model.

[0075] This embodiment establishes a quantitative correlation between deformation occurrence and sudden changes in computing power demand by calculating the time offset and amplitude ratio coefficient between historical curvature change rate extreme points and peak rendering computation points. By predicting the peak rendering computation point of the next frame based on the curvature change rate extreme points before the current moment, the computing power demand prediction achieves both temporal foresight and amplitude determinism. This embodiment can anticipate impending sudden changes in computing power demand in local areas, allowing the system sufficient time to reallocate rendering resources.

[0076] In one embodiment, reference Figure 4 The process involves edge computing nodes dynamically reconstructing the bandwidth allocation queue for frame buffer data based on rendering computing power requirements. This includes sorting the rendering data stream write priorities in descending order based on the predicted rendering computing power requirements of each display sub-region, generating a dynamic priority list. Continuous frame buffer write slots and high-bandwidth transmission channels are allocated to display sub-regions ranked higher in the dynamic priority list, while discrete frame buffer write slots and low-bandwidth transmission channels are allocated to display sub-regions ranked lower. The duty cycle of continuous and discrete frame buffer write slots is adjusted based on real-time updates to the dynamic priority list.

[0077] In this embodiment, the process of the edge computing node generating a dynamic priority list is as follows: The edge computing node obtains the predicted rendering computing power requirements of all display sub-regions, forming a rendering computing power requirement vector. The edge computing node sorts the elements in the rendering computing power requirement vector in descending order, and records the display sub-region identifier corresponding to each element. The edge computing node generates a dynamic priority list, the length of which is equal to the number of display sub-regions. The first element in the dynamic priority list corresponds to the display sub-region with the highest predicted rendering computing power requirement, and the last element corresponds to the display sub-region with the lowest predicted rendering computing power requirement. The edge computing node updates the dynamic priority list every frame period to ensure that the priority list can reflect the changes in computing power requirements of each display sub-region in real time.

[0078] In this embodiment, the process of the edge computing node allocating frame buffer write slots to display sub-regions is as follows: The edge computing node divides each frame period into multiple equal time slices, and each time slice corresponds to one frame buffer write slot. The total number of frame buffer write slots is S, where S is a positive integer greater than or equal to the number of display sub-regions. The edge computing node divides the frame buffer write slots into continuous slot groups and discrete slot groups. The continuous slot group consists of L consecutive time slices, and the discrete slot group consists of SL non-consecutive time slices, where L is a positive integer less than S. The edge computing node allocates slots from the continuous slot group to the top K display sub-regions in the dynamic priority list, with each display sub-region allocated one continuous slot block. The length of the continuous slot block is dynamically adjusted based on the predicted rendering computing power requirement of the display sub-region; the higher the predicted rendering computing power requirement, the longer the allocated continuous slot block. The edge computing node allocates slots from the discrete slot group to the bottom MK display sub-regions in the dynamic priority list, with each display sub-region allocated one discrete slot.

[0079] In this embodiment, the process of allocating transmission channels for the display sub-region by the edge computing node is as follows. Two independent transmission channels are established between the edge computing node and the frame buffer: a high-bandwidth transmission channel and a low-bandwidth transmission channel. The transmission rate of the high-bandwidth transmission channel is... The transmission rate of the low-bandwidth transmission channel is ,in Greater than Edge computing nodes allocate high-bandwidth transmission channels to the top K display sub-regions in the dynamic priority list and low-bandwidth transmission channels to the bottom MK display sub-regions. High-bandwidth and low-bandwidth transmission channels use time-division multiplexing for data transmission, with each channel transmitting the rendering data stream for its assigned display sub-region within its allocated time slot.

[0080] In this embodiment, the process by which the edge computing node adjusts the duty cycle of consecutive frame buffer write slots to discrete frame buffer write slots based on the real-time update of the dynamic priority list is as follows: The edge computing node calculates the ratio of the sum of the predicted rendering computing power requirements of the top K display sub-regions in the dynamic priority list to the sum of the predicted rendering computing power requirements of all display sub-regions, denoted as [the ratio is missing from the original text]. . The value range is [0,1]. Edge computing nodes are determined based on the ratio... Adjust the size of the continuous time slot group and the discrete time slot group. The length L of the continuous time slot group and its ratio. Proportional to the length SL of the discrete time slot group and the ratio Proportional. The adjusted formula is:

[0081]

[0082] in, This is a rounding function. Edge computing nodes reallocate frame buffer write slots to each display sub-region based on the adjusted sizes of the continuous and discrete time slot groups. If the ratio... If the ratio is greater than the first threshold, the edge computing node increases the number of high-bandwidth transmission channels; if the ratio is greater than the first threshold, the edge computing node increases the number of high-bandwidth transmission channels. If the threshold is less than the second threshold, the number of high-bandwidth transmission channels for edge computing nodes is reduced. The first threshold is greater than the second threshold. Examples of time slot and channel allocation under different computing power demand distributions are shown in Table 1.

[0083] Table 1 Examples of time slot and channel allocation under different computing power demand distributions

[0084] Local high computing power demand 0.3 3 7 2 1 Even computing power requirements 0.5 5 5 1 1 Global high computing power requirements 0.8 8 2 3 1

[0085] Table 1 shows examples of time slot and channel allocation under different computing power demand distributions. In this example, the total number S of frame buffer write time slots is 10. When local deformation of the flexible screen leads to local high computing power demand, the ratio... The ratio is 0.3, the continuous time slot length L is 3, the discrete time slot length is 7, and the number of high-bandwidth channels is 2. When the flexible screen is in a flat state, resulting in uniform computing power requirements, the ratio is... The ratio is 0.5, with both continuous and discrete time slot lengths being 5, and both the number of high-bandwidth and low-bandwidth channels being 1. When the flexible screen undergoes global deformation leading to high global computing power requirements, the ratio... The value is 0.8, the continuous time slot length L is 8, the discrete time slot length is 2, and the number of high-bandwidth channels is 3.

[0086] This embodiment generates a dynamic priority list based on predicted rendering computing power requirements, allocating different frame buffer write slots and transmission channels to display sub-regions of different priorities, thus achieving adaptive allocation of frame buffer access bandwidth. By adjusting the duty cycle of continuous and discrete slots in real time based on the dynamic priority list, it ensures that bandwidth resources are used efficiently. This embodiment guarantees that rendering data streams from high-computing-power-demand areas are written to the frame buffer first, avoiding write delays caused by insufficient bandwidth.

[0087] In one embodiment, the process of an edge computing node dynamically reconstructing the rendering task scheduling sequence based on rendering computing power requirements includes extracting the image data corresponding to each display sub-region in the rendering data stream, and calculating the rendering load density of each display sub-region by combining the predicted rendering computing power requirements. Rendering tasks corresponding to display sub-regions with rendering load densities exceeding a preset density threshold are migrated from the regular execution queue of the rendering task scheduling sequence to an accelerated execution queue. The edge computing node then prioritizes routing the data streams corresponding to the rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side.

[0088] In this embodiment, the process by which the edge computing node extracts the image data quantity corresponding to each display sub-region in the rendering data stream is as follows: The rendering data stream contains all image data of the image frame to be rendered, including vertex data, texture data, and shader data. The edge computing node divides the rendering data stream according to the display sub-region division method, obtaining a sub-rendering data stream corresponding to each display sub-region. The edge computing node counts the number of vertices, texture data size, and shader instruction count in each sub-rendering data stream. The edge computing node performs a weighted sum of the number of vertices, texture data size, and shader instruction count to obtain the image data quantity corresponding to that display sub-region. The weighting coefficient is set according to the degree of influence of different types of data on the rendering computation volume; the weighting coefficient for vertex data is... The weighting coefficients for the texture data are The weighting factor for shader instructions is ,in , , All are real numbers greater than 0.

[0089] In this embodiment, the process of edge computing nodes calculating the rendering load density of each display sub-region is as follows. Rendering load density is defined as the amount of rendering computation per unit display area. For each display sub-region... Its rendering load density The calculation formula is:

[0090]

[0091] in, To display sub-regions The corresponding amount of image and data data, To display sub-regions The predicted value of rendering computing power requirements, To display sub-regions The physical area. Edge computing nodes calculate the rendering load density of all display sub-regions, forming a rendering load density matrix.

[0092] In this embodiment, the process by which the edge computing node migrates rendering tasks corresponding to display sub-regions with rendering load densities exceeding a preset density threshold from the regular execution queue to the accelerated execution queue is as follows. The preset density threshold for the edge computing node is... , A real number greater than 0. The edge computing node traverses the rendering load density matrix, assigning the rendering load density of each display sub-region... With preset density threshold Compare. If Greater than If so, then the display sub-area is a high-load area; if Less than or equal to If a region is high-load, then that sub-region is considered a low-load region. The edge computing node retrieves all rendering tasks corresponding to high-load regions from the regular execution queue, removes these tasks from the regular execution queue, and adds them to the head of the accelerated execution queue. Rendering tasks in the accelerated execution queue are sorted in descending order of rendering load density.

[0093] In this embodiment, the edge computing node maintains two independent task scheduling queues: a regular execution queue and an accelerated execution queue. The regular execution queue stores rendering tasks for low-load areas, while the accelerated execution queue stores rendering tasks for high-load areas. The edge computing node uses a priority scheduling algorithm for task scheduling, with tasks in the accelerated execution queue having a higher scheduling priority. The edge computing node prioritizes scheduling tasks in the accelerated execution queue, and only schedules tasks in the regular execution queue when the accelerated execution queue is empty. The edge computing node allocates more computing resources to tasks in the accelerated execution queue, including more computing cores and larger cache space.

[0094] In this embodiment, the process by which the edge computing node prioritizes routing the data stream corresponding to the rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side is as follows: The edge computing node monitors the available computing resources of the idle computing power cluster on the edge side in real time. If the idle computing power cluster on the edge side has available computing resources, the edge computing node prioritizes sending the rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side for execution. If the idle computing power cluster on the edge side does not have available computing resources, the edge computing node allocates the rendering tasks in the accelerated execution queue to the local graphics processing unit for execution. The edge computing node sets a higher transmission priority for the rendering tasks sent to the idle computing power cluster on the edge side to ensure that the rendering data stream can be quickly transmitted to the idle computing power cluster on the edge side. The task scheduling strategies under different rendering load densities are shown in Table 2.

[0095] Table 2 Task scheduling strategies under different rendering load densities

[0096] Execution acceleration queue Edge-side idle computing power cluster 4 compute cores, 256KB cache Execution acceleration queue Local graphics processing unit 2 compute cores, 128KB cache Regular execution queue Local graphics processing unit One compute core, 64KB cache

[0097] Table 2 shows the task scheduling strategies under different rendering load densities. In this example, the preset density threshold is... When the rendering load density is greater than When the rendering task is placed in the accelerated execution queue, it is routed to an idle computing cluster on the edge side for execution, allocated 4 computing cores and 256KB of cache. When the rendering load density is greater than and less than or equal to At this time, rendering tasks are placed in the accelerated execution queue and executed by the local graphics processing unit, allocated 2 compute cores and 128KB cache. When the rendering load density is less than or equal to At that time, the rendering task is placed in the regular execution queue and executed by the local graphics processing unit, which is allocated 1 computing core and 64KB cache.

[0098] This embodiment calculates the rendering load density by combining image data volume with predicted rendering computing power requirements, thus more accurately assessing the actual rendering load of each display sub-region. By migrating rendering tasks corresponding to high-load areas to an accelerated execution queue and prioritizing routing them to idle computing power clusters on the edge side, dynamic scheduling of rendering tasks is achieved. This embodiment fully utilizes idle computing resources on the edge side, improving the overall rendering performance of the system.

[0099] In one embodiment, reference Figure 5 When the time offset is less than a preset time threshold, the flexible screen is determined to have undergone a rapid deformation. The edge computing node triggers the interruption and reconstruction mechanism of the rendering pipeline, freezing any unexecuted rendering tasks in the current rendering task scheduling sequence and truncating the rendering data stream being written to the frame buffer. The edge computing node re-collects curvature sensing data after the rapid deformation, recalculates the predicted rendering computing power requirement for the next frame based on the curvature-computing power requirement prediction model, overwrites the current frame buffer data access bandwidth allocation queue and rendering task scheduling sequence with the recalculated predicted rendering computing power requirement, and resumes the writing of the truncated rendering data stream.

[0100] In this embodiment, the process by which the edge computing node determines that the flexible screen has undergone a rapid deformation is as follows: The edge computing node calculates in real time the time offset between the extreme point of the rate of curvature change of each display sub-region and the corresponding peak point of the rendering computation. The preset time threshold for edge computing nodes is... , A real number greater than 0. If the time offset of any display sub-region... Less than the preset time threshold If a display sub-region experiences a rapid deformation, the entire flexible screen is determined to have undergone a rapid deformation. If more than a preset number of display sub-regions experience rapid deformation, the entire flexible screen is determined to have undergone rapid deformation. The preset number is a positive integer greater than or equal to 1.

[0101] In this embodiment, the process of the edge computing node triggering the interruption and reconstruction mechanism of the rendering pipeline is as follows: When it is determined that the flexible screen has undergone a sudden deformation, the edge computing node sends an interrupt signal to each stage of the rendering pipeline. Upon receiving the interrupt signal, the geometry processing stage, rasterization stage, and pixel shading stage of the rendering pipeline immediately stop the currently executing operation. The edge computing node freezes the rendering tasks that have not yet been executed in the current rendering task scheduling sequence and saves the status information of these tasks to the edge computing node's random access memory. The status information includes the task identifier, task execution progress, context data, and register values. The edge computing node sends a truncation signal to the frame buffer controller. Upon receiving the truncation signal, the frame buffer controller immediately stops the ongoing rendering data stream writing operation and records the current writing position and the amount of data already written.

[0102] In this embodiment, the process of the edge computing node re-acquiring curvature sensing data after abrupt deformation and recalculating the rendering computing power requirements is as follows: The edge computing node increases the sampling frequency of the curvature sensor to acquire curvature sensing data of each display sub-region at a higher rate. The edge computing node filters the acquired curvature sensing data to remove noise caused by abrupt deformation. The filtering process uses a median filtering algorithm with a filtering window size of 3×3. The edge computing node extracts the spatial gradient features and time series rate of change of the filtered curvature sensing data and fuses them into a new multidimensional deformation feature vector. The edge computing node inputs the new multidimensional deformation feature vector into the curvature-computing power requirement prediction model and outputs the recalculated rendering computing power requirement prediction values ​​for each display sub-region in the next frame.

[0103] In this embodiment, the process by which the edge computing node overwrites the current frame buffer data access bandwidth allocation queue and rendering task scheduling sequence based on the recalculated rendering computing power demand prediction is as follows: The edge computing node clears the current frame buffer data access bandwidth allocation queue and rendering task scheduling sequence. Based on the recalculated rendering computing power demand prediction, the edge computing node generates a new dynamic priority list. Based on the new dynamic priority list, the edge computing node reconstructs the frame buffer data access bandwidth allocation queue, allocating new frame buffer write time slots and transmission channels to each display sub-region. Based on the new dynamic priority list and the rendering load density of each display sub-region, the edge computing node reconstructs the rendering task scheduling sequence, adding rendering tasks corresponding to high-load regions to the accelerated execution queue and adding rendering tasks corresponding to low-load regions to the regular execution queue.

[0104] In this embodiment, the process of the edge computing node resuming the writing of the truncated rendering data stream is as follows: The edge computing node obtains the write position and the amount of data already written from the frame buffer controller. Starting from the truncated position, the edge computing node continues to transmit the remaining rendering data stream to the frame buffer controller. The frame buffer controller writes the remaining rendering data stream into the corresponding physical address space of the frame buffer according to the new access bandwidth allocation queue. For display sub-regions that have been partially written to the frame buffer, the edge computing node checks whether the written data is complete. If the written data is incomplete, the edge computing node retransmits the complete rendering data stream for that display sub-region to the frame buffer controller.

[0105] This embodiment enables rapid response to sudden deformations of flexible screens by triggering an interruption and reconstruction mechanism in the rendering pipeline upon detecting abrupt deformation. By freezing unexecuted rendering tasks and truncating the data stream being written, outdated rendering data is prevented from being written to the frame buffer. By re-acquiring curvature sensor data and recalculating rendering computational requirements, the new scheduling strategy ensures that it can adapt to the state after abrupt deformation. This embodiment effectively avoids severe screen tearing and rendering errors that occur during abrupt deformation.

[0106] In one embodiment, the process of an edge computing node preferentially routing the data streams corresponding to rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side includes monitoring the real-time load rate and cache occupancy rate of each computing core in the idle computing power cluster on the edge side, and selecting computing cores whose real-time load rate and cache occupancy rate are both lower than the corresponding preset thresholds to form an available computing power pool. Based on the predicted computing power demand of each rendering task in the accelerated execution queue, the rendering tasks in the accelerated execution queue are split into multiple rendering sub-tasks, and the multiple rendering sub-tasks are distributed to computing cores in the available computing power pool for parallel execution. The rendering result data streams output by each computing core are then aggregated and written to the frame buffer.

[0107] In this embodiment, the process by which the edge computing node monitors the real-time load rate and cache occupancy of each computing core in the idle computing power cluster on the edge side is as follows: Each computing core in the idle computing power cluster on the edge side has a built-in performance monitoring unit, which can collect the operating status information of the computing core in real time. The performance monitoring unit collects performance indicators such as instruction execution rate, pipeline pause rate, and cache hit rate of the computing core at fixed time intervals. The edge computing node periodically reads the performance monitoring data of each computing core through the high-speed interconnect bus. The edge computing node calculates the real-time load rate of the computing core based on the instruction execution rate and pipeline pause rate, with the real-time load rate ranging from [0,1]. The edge computing node calculates the cache occupancy of the computing core based on the cache hit rate and total cache capacity, with the cache occupancy ranging from [0,1].

[0108] In this embodiment, the process of selecting available computing cores from edge computing nodes to form an available computing power pool is as follows. The preset load rate threshold for edge computing nodes is... The preset cache occupancy threshold is ,in and All are real numbers greater than 0 and less than 1. The edge computing node iterates through all computing cores in the idle computing power cluster on the edge side, comparing the real-time load rate L of each computing core with the load rate threshold. Compare the cache occupancy C of each computing core with the cache occupancy threshold. Compare them. If L is less than... And C is less than If a computing core is found to be available, it is considered an available computing core. Edge computing nodes add all available computing cores to the available computing power pool and record the identifier, computing power, and remaining cache space of each available computing core. The size of the available computing power pool dynamically changes according to the operating status of the idle computing power cluster on the edge side.

[0109] In this embodiment, the process by which the edge computing node splits a rendering task in the accelerated execution queue into multiple rendering subtasks is as follows: The edge computing node retrieves a rendering task from the accelerated execution queue and obtains the corresponding display sub-region identifier and the predicted computing power requirement. Based on the size of the available computing power pool and the computing power of each available computing core, the edge computing node determines the number of subtasks that the rendering task needs to be split into. Number of subtasks The calculation formula is:

[0110]

[0111] Where P is the predicted computing power requirement for the rendering task. This represents the average computing power of all computing cores in the available computing pool. This is a rounding function. Edge computing nodes uniformly divide the rendering area of ​​the displayed sub-region into... Each sub-region corresponds to a rendering subtask. Each rendering subtask contains the corresponding sub-region's image data, rendering parameters, and target address information.

[0112] In this embodiment, the process by which the edge computing node distributes multiple rendering subtasks to computing cores in the available computing power pool for parallel execution is as follows: The edge computing node uses a round-robin scheduling algorithm to distribute the rendering subtasks to computing cores in the available computing power pool. The edge computing node sequentially sends each rendering subtask to the next computing core in the available computing power pool. If a computing core completes its current rendering subtask, the edge computing node immediately sends the next waiting rendering subtask to that computing core. The edge computing node sets a timeout for each rendering subtask. If a rendering subtask is not completed within the timeout period, the edge computing node redistributes the subtask to other available computing cores.

[0113] In this embodiment, the process of the edge computing node aggregating the rendering result data streams output by each computing core and writing them into the frame buffer is as follows: After each computing core completes its rendering subtask, it returns the rendering result data to the edge computing node via a high-speed interconnect bus. The rendering result data includes the color and depth values ​​of all pixels within the sub-region. The edge computing node concatenates the result data of all rendering subtasks into a complete rendering result data stream for the display sub-region, according to the spatial order of the sub-regions. The edge computing node sends the complete rendering result data stream to the frame buffer controller. The frame buffer controller allocates queues based on the frame buffer data access bandwidth and writes the rendering result data stream into the physical address space of the corresponding display sub-region in the frame buffer.

[0114] This embodiment monitors the operating status of each computing core in the idle computing power cluster on the edge side, and filters out available computing resources to form an available computing power pool. By splitting the rendering task into multiple sub-tasks and distributing them to the available computing power pool for parallel execution, the idle computing resources on the edge side are fully utilized. By aggregating the rendering results from each computing core and writing them to the frame buffer, the integrity and correctness of the rendering data are ensured. This embodiment improves the rendering speed in high-load areas and shortens the processing time of the rendering data stream.

[0115] In one embodiment, the process of an edge computing node transmitting rendering data streams through a high-bandwidth transmission channel and a low-bandwidth transmission channel includes performing lossless compression encoding on the rendering data stream entering the high-bandwidth transmission channel and inserting a high-priority transmission identifier and a first target physical address field into the header of the compressed data packet. For the rendering data stream entering the low-bandwidth transmission channel, lossy compression encoding is performed, discarding detail components below the visual resolution threshold, and a low-priority transmission identifier and a second target physical address field are inserted into the header of the compressed data packet. The frame buffer controller arbitrates the execution order of writing corresponding compressed data packets into the frame buffer based on the high-priority and low-priority transmission identifiers.

[0116] In this embodiment, the process of the edge computing node performing lossless compression encoding on the rendering data stream entering the high-bandwidth transmission channel is as follows: The edge computing node uses the LZ4 lossless compression algorithm to compress the rendering data stream of the high-bandwidth transmission channel. The LZ4 algorithm is a dictionary-based lossless compression algorithm with extremely high compression and decompression speeds. The edge computing node divides the rendering data stream into multiple fixed-size data blocks, each 64KB in size. The edge computing node independently performs LZ4 compression encoding on each data block to generate compressed data blocks. The edge computing node adds a block identifier, block length, and checksum information to the header of each compressed data block. The block identifier identifies the type of compressed data block, the block length indicates the size of the compressed data block, and the checksum is used to detect data errors that may occur during transmission.

[0117] In this embodiment, the process by which the edge computing node inserts a high-priority transmission identifier and a first target physical address segment into the header of the compressed data packet is as follows: The edge computing node encapsulates multiple compressed data blocks into a single transmission data packet. The header of the transmission data packet includes a transmission identifier field, a target physical address field, a data packet length field, and a sequence number field. The edge computing node writes a high-priority transmission identifier into the transmission identifier field; the high-priority transmission identifier is a preset binary value. The edge computing node writes the starting physical address of the display sub-region corresponding to the rendered data stream in the frame buffer into the first target physical address segment. The first target physical address segment includes the starting physical address and data length information. The edge computing node assigns a unique sequence number to each transmission data packet and writes it into the sequence number field, which is used by the receiving end to sort and reassemble the data packets.

[0118] In this embodiment, the process of the edge computing node performing lossy compression encoding on the rendering data stream entering the low-bandwidth transmission channel is as follows: The edge computing node uses a lossy compression algorithm based on the human visual system to compress the rendering data stream in the low-bandwidth transmission channel. This algorithm utilizes the characteristic that the human visual system is insensitive to high-frequency detail components, discarding detail components in the rendering data stream that are below the visual resolution threshold. The edge computing node first converts the rendering data stream from the RGB color space to the YUV color space. The YUV color space separates luminance information Y from chrominance information U and V. The human visual system is more sensitive to luminance information than to chrominance information. The edge computing node compresses the luminance component Y at a lower ratio and the chrominance components U and V at a higher ratio. The edge computing node uses discrete cosine transform to convert the pixel data in the spatial domain into coefficient data in the frequency domain. The edge computing node quantizes the coefficient data in the frequency domain, discarding high-frequency coefficients with amplitudes below the quantization threshold. The quantization threshold is set according to the visual resolution threshold. The edge computing node performs entropy encoding on the quantized coefficient data to generate compressed data.

[0119] In this embodiment, the process by which edge computing nodes discard detail components in the rendering data stream that are below the visual resolution threshold is as follows. The visual resolution threshold refers to the minimum brightness and chromaticity differences that the human visual system can distinguish. The visual resolution threshold is related to the curvature value of the display sub-region, viewing distance, and ambient brightness. The edge computing node calculates the pixel density change rate of the current display sub-region based on its curvature value. The higher the pixel density change rate, the higher the visual resolution threshold. The edge computing node adjusts the visual resolution threshold based on the viewing distance and ambient brightness. The farther the viewing distance, the higher the visual resolution threshold; the higher the ambient brightness, the higher the visual resolution threshold. The edge computing node determines the quantization threshold based on the adjusted visual resolution threshold. The edge computing node sets high-frequency coefficients with amplitudes below the quantization threshold to 0, thereby discarding detail components below the visual resolution threshold.

[0120] In this embodiment, the process of the edge computing node inserting a low-priority transmission identifier and a second target physical address segment into the header of the compressed data packet is similar to the process of inserting a high-priority transmission identifier and a first target physical address segment. The edge computing node writes the low-priority transmission identifier into the transmission identifier field of the transmitted data packet. The low-priority transmission identifier is a different preset binary value than the high-priority transmission identifier. The edge computing node then writes the starting physical address of the display sub-region corresponding to the rendered data stream in the frame buffer into the second target physical address segment.

[0121] In this embodiment, the process by which the frame buffer controller arbitrates the execution order of writing compressed data packets into the frame buffer based on the high-priority transmission identifier and the low-priority transmission identifier is as follows: The frame buffer controller internally sets up two independent receive buffers: a high-priority buffer and a low-priority buffer. The high-priority buffer stores compressed data packets with high-priority transmission identifiers, and the low-priority buffer stores compressed data packets with low-priority transmission identifiers. The frame buffer controller uses a fixed-priority arbitration algorithm to arbitrate the data packets in the two buffers. The frame buffer controller processes data packets in the high-priority buffer first, and only processes data packets in the low-priority buffer when the high-priority buffer is empty. The frame buffer controller decompresses the compressed data packets and writes the decompressed pixel data into the corresponding physical address space of the frame buffer. A comparison of the compression encoding parameters for the high-bandwidth and low-bandwidth transmission channels is shown in Table 3.

[0122] Table 3 Comparison of compression coding parameters for high-bandwidth and low-bandwidth transmission channels

[0123] Parameters High-bandwidth transmission channel Low bandwidth transmission channel Compression algorithm LZ4 lossless compression HVS-based lossy compression Compression ratio 2:1 3:1 10:1 20:1 Color space RGB YUV Quantification none Yes, adaptive quantization threshold Details preserved Keep all Discard details below the visual resolution threshold Transmission priority high Low

[0124] Table 3 shows a comparison of compression coding parameters for high-bandwidth and low-bandwidth transmission channels. The high-bandwidth transmission channel uses the LZ4 lossless compression algorithm with a compression ratio of 2:1 to 3:1, preserving all image details and having a higher transmission priority. The low-bandwidth transmission channel uses a lossy compression algorithm based on the human visual system with a compression ratio of 10:1 to 20:1, discarding details below the visual resolution threshold and having a lower transmission priority.

[0125] This embodiment achieves efficient data transmission under bandwidth-constrained conditions by employing different compression encoding methods for rendering data streams of different priorities. Lossless compression is performed on the data stream of high-bandwidth transmission channels to ensure image quality in areas with high computing power requirements. Lossy compression is performed on the data stream of low-bandwidth transmission channels, discarding detail components below the visual resolution threshold and reducing the amount of transmitted data. By inserting a priority transmission identifier into the packet header, the frame buffer controller can arbitrate the writing order of data packets according to priority, ensuring that high-priority data is written first.

[0126] In one embodiment, the edge computing-based flexible display real-time image quality rendering control system further includes a cloud server, which deploys cloud-integrated application runtime support platform software. Edge computing nodes upload logs of the mapping relationship between curvature sensing data and rendering computing power requirements to the cloud-integrated application runtime support platform software. The cloud-integrated application runtime support platform software aggregates the mapping relationship logs uploaded by multiple edge computing nodes, periodically iteratively trains the global curvature-computing power requirement baseline model, and distributes the updated global curvature-computing power requirement baseline model to the edge computing nodes. The edge computing nodes use the updated global curvature-computing power requirement baseline model to replace the excess weight parameters in the curvature-computing power requirement prediction model.

[0127] In this embodiment, the process of the edge computing node generating and uploading a mapping log between curvature sensing data and rendering computational requirements is as follows: During operation, the edge computing node records the curvature sensing data and corresponding actual rendering computation volume for each display sub-region in real time. At preset time intervals, the edge computing node organizes the recorded curvature sensing data and rendering computation volume data into a mapping log. The mapping log uses a structured data format, including fields such as timestamp, display sub-region identifier, curvature value, curvature change rate, and actual rendering computation volume. The edge computing node uploads the mapping log to the cloud-based converged application runtime support platform software on the cloud server via a wireless network. When uploading the log, the edge computing node encrypts the log data to ensure data transmission security.

[0128] In this embodiment, the process by which the cloud-integrated application runtime support platform software aggregates mapping relationship logs uploaded by multiple edge computing nodes is as follows: The cloud-integrated application runtime support platform software is deployed in a distributed computing cluster on a cloud server and can simultaneously receive mapping relationship logs from multiple edge computing nodes. The cloud-integrated application runtime support platform software stores the received log data in a distributed database. The distributed database uses a sharded storage method, storing log data from different edge computing nodes in different database shards. The cloud-integrated application runtime support platform software performs data cleaning and preprocessing on the stored log data, removing abnormal and duplicate data and unifying the data format. The cloud-integrated application runtime support platform software integrates the preprocessed log data into a global training dataset.

[0129] In this embodiment, the process of periodically iteratively training the global curvature-computing power demand baseline model using the cloud-integrated application runtime support platform software is as follows: The global curvature-computing power demand baseline model adopts the same multilayer perceptron structure as the curvature-computing power demand prediction model local on the edge computing node. Every preset training cycle, the cloud-integrated application runtime support platform software iteratively trains the global curvature-computing power demand baseline model using the global training dataset. The training process employs the backpropagation algorithm and a stochastic gradient descent optimizer. The loss function uses the mean squared error loss function, calculated as follows:

[0130]

[0131] in, The number of training samples, # represents the actual rendering computation cost for the i-th sample, and #imgpt120# represents the rendering computation cost predicted by the model. The cloud-based integrated application runtime support platform software divides the training dataset into a training set and a validation set. The training set is used to update model parameters, and the validation set is used to evaluate the model's generalization ability. Training stops when the model's loss function value on the validation set no longer decreases.

[0132] In this embodiment, the process by which the cloud-integrated application runtime support platform software distributes the updated global curvature-computing power requirement baseline model to edge computing nodes is as follows: After completing model training, the cloud-integrated application runtime support platform software saves the updated model parameters as a model file. The model file contains the model's structural information and weight parameters. The cloud-integrated application runtime support platform software distributes the model file to all connected edge computing nodes via a wireless network. When distributing the model file, the cloud-integrated application runtime support platform software digitally signs the model file to ensure its integrity and authenticity. After receiving the model file, the edge computing nodes verify the validity of the digital signature. If the digital signature is valid, the edge computing nodes save the model file to a non-volatile storage medium.

[0133] In this embodiment, the process by which the edge computing node replaces the outdated weight parameters in the curvature-computing power demand prediction model with the updated global curvature-computing power demand baseline model is as follows: During system initialization or at a preset model update time, the edge computing node loads the updated global curvature-computing power demand baseline model. The edge computing node compares the weight parameters of the global curvature-computing power demand baseline model with the weight parameters of the local curvature-computing power demand prediction model. The edge computing node identifies layers in the local model whose weight parameters have changed beyond a preset threshold; these layers are considered outdated weight parameters. The edge computing node replaces the outdated weight parameters in the local model with the weight parameters of the corresponding layers in the global model. The edge computing node retains layers in the local model whose weight parameters have not changed beyond the preset threshold; these layers contain the personalized characteristics of the specific flexible screen hardware.

[0134] This embodiment uploads mapping relationship logs from edge computing nodes to a cloud server. The cloud server aggregates data from multiple edge nodes for global model training, enabling continuous optimization of the curvature-computing demand prediction model. By distributing the updated global model to edge computing nodes and replacing excessive weight parameters in the local model, the edge-side prediction model can continuously adapt to the deformation and aging characteristics of different flexible screen hardware and complex scene changes. This embodiment maintains the long-term stability of prediction accuracy and improves the system's adaptability and robustness.

[0135] In one embodiment, reference Figure 6 The cloud server also deploys cloud computing software. When the flexible screen is in a static state and undergoes slow deformation, the edge computing nodes only extract the contour coordinates of areas in the curvature sensing data whose rate of change is lower than the slow deformation threshold, and upload these coordinates to the cloud computing software. The cloud computing software calculates the pixel offset caused by the slow deformation based on the contour coordinates, generates a local coordinate remapping table, and sends it to the edge computing nodes. The edge computing nodes, based on the local coordinate remapping table, read the pixel data of the original static image from the frame buffer, perform coordinate translation operations on the pixel data, and then write it back to the corresponding physical address in the frame buffer.

[0136] In this embodiment, the process by which the edge computing node determines that the flexible screen is in a static state and undergoing slow deformation is as follows: The edge computing node compares the image data of the current image frame to be displayed with that of the previous frame, and calculates the pixel difference value between the two frames. The pixel difference value is the sum of the absolute values ​​of the differences in the color values ​​of corresponding pixels in the two frames. If the pixel difference value is lower than a preset static state threshold, the flexible screen is determined to be in a static state. The edge computing node calculates the curvature time change rate of each display sub-region. If the curvature time change rate of all display sub-regions is lower than a preset gradual change threshold, the flexible screen is determined to undergo slow deformation. The gradual change threshold is less than the curvature change rate corresponding to the time threshold used in the abrupt deformation determination.

[0137] In this embodiment, the process by which the edge computing node extracts the contour coordinates of regions with a curvature change rate lower than the gradual change threshold from the curvature sensing data is as follows: The edge computing node traverses all display sub-regions and filters out those with a curvature change rate lower than the gradual change threshold. These display sub-regions constitute slowly deformable regions. The edge computing node extracts the boundary point coordinates of the slowly deformable regions to form the region contour. The region contour consists of a series of continuous pixel coordinate points. The edge computing node simplifies the region contour, removing redundant coordinate points and retaining key contour feature points. The simplification process uses the Douglas-Puk algorithm, which can reduce the number of contour points while preserving the basic shape of the contour.

[0138] In this embodiment, the process of an edge computing node uploading region contour coordinates to cloud computing software is as follows: The edge computing node encapsulates the simplified region contour coordinate data into an upload data packet. The upload data packet includes a timestamp, flexible screen identifier, number of region contour points, and a list of region contour coordinates. The edge computing node sends the upload data packet to the cloud computing software on the cloud server via a wireless network. The edge computing node uses a low-power transmission mode to upload data, reducing system power consumption.

[0139] In this embodiment, the process by which the cloud computing software calculates the pixel offset caused by slow deformation based on the region contour coordinates is as follows: The cloud computing software determines the range and shape of the slowly deforming region based on the region contour coordinates. The cloud computing software establishes a physical deformation model of the flexible screen, which adopts a thin-plate spline function model. The thin-plate spline function model can accurately describe the bending and folding deformation of the flexible screen. The cloud computing software uses the region contour coordinates as constraint points of the deformation model and solves for the parameters of the deformation model. Based on the solved deformation model parameters, the cloud computing software calculates the offset of each pixel within the slowly deforming region. The pixel offset includes horizontal and vertical offsets.

[0140] In this embodiment, the process of the cloud computing software generating a local coordinate remapping table and sending it to the edge computing nodes is as follows: The cloud computing software generates a local coordinate remapping table based on the calculated offset of each pixel. The local coordinate remapping table is a two-dimensional array, where each element corresponds to the mapping relationship between the original coordinates and the target coordinates of a pixel within the slowly deformed region. The size of the local coordinate remapping table is the same as the number of pixels in the slowly deformed region. The cloud computing software compresses the local coordinate remapping table into a compressed file and sends it to the edge computing nodes via a wireless network. After receiving the compressed file, the edge computing nodes decompress it to obtain the complete local coordinate remapping table.

[0141] In this embodiment, the process by which the edge computing node performs coordinate translation on pixel data based on the local coordinate remapping table is as follows: The edge computing node reads the pixel data of the original still image from the frame buffer and stores it in a temporary buffer. The edge computing node traverses each mapping entry in the local coordinate remapping table. For each mapping entry, the edge computing node reads the pixel data corresponding to the original coordinates from the temporary buffer. The edge computing node writes the read pixel data to the physical address corresponding to the target coordinates in the frame buffer. After processing all mapping entries, the edge computing node clears the temporary buffer. The pixel data stored in the frame buffer is the still image data after coordinate translation. The display controller reads data from the frame buffer and outputs it for display according to a fixed refresh sequence.

[0142] This embodiment avoids transmitting the entire image data by extracting only the outline coordinates of a region when the flexible screen is stationary and undergoing slow deformation. The cloud calculates pixel offsets and generates a local coordinate remapping table; edge computing nodes only need to perform simple coordinate translation operations to update the image. This embodiment avoids full re-rendering of the static image under slow deformation, reducing the overall power consumption and data processing load of the system in low-dynamic scenarios.

Claims

1. A real-time image quality rendering control system for flexible displays based on edge computing, characterized in that, include: Edge computing nodes collect real-time curvature sensing data from various areas of the flexible screen; Data mining software, deployed on the edge computing node, performs association rule mining on the historical curvature change sequence and local pixel rendering computation volume to construct a curvature-computing power demand prediction model. The edge computing node processes the curvature sensing data according to the curvature-computing power demand prediction model to predict the rendering computing power demand of each display sub-region in the next frame. The edge computing node dynamically reconstructs the frame buffer data access bandwidth allocation queue and the rendering task scheduling sequence based on the rendering computing power requirements; The edge computing node prioritizes routing the rendering data stream of high-computing-power-demand sub-regions to the idle computing power cluster on the edge side, and writes the rendering data stream of low-computing-power-demand sub-regions into the frame buffer according to the normal timing.

2. The system according to claim 1, characterized in that, The process of the edge computing node collecting curvature sensing data of each region of the flexible screen in real time includes extracting the spatial gradient features and time series rate of change of the curvature sensing data, and fusing the spatial gradient features and the time series rate of change into a multidimensional deformation feature vector. The data mining software performs deep association rule mining based on the multidimensional deformation feature vector and the historical local pixel rendering computation volume to generate the curvature-computing power demand prediction model in the multidimensional feature space. The edge computing node inputs the multidimensional deformation feature vector obtained in real time into the curvature-computing power demand prediction model and outputs the predicted computing power demand value of each display sub-region in the next frame.

3. The system according to claim 2, characterized in that, The data mining software performs association rule mining on the historical curvature change sequence and the local pixel rendering computation volume, including calculating the time offset and amplitude ratio coefficient between the extreme point of curvature change rate and the peak point of rendering computation volume in the historical curvature change sequence, and storing the time offset and amplitude ratio coefficient as an association weight coefficient matrix in the curvature-computing power demand prediction model. The edge computing node, based on the extreme point of curvature change rate before the current moment and combined with the associated weight coefficient matrix, calculates the peak position and amplitude of the rendering computation amount in a specific frame interval after the current moment, thus forming the rendering computational power requirement of each display sub-region in the next frame.

4. The system according to claim 1, characterized in that, The process of the edge computing node dynamically reconstructing the access bandwidth allocation queue of frame buffer data according to the rendering computing power requirements includes sorting the rendering data stream writing priority in descending order according to the rendering computing power requirements of each display sub-region, and generating a dynamic priority list. The display sub-regions ranked first in the dynamic priority list are allocated continuous frame buffer write time slots and high-bandwidth transmission channels, while the display sub-regions ranked later are allocated discrete frame buffer write time slots and low-bandwidth transmission channels. The duty cycle of the continuous frame buffer write time slots and the discrete frame buffer write time slots is adjusted according to the real-time update of the dynamic priority list.

5. The system according to claim 4, characterized in that, The process of the edge computing node dynamically reconstructing the rendering task scheduling sequence according to the rendering computing power requirement includes extracting the image data quantity corresponding to each display sub-region in the rendering data stream, and calculating the rendering load density of each display sub-region in combination with the rendering computing power requirement prediction value. The rendering tasks corresponding to the display sub-regions whose rendering load density exceeds a preset density threshold are migrated from the regular execution queue of the rendering task scheduling sequence to the accelerated execution queue. The edge computing node prioritizes routing the data streams corresponding to the rendering tasks in the accelerated execution queue to the idle computing power cluster on the edge side.

6. The system according to claim 3, characterized in that, When the time offset is less than a preset time threshold, it is determined that the flexible screen has undergone a rapid deformation; The edge computing node triggers the interruption and reconstruction mechanism of the rendering pipeline, freezes the rendering tasks that have not been executed in the current rendering task scheduling sequence, and truncates the rendering data stream that is being written to the frame buffer. The edge computing node re-collects curvature sensing data after the sharp deformation, recalculates the rendering computing power requirement prediction value for the next frame based on the curvature-computing power requirement prediction model, overwrites the current frame buffer data access bandwidth allocation queue and rendering task scheduling sequence with the recalculated rendering computing power requirement prediction value, and restores the writing of the truncated rendering data stream.

7. The system according to claim 5, characterized in that, The process by which the edge computing node prioritizes routing the data stream corresponding to the rendering task in the accelerated execution queue to the idle computing power cluster on the edge side includes monitoring the real-time load rate and cache occupancy rate of each computing core in the idle computing power cluster on the edge side, and selecting computing cores whose real-time load rate and cache occupancy rate are both lower than the corresponding preset threshold to form an available computing power pool. Based on the predicted computing power requirements of each rendering task in the accelerated execution queue, the rendering tasks in the accelerated execution queue are split into multiple rendering subtasks. The multiple rendering subtasks are distributed to the computing cores in the available computing power pool for parallel execution, and the rendering result data streams output by each computing core are aggregated and written into the frame buffer.

8. The system according to claim 4, characterized in that, The process of the edge computing node transmitting rendering data streams through the high-bandwidth transmission channel and the low-bandwidth transmission channel includes performing lossless compression encoding on the rendering data streams entering the high-bandwidth transmission channel and inserting a high-priority transmission identifier and a first target physical address segment into the header of the compressed data packet; Lossy compression encoding is performed on the rendering data stream entering the low-bandwidth transmission channel, discarding detail components in the rendering data stream that are below the visual resolution threshold, and inserting a low-priority transmission identifier and a second target physical address segment into the header of the compressed data packet. The frame buffer controller arbitrates the execution order of writing corresponding compressed data packets into the frame buffer based on the high-priority transmission identifier and the low-priority transmission identifier.

9. The system according to claim 1, characterized in that, It also includes a cloud server, which is equipped with cloud-integrated application runtime support platform software; The edge computing node uploads the mapping relationship log between the curvature sensing data and the rendering computing power requirement to the cloud-based integrated application runtime support platform software; The cloud-integrated application operation support platform software aggregates mapping relationship logs uploaded by multiple edge computing nodes, performs periodic iterative training on the global curvature-computing power demand baseline model, and distributes the updated global curvature-computing power demand baseline model to the edge computing nodes. The edge computing node replaces the excess weighting parameters in the curvature-computing demand prediction model with the updated global curvature-computing demand baseline model.

10. The system according to claim 9, characterized in that, The cloud server is also deployed with cloud computing software. When the flexible screen is in a static state and undergoes slow deformation, the edge computing node only extracts the contour coordinates of the region whose rate of change is lower than the slow change threshold from the curvature sensing data, and uploads the contour coordinates of the region to the cloud computing software. The cloud computing software calculates the pixel offset caused by slow deformation based on the region contour coordinates, generates a local coordinate remapping table, and sends it to the edge computing node. The edge computing node reads the pixel data of the original still image from the frame buffer according to the local coordinate remapping table, performs a coordinate translation operation on the pixel data, and then writes it back to the corresponding physical address of the frame buffer.

Images