A three-dimensional engine rendering control method based on adaptive deep learning

Abstract

The application discloses a three-dimensional engine rendering control method based on adaptive deep learning. By collecting scene geometric complexity, texture complexity, processor utilization, video memory occupation and historical rendering performance data, a rendering state feature vector is constructed, and an initial rendering parameter is generated by inputting the prediction model; on this basis, an exploration search mechanism is introduced to perform multidimensional trial on the parameter, and candidate parameters are screened in combination with the trial rendering result; further based on a directional search strategy, the parameter optimization direction is determined and iterative search is performed to obtain refined rendering parameters; the optimized parameters are applied to the three-dimensional engine to complete rendering processing, and performance and image quality data are synchronously acquired; meanwhile, a rendering cycle data feedback mechanism is constructed to realize dynamic updating of the parameters and the state. The method fuses the prediction and search strategies, realizes collaborative control of rendering efficiency and image quality, and has adaptive adjustment and cyclic optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of computer graphics processing technology, and in particular to a 3D engine rendering control method based on adaptive deep learning. Background Technology

[0002] With the widespread application of 3D graphics rendering technology in game engines, virtual reality, and digital twins, achieving a dynamic balance between rendering efficiency and image quality in complex scenes has become an important research direction for optimizing the performance of graphics processing systems. Currently, 3D engines typically rely on preset rendering parameters or dynamic adjustment methods based on empirical rules to control key parameters such as resolution, pixel sampling rate, and level of detail to meet rendering needs under different hardware conditions and scene complexities.

[0003] Existing rendering control methods still have significant limitations in complex dynamic scenes. On the one hand, traditional methods based on fixed strategies or simple threshold adjustments struggle to comprehensively consider the multidimensional impact of scene geometric complexity, texture complexity, and hardware resource status, leading to lag or inaccuracy in parameter adjustments and easily resulting in large frame latency fluctuations or image quality degradation. On the other hand, single optimization strategies typically lack the ability to effectively explore the rendering parameter search space, making it difficult to quickly converge to the optimal solution in discrete parameter combinations. This can easily lead to getting stuck in local optima or generating ineffective trials, reducing overall rendering efficiency.

[0004] Therefore, how to provide a 3D engine rendering control method based on adaptive deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a 3D engine rendering control method based on adaptive deep learning. This invention comprehensively utilizes deep learning prediction models, exploratory search algorithms, and direction optimization strategies to dynamically adjust rendering resolution, pixel sampling times, and detail levels. It also combines rendering performance and image quality data to construct a closed-loop feedback mechanism, thereby achieving continuous optimization and adaptive updating of rendering parameters.

[0006] A 3D engine rendering control method based on adaptive deep learning according to an embodiment of the present invention includes the following steps: S1. Collect scene data for the current rendering cycle and generate a rendering state feature vector; S2. Input the rendering state feature vector into the adaptive deep learning prediction network and output the initial rendering parameter set. S3. Use the initial set of rendering parameters as the initial point for Hooke-Jeeves exploration and search, and generate trial parameter points; S4. Input the trial parameter points into the 3D engine for trial rendering, obtain the frame latency data and image quality evaluation data corresponding to the trial parameter points, and determine the candidate parameter point set. S5. Use the set of candidate parameter points as input for the Rosenbrock direction search to determine the parameter search direction; S6. Perform iterative search on the candidate parameter point set along the parameter search direction to obtain the refined rendering parameter set; S7. Load the refined rendering parameter set into the 3D engine rendering process, generate the rendering image frame of the current rendering cycle, and obtain the frame latency data, hardware processor utilization data, video memory usage data and image quality evaluation data of the current rendering cycle. S8. Use the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle as input data for the next rendering cycle.

[0007] Optionally, S3 includes: S31. Read the initial rendering parameter set and form a discretized initial rendering parameter set; S32. Constructing the initial point construction process for the Hooke-Jeeves exploration search based on the discretized initial rendering parameter set. The initial point construction process includes a joint feasible region constraint correction step: Read the hardware processor utilization data, video memory usage data, scene geometry complexity data, and texture complexity data from the rendering state feature vector; The hardware processor budget coefficient is determined in the preset hardware processor budget mapping table based on the hardware processor utilization data, and the video memory budget coefficient is determined in the preset video memory budget mapping table based on the video memory usage data. The geometric budget coefficients are determined in the preset geometric budget mapping table based on the scene geometric complexity data, and the texture budget coefficients are determined in the preset texture budget mapping table based on the texture complexity data. According to a fixed combination rule, the hardware processor budget coefficient, video memory budget coefficient, geometric budget coefficient, and texture budget coefficient are written into the budget constraint vector, and the pixel sampling budget threshold and detail level budget threshold are calculated based on the budget constraint vector. The discretized initial rendering parameter set is substituted into the pixel sampling budget threshold and the level of detail budget threshold for consistency verification; if the consistency verification fails, the parameters are adjusted downwards in a fixed order: First, select the next lower value from the preset set of pixel sampling number parameter values ​​to replace the pixel sampling number parameter. Then, select the next lower value from the preset set of rendering resolution parameter values ​​to replace the rendering resolution parameter. Finally, select the next lower value from the preset set of detail level parameter values ​​to replace the detail level parameter. This process continues until the consistency check is successful. The initial point for the Hooke-Jeeves exploration search is formed by combining the rendering resolution parameters, pixel sampling count parameters, and detail level parameters after the consistency check is successful. S33. Constructing the exploration step size construction process for Hooke–Jeeves exploratory search: Calculate the fixed interval between two adjacent values ​​in the preset set of rendering resolution parameters, and define the fixed interval as the rendering resolution step size; Calculate the fixed interval between two adjacent values ​​in the preset set of pixel sampling number parameters, and define the fixed interval as the pixel sampling number step size; Calculate the fixed interval between two adjacent values ​​in the preset set of parameters at the level of detail, and define the fixed interval as the step size at the level of detail. Based on the scene geometric complexity data and texture complexity data, determine the step size level number in the preset step size level mapping table, and multiply the integer multiple coefficient corresponding to the step size level number by the rendering resolution step size, pixel sampling number step size and detail level step size respectively to obtain the initial exploration step size set; S34. Perform the exploration and movement steps of Hooke–Jeeves' exploration search: Using the initial point of the Hooke–Jeeves exploration search as the current reference point, single-dimensional exploration movement is performed sequentially according to the rendering resolution parameters, pixel sampling number parameters, and detail level parameters. For each parameter dimension, perform one forward exploration move and one reverse exploration move. The forward exploration move is to keep the other parameters unchanged at the current reference point and move the current parameter only in the direction of increasing its preset value set by one initial exploration step. The reverse exploration move is to keep the other parameters unchanged at the current reference point and move the current parameter only in the direction of decreasing its preset value set by one initial exploration step. The joint feasible region constraint correction step is performed on both the exploration movement parameter points obtained from each forward exploration movement and the exploration movement parameter points obtained from each reverse exploration movement. Write the parameter points output by the joint feasible region constraint correction step into the set of trial parameter points in the order of generation. S35. Perform evaluation calculations on the set of trial parameter points and determine the exploration result points; S36. After the exploration result points are determined, execute the asymmetric coupling trial step to generate exploration result points; S37. Execute the Hooke-Jeeves exploration search pattern movement step: Construct pattern movement parameter points using the current reference point and the exploration result point. The direction of change of each parameter level of the pattern movement parameter point is consistent with the direction of change of the parameter level from the current reference point to the exploration result point, and the magnitude of the change of the parameter level is equal to the magnitude of the change of the parameter level from the current reference point to the exploration result point. After processing the pattern movement parameter points through the joint feasible region constraint correction step, input them into the 3D engine for trial rendering, obtain the frame latency data and image quality evaluation data corresponding to the pattern movement parameter points, and calculate the pattern movement parameter evaluation value according to the preset evaluation function. When the pattern movement parameter evaluation value is less than the exploration result evaluation value, update the pattern movement parameter points to the new current reference point and return to the exploration movement step of the Hooke-Jeeves exploration search; otherwise, update the exploration result point to the new current reference point and return to the exploration movement step of the Hooke-Jeeves exploration search. S38. When no update point with a reduced evaluation value is obtained in a consecutive round, the initial exploration step size set is updated according to the fixed demotion rule: the step size level number is replaced with the next smaller step size level number, and the initial exploration step size set is rebuilt; when the step size level number reaches the preset minimum step size level number, all the generated trial parameter points are output as trial parameter points.

[0008] Optionally, the step of performing asymmetric coupling probing after the exploration result point is determined includes: When the exploration result point is adjusted up in terms of rendering resolution parameters relative to the current reference point, a lower value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, forming a coupled exploration parameter point; When the exploration result point is downgraded in terms of rendering resolution parameters relative to the current baseline point, a higher value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, thus forming a coupled exploration parameter point. After the coupling trial parameter points are processed by the joint feasible region constraint correction step, they are input into the 3D engine for trial rendering. The frame delay data and image quality evaluation data corresponding to the coupling trial parameter points are obtained, and the evaluation value of the coupling trial parameters is calculated according to the preset evaluation function. When the evaluation value of the coupling trial parameters is less than the evaluation value of the exploration result, the coupling trial parameter points are updated to exploration result points.

[0009] Optionally, S4 includes: S41. Read the set of trial parameter points and establish a trial rendering execution list in the order of generation. At the same time, write the camera pose matrix, light source parameter set, animation timestamp and visible object set of the current rendering cycle into the read-only buffer to form a fixed rendering input state. S42. Write each trial parameter point into the 3D engine control register unit in sequence according to the trial rendering execution list. Execute the complete rendering process and output the trial rendering image frame while keeping the fixed rendering input state unchanged. At the same time, calculate the corresponding frame delay data by recording the timestamp. S43. After outputting the trial rendered image frame, perform differential stability evaluation on the trial rendered image frame: Perform a pixel-by-pixel difference calculation between the trial rendered image frame and the baseline image frame cached at the beginning of the current rendering cycle; The percentage of pixels whose absolute pixel difference is greater than a preset difference threshold is used to obtain the image perturbation ratio. Cross-validate the image perturbation ratio with the pixel sampling number parameter. When the pixel sampling number parameter increases but the image perturbation ratio does not decrease, mark the trial parameter point as a structural invalid point. S44. Establish a tiered screening mechanism: The first filtering condition is that the frame delay data is less than or equal to the frame delay data corresponding to the current reference point; The second layer of filtering criteria is that the image perturbation ratio is less than or equal to the image perturbation ratio corresponding to the current reference point; Only trial parameter points that simultaneously meet the first-level and second-level screening conditions are included in the candidate parameter point set. S45. When the candidate parameter point set is empty, select the trial parameter point with the smallest frame delay data as the unique element of the candidate parameter point set; when the candidate parameter point set is not empty, output the candidate parameter point set in ascending order of frame delay data.

[0010] Optionally, S5 includes: S51. Based on the candidate parameter point set and the previous rendering cycle data, feasible candidate parameter points are selected according to the dual constraints of frame latency and image quality, and reference candidate parameter points are determined according to the priority rules. S52. Construct a reference parameter vector based on the reference candidate parameter points, and construct corresponding parameter vectors by selecting quality-priority candidate parameter points and delay-priority candidate parameter points respectively. S53. Based on the difference between the quality-priority parameter vector and the reference parameter vector, and the difference between the delay-priority parameter vector and the reference parameter vector, the first parameter search direction, the second parameter search direction, and the third parameter search direction are constructed by taking the sign, eliminating components, and enumerating directions. S54. Based on the preset value set of rendering resolution parameters, pixel sampling number parameters, and detail level parameters, determine the directional step size of each parameter search direction in the corresponding dimension and form a set of discrete step sizes. S55. Perform Rosenbrock direction search by successive linear attempts to form forward discrete trial parameter points and reverse discrete trial parameter points. S56. Input the forward discrete trial parameter points and the reverse discrete trial parameter points into the 3D engine for trial rendering and record the corresponding data. When the frame delay and image quality constraints are met, select the discrete trial parameter point with smaller frame delay to update the current parameter vector. Otherwise, keep the current parameter vector unchanged and process the next parameter search direction. S57. Execution Direction Reconstruction: Accumulate the parameter changes of all adopted and updated discrete trial parameter points relative to the reference parameter vector in the direction search evaluation record table to form a cumulative change vector. Sign the cumulative change vector in each of the three dimensions and limit it to [missing information]. Obtain the reconstructed principal direction; replace the reconstructed principal direction with the first parameter search direction, and sequentially perform component elimination processing on the second and third parameter search directions to remove non-zero components in the same direction as the determined direction; when the second or third parameter search direction becomes a zero vector, from In the middle, a direction that is different from the determined direction and is non-zero is selected and replaced according to a fixed enumeration order; S58. Execution direction acceptance lock: In the direction search evaluation record table, count the number of times the first parameter search direction, the second parameter search direction, and the third parameter search direction are used to update the current parameter vector. When the update count of any parameter search direction is zero, mark the parameter search direction as a locked elimination direction and do not construct a forward trial parameter vector and a reverse trial parameter vector for the parameter search direction in the next round of direction successive linear trial. S59. Output the first parameter search direction, the second parameter search direction, and the third parameter search direction as parameter search directions, and combine the reference parameter vector with the updated current parameter vector.

[0011] Optionally, the Rosenbrock direction search includes successive linear exploration of the direction, comprising: Using the reference parameter vector as the current parameter vector, process them one by one in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction; For each parameter search direction, construct a forward trial parameter vector and a reverse trial parameter vector. The forward trial parameter vector is the parameter vector obtained by moving the current parameter vector along the parameter search direction by one step size corresponding to the parameter search direction. The reverse trial parameter vector is the parameter vector obtained by moving the current parameter vector in the opposite direction along the parameter search direction by one step size corresponding to the parameter search direction. The forward and reverse trial parameter vectors are mapped to the nearest values ​​of the preset sets of rendering resolution parameters, pixel sampling times parameters, and detail level parameters, respectively, to form forward discrete trial parameter points and reverse discrete trial parameter points.

[0012] Optionally, S6 includes: S61. Read the reference parameter vector, the updated current parameter vector, and the first parameter search direction, the second parameter search direction, and the third parameter search direction. Establish an iterative control table that records the frame delay data and image quality evaluation data corresponding to the current parameter vector, and write the current parameter vector into the initial position of the refined rendering parameter set. S62. Perform a single-direction iterative search in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction, and set the corresponding direction search direction step size for each parameter search direction. S63. Starting from the current parameter vector, construct forward discrete iterative parameter points along the current parameter search direction and perform trial rendering. When the forward discrete iterative parameter points satisfy the conditions of reduced frame latency and no reduction in image quality evaluation data, update the current parameter vector and write it into the refined rendering parameter set. S64. When the forward discrete iteration parameter points do not meet the update conditions, construct the reverse discrete iteration parameter points and perform trial rendering and judgment. When the reverse discrete iteration parameter points meet the update conditions, update the current parameter vector and write it into the refined rendering parameter set. S65. When neither the forward discrete iteration parameter point nor the reverse discrete iteration parameter point satisfies the update condition, the step size of the current parameter search direction is replaced with the next smaller adjacent value interval and the iteration search is re-executed until no update occurs in the three directions, and then the refined rendering parameter set is output.

[0013] Optionally, S7 includes: S71. Read the refined rendering parameter set and select the current parameter vector obtained from the last update as the target rendering parameter vector for the current rendering cycle. Write each parameter into the corresponding control unit and complete parameter latching. S72. After completing parameter latching, start the complete rendering process of the 3D engine and calculate the frame latency data of the current rendering cycle by recording the rendering start timestamp and end timestamp. S73. During the execution of the complete rendering process of the 3D engine, the hardware processor utilization sampling value and the video memory usage sampling value are collected through the hardware monitoring interface and the arithmetic average is performed to obtain the hardware processor utilization data and video memory usage data of the current rendering cycle. S74. After the frame buffer output stage is completed, perform pixel-by-pixel difference calculation and edge detection operation on the rendered image frame of the current rendering cycle, and obtain the image quality evaluation data of the current rendering cycle based on the first quality component and the second quality component. S75. Establish a correspondence between the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle and the target rendering parameter vector, write it into the rendering cycle record table, and output the rendered image frame and its corresponding data of the current rendering cycle.

[0014] The beneficial effects of this invention are: This invention introduces Hooke–Jeeves exploratory search to explore the initial rendering parameters predicted by adaptive deep learning in multiple dimensions. By combining joint feasible region constraint correction, asymmetric coupling exploration, and mode shifting mechanism, it can quickly find effective parameter combinations that match the current scene state in the discrete parameter space, reduce the number of invalid explorations, and improve the efficiency of initial parameter correction and the coverage of parameter search.

[0015] This invention introduces Rosenbrock direction search, constructs parameter search directions based on frame delay data and image quality evaluation data corresponding to candidate parameter points, and combines direction successive linear trial, direction reconstruction and direction acceptance locking mechanisms to improve the directionality and specificity of the parameter iteration process, enhance the ability to approximate the optimal parameter region, reduce local oscillation and repeated search phenomena, thereby improving the convergence efficiency and stability of rendering parameter refinement optimization. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a 3D engine rendering control method based on adaptive deep learning proposed in this invention; Figure 2 This is a flowchart of the Hooke–Jeeves-based exploratory search proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 - Figure 2 A 3D engine rendering control method based on adaptive deep learning includes the following steps: S1. Collect scene geometric complexity data, texture complexity data, hardware processor utilization data, video memory usage data, frame latency data of the previous rendering cycle, and image quality evaluation data of the previous rendering cycle, and generate rendering state feature vectors. S2. Input the rendering state feature vector into the adaptive deep learning prediction network, and output an initial set of rendering parameters including rendering resolution parameters, pixel sampling number parameters, and detail level parameters. S3. Using the initial set of rendering parameters as the starting point for Hooke-Jeeves exploration and search, perform trials along the corresponding dimensions of rendering resolution parameters, pixel sampling times parameters, and detail level parameters to obtain trial parameter points; S4. Input the trial parameter points into the 3D engine for trial rendering, obtain the frame latency data and image quality evaluation data corresponding to the trial parameter points, and determine the candidate parameter point set. S5. Using the set of candidate parameter points as input for the Rosenbrock direction search, the parameter search direction is determined based on the frame latency data and image quality evaluation data corresponding to the set of candidate parameter points, combined with the frame latency data and image quality evaluation data of the previous rendering cycle. S6. Perform iterative search on the candidate parameter point set along the parameter search direction to obtain the refined rendering parameter set; S7. Load the refined rendering parameter set into the 3D engine rendering process, generate the rendering image frame of the current rendering cycle, and obtain the frame latency data, hardware processor utilization data, video memory usage data and image quality evaluation data of the current rendering cycle. S8. Use the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle as input data for the next rendering cycle.

[0019] In this embodiment, S1 includes: S11. At the start of the current rendering cycle, extract the set of mesh objects currently participating in rendering from the scene management structure of the 3D engine, traverse the set of mesh objects, count the number of triangles in each mesh object and sum them to obtain the total number of triangles, and at the same time count the number of mesh objects participating in rendering and the number of mesh objects that have switched levels of detail; divide the total number of triangles by a preset triangle capacity threshold, divide the number of mesh objects by a preset object number threshold, divide the number of mesh objects that have switched levels of detail by a preset switching number threshold, and perform a weighted sum of the three ratios to obtain the scene geometric complexity data; S12. Extract the currently loaded texture set from the texture management structure of the 3D engine, traverse the texture set, count the resolution of each texture and convert it into the number of pixels, and then sum them up to obtain the total number of texture pixels. At the same time, count the number of materials involved in rendering and the number of texture sampling calls in the current rendering cycle. Divide the total number of texture pixels by the preset pixel capacity threshold, the number of materials by the preset material number threshold, and the number of texture sampling calls by the preset sampling number threshold. Sum the three ratios with weights to obtain the texture complexity data. S13. Collect hardware processor utilization data and video memory usage data at a fixed sampling frequency through the hardware monitoring interface at multiple moments in the current rendering cycle. Perform arithmetic mean processing on each sample value and divide it by the corresponding rated maximum utilization and video memory capacity to obtain normalized hardware processor utilization data and normalized video memory usage data. S14. Read the frame latency data and image quality evaluation data of the previous rendering cycle from the historical rendering record cache, and perform time alignment processing according to the start time of the current rendering cycle. S15. According to the preset field order, the scene geometric complexity data, texture complexity data, normalized hardware processor utilization data, normalized video memory usage data, frame latency data of the previous rendering cycle, and image quality evaluation data of the previous rendering cycle are sequentially written into a fixed-length array structure to generate a rendering state feature vector.

[0020] In this embodiment, S2 includes: S21. Based on the field order and number of fields in the rendering state feature vector, construct an adaptive deep learning prediction network. The adaptive deep learning prediction network includes an input layer, a first fully connected layer, a second fully connected layer, and an output layer. The number of nodes in the input layer is the same as the number of fields in the rendering state feature vector. The output layer is set to three output nodes, which correspond to the rendering resolution parameter, the pixel sampling number parameter, and the detail level parameter, respectively. S22. Establish a historical sample buffer within the 3D engine process. Read the rendering state feature vector, the corresponding set of refined rendering parameters, the corresponding frame latency data, and the corresponding image quality evaluation data from the historical rendering record cache in the order of the rendering cycle. Use the rendering state feature vector as the training input of the adaptive deep learning prediction network, and use the rendering resolution parameter, pixel sampling number parameter, and detail level parameter in the set of refined rendering parameters as training labels. S23. Perform linear scaling on the rendering state feature vector in the historical sample buffer according to the preset minimum and preset maximum values ​​of each field to generate the scaled rendering state feature vector, and input the scaled rendering state feature vector into the adaptive deep learning prediction network. S24. Randomly initialize the network weight parameters of the adaptive deep learning prediction network, input the scaled rendering state feature vector in batches to obtain the predicted values ​​of the rendering resolution parameter, pixel sampling number parameter, and detail level parameter, and calculate the batch loss according to the following formula: ; in, For training labels, For predicted values, For frame delay data, For image quality assessment data, The preset frame delay threshold, To preset the image quality threshold, The network weight parameters are updated by backpropagation based on the batch loss, with preset weight coefficients. S25. In the online inference phase, the rendering state feature vector of the current rendering cycle is subjected to linear scaling processing consistent with S23. The scaled rendering state feature vector is input into the adaptive deep learning prediction network, and the predicted values ​​of the rendering resolution parameter, pixel sampling number parameter, and detail level parameter are output. S26. Select the value with the smallest difference from the predicted value of the rendering resolution parameter from the preset set of rendering resolution parameter values, select the value with the smallest difference from the predicted value of the pixel sampling number parameter from the preset set of pixel sampling number parameter values, and select the value with the smallest difference from the predicted value of the detail level parameter from the preset set of detail level parameter values ​​to generate the initial rendering parameter set. S27. When the obtained frame latency data of the current rendering cycle is greater than the preset frame latency threshold or the image quality evaluation data of the current rendering cycle is less than the preset image quality threshold, the rendering state feature vector, refined rendering parameter set, frame latency data and image quality evaluation data corresponding to the current rendering cycle are written into the historical sample buffer, and the network weight parameter update is performed again.

[0021] In this embodiment, S3 includes: S31. Read the initial rendering parameter set and obtain the rendering resolution parameter, pixel sampling number parameter, and level of detail parameter respectively; in the preset value set of the rendering resolution parameter, the preset value set of the pixel sampling number parameter, and the preset value set of the level of detail parameter, select the value with the smallest difference from the corresponding parameter value to form a discretized initial rendering parameter set. S32. Constructing the initial point construction process for the Hooke-Jeeves exploration search based on the discretized initial rendering parameter set. The initial point construction process includes a joint feasible region constraint correction step: Read the hardware processor utilization data, video memory usage data, scene geometry complexity data, and texture complexity data from the rendering state feature vector; The hardware processor budget coefficient is determined in the preset hardware processor budget mapping table based on the hardware processor utilization data, and the video memory budget coefficient is determined in the preset video memory budget mapping table based on the video memory usage data. The geometric budget coefficients are determined in the preset geometric budget mapping table based on the scene geometric complexity data, and the texture budget coefficients are determined in the preset texture budget mapping table based on the texture complexity data. According to a fixed combination rule, the hardware processor budget coefficient, video memory budget coefficient, geometric budget coefficient, and texture budget coefficient are written into the budget constraint vector, and the pixel sampling budget threshold and detail level budget threshold are calculated based on the budget constraint vector. The discretized initial rendering parameter set is substituted into the pixel sampling budget threshold and the level of detail budget threshold for consistency verification. When the consistency verification fails, the parameters are adjusted in a fixed order: first, a lower value is selected from the preset value set of the pixel sampling number parameter to replace the pixel sampling number parameter; then, a lower value is selected from the preset value set of the rendering resolution parameter to replace the rendering resolution parameter; finally, a lower value is selected from the preset value set of the level of detail parameter to replace the level of detail parameter, until the consistency verification is successful. The initial point for the Hooke-Jeeves exploration search is formed by combining the rendering resolution parameters, pixel sampling count parameters, and detail level parameters after the consistency check is successful. S33. Constructing the exploration step size construction process for Hooke–Jeeves exploratory search: Calculate a fixed interval between two adjacent values ​​in the preset set of rendering resolution parameters, and define the fixed interval as the rendering resolution step size; calculate a fixed interval between two adjacent values ​​in the preset set of pixel sampling times parameters, and define the fixed interval as the pixel sampling times step size; calculate a fixed interval between two adjacent values ​​in the preset set of detail level parameters, and define the fixed interval as the detail level step size. Based on the scene's geometric complexity data and texture complexity data, determine the step size level number in the preset step size level mapping table, and multiply the integer multiple coefficient corresponding to the step size level number by the rendering resolution step size, pixel sampling number step size, and detail level step size respectively to obtain the initial exploration step size set; S34. Execute the exploration movement steps of the Hooke-Jeeves exploration search: Using the initial point of the obtained Hooke-Jeeves exploration search as the current reference point, perform single-dimensional exploration movement in the order of rendering resolution parameters, pixel sampling count parameters, and detail level parameters; perform one forward exploration movement and one reverse exploration movement for each parameter dimension. The forward exploration movement is to keep the other parameters unchanged at the current reference point and only move the current parameter in the direction of increasing its preset value set by one initial exploration step. The reverse exploration movement is to keep the other parameters unchanged at the current reference point and only move the current parameter in the direction of decreasing its preset value set by one initial exploration step. Perform the joint feasible region constraint correction step described in S32 on the exploration movement parameter points obtained by each forward exploration movement and each reverse exploration movement. Write the parameter points output by the joint feasible region constraint correction step into the trial parameter point set in the order of generation. S35. Perform evaluation calculations on the set of trial parameter points and determine the exploration result points: Input the set of trial parameter points into the 3D engine point by point for trial rendering, and obtain the frame latency data and image quality evaluation data corresponding to each trial parameter point; calculate the evaluation value for each trial parameter point according to the preset evaluation function, and select the trial parameter point with the smallest evaluation value as the exploration result point; S36. After the exploration result points are determined, perform the asymmetric coupling probing step to avoid the conventional one-dimensional probing missing the discrete coupling combination of resolution parameters and pixel sampling number parameters: When the exploration result point is adjusted up in terms of rendering resolution parameters relative to the current reference point, a lower value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, forming a coupled exploration parameter point; When the exploration result point is downgraded in terms of rendering resolution parameters relative to the current baseline point, a higher value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, thus forming a coupled exploration parameter point. After the coupled trial parameter points are processed by the joint feasible region constraint correction step, they are input into the 3D engine for trial rendering. The frame delay data and image quality evaluation data corresponding to the coupled trial parameter points are obtained, and the evaluation value of the coupled trial parameters is calculated according to the preset evaluation function. When the evaluation value of the coupled trial parameters is less than the evaluation value of the exploration result, the coupled trial parameter points are updated to exploration result points. S37. Execute the Hooke-Jeeves exploration search pattern movement step: Construct pattern movement parameter points using the current reference point and the exploration result point. The direction of change of each parameter level of the pattern movement parameter point is consistent with the direction of change of the parameter level from the current reference point to the exploration result point, and the magnitude of the change of the parameter level is equal to the magnitude of the change of the parameter level from the current reference point to the exploration result point. After processing by the joint feasible region constraint correction step, the pattern movement parameter points are input into the 3D engine for trial rendering to obtain the frame delay data and image quality evaluation data corresponding to the pattern movement parameter points. The evaluation value of the pattern movement parameters is calculated according to the preset evaluation function. When the evaluation value of the pattern movement parameters is less than the evaluation value of the exploration result point, the pattern movement parameter points are updated to the new current reference point and the exploration movement step of the Hooke-Jeeves exploration search is returned. Otherwise, the exploration result point is updated to the new current reference point and the exploration movement step of the Hooke-Jeeves exploration search is returned. S38. When no update point with a reduced evaluation value is obtained in a consecutive round, the initial exploration step size set is updated according to the fixed demotion rule: the step size level number is replaced with the next smaller step size level number, and the initial exploration step size set is reconstructed according to S33. When the step size level number reaches the preset minimum step size level number, all trial parameter points are output as trial parameter points.

[0022] In this embodiment, S4 includes: S41. Read the set of trial parameter points, assign a sequence number to each trial parameter point according to the generation order, and establish a trial rendering execution list; before executing trial rendering, write the camera pose matrix, light source parameter set, animation timestamp and visible object set of the current rendering cycle into the read-only buffer to form a fixed rendering input state. S42. Process each trial parameter point one by one according to the order of the trial rendering execution list; while keeping the fixed rendering input state unchanged, write the rendering resolution parameter, pixel sampling number parameter and detail level parameter in the trial parameter point into the corresponding control register unit of the 3D engine rendering process, execute a complete rendering pipeline process, output the trial rendering image frame, and record the rendering start time stamp and rendering end time stamp, and calculate the frame delay data. S43. After outputting the trial rendered image frame, perform differential stability evaluation on the trial rendered image frame: Perform a pixel-by-pixel difference calculation between the trial rendered image frame and the baseline image frame cached at the beginning of the current rendering cycle; The percentage of pixels whose absolute pixel difference is greater than a preset difference threshold is used to obtain the image perturbation ratio. Cross-validate the image perturbation ratio with the pixel sampling number parameter. If the pixel sampling number parameter increases but the image perturbation ratio does not decrease, mark the probe parameter point as a structural invalid point. S44. Establish a tiered screening mechanism: The first filtering condition is that the frame delay data is less than or equal to the frame delay data corresponding to the current reference point; The second layer of filtering criteria is that the image perturbation ratio is less than or equal to the image perturbation ratio corresponding to the current reference point; Only trial parameter points that simultaneously meet the first-level and second-level screening conditions are included in the candidate parameter point set. S45. When the candidate parameter point set is empty, select the trial parameter point with the smallest frame delay data as the unique element of the candidate parameter point set; when the candidate parameter point set is not empty, output the candidate parameter point set in ascending order of frame delay data.

[0023] In this embodiment, S5 includes: S51. Read the candidate parameter point set and the frame delay data and image quality evaluation data corresponding to each candidate parameter point, and read the frame delay data and image quality evaluation data of the previous rendering cycle; select candidate parameter points from the candidate parameter point set that simultaneously satisfy "frame delay data is less than or equal to the frame delay data of the previous rendering cycle and image quality evaluation data is greater than or equal to the image quality evaluation data of the previous rendering cycle" to form a feasible candidate parameter point set; when the feasible candidate parameter point set is empty, select the candidate parameter point with the largest image quality evaluation data as the reference candidate parameter point, and select the candidate parameter point with the smallest frame delay data when the image quality evaluation data is the same; when the feasible candidate parameter point set is not empty, select the candidate parameter point with the smallest frame delay data as the reference candidate parameter point. S52. Write the rendering resolution parameter, pixel sampling number parameter, and detail level parameter of the reference candidate parameter point into the reference parameter vector in a fixed field order; select quality-priority candidate parameter points and latency-priority candidate parameter points that are different from the reference candidate parameter points from the candidate parameter point set. The quality-priority candidate parameter point is the candidate parameter point with the largest image quality evaluation data, and in the case of the same image quality evaluation data, the candidate parameter point with the smallest frame latency data is selected. The latency-priority candidate parameter point is the candidate point with the smallest frame latency data, and in the case of the same frame latency data, the candidate point with the largest image quality evaluation data is selected; write the quality-priority candidate parameter point into the quality-priority parameter vector, and write the latency-priority candidate point into the latency-priority parameter vector. S53. Construct a set of parameter search directions for the Rosenbrook directional search. This set includes a first parameter search direction, a second parameter search direction, and a third parameter search direction. The construction steps include: calculating the quality difference vector as the difference between the quality-priority parameter vector and the reference parameter vector; calculating the latency difference vector as the difference between the latency-priority parameter vector and the reference parameter vector; and resizing the quality difference vector in terms of the rendering resolution parameter, pixel sampling rate parameter, and level of detail parameter, and limiting it to... Obtain the first parameter, the search direction; take the sign of the time delay difference vector in each of the three dimensions and limit it to... A temporary second direction is obtained, and component elimination processing is performed on the temporary second direction to remove non-zero components in the same direction as the first parameter search direction, thus obtaining the second parameter search direction; The third parameter search direction is selected according to a fixed enumeration order, and is non-zero and is different from both the first and second parameter search directions. S54. Construct the set of discrete step sizes for Rosenbrook direction search: Read the interval between two adjacent values ​​from the preset set of rendering resolution parameters, the preset set of pixel sampling number parameters, and the preset set of detail level parameters, and determine the direction step size for rendering resolution parameters, pixel sampling number parameters, and detail level parameters, respectively; For the first parameter search direction, the second parameter search direction, and the third parameter search direction, take the direction step size corresponding to the non-zero components of the three dimensions as the parameter search direction step size for that direction; S55. Perform successive linear exploration of the Rosenbrock direction search: Using the reference parameter vector as the current parameter vector, process them one by one in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction; construct a forward exploration parameter vector and a reverse exploration parameter vector for each parameter search direction. The forward exploration parameter vector is the parameter vector obtained by moving the current parameter vector along the parameter search direction by one step of the corresponding parameter search direction, and the reverse exploration parameter vector is the parameter vector obtained by moving the current parameter vector in the opposite direction along the parameter search direction by one step of the corresponding parameter search direction; map the forward exploration parameter vector and the reverse exploration parameter vector to the nearest value of the preset value set of rendering resolution parameter, the preset value set of pixel sampling number parameter, and the preset value set of detail level parameter, respectively, to form forward discrete exploration parameter points and reverse discrete exploration parameter points; S56. Input the forward discrete trial parameter points and the reverse discrete trial parameter points into the 3D engine for trial rendering, obtain the corresponding frame delay data and image quality evaluation data, and write the current parameter vector, parameter search direction, forward discrete trial parameter points, reverse discrete trial parameter points, frame delay data, and image quality evaluation data into the direction search evaluation record table; when either the forward discrete trial parameter point or the reverse discrete trial parameter point satisfies the condition that "the frame delay data is less than or equal to the frame delay data corresponding to the current parameter vector and the image quality evaluation data is greater than or equal to the image quality evaluation data corresponding to the current parameter vector", select the discrete trial parameter point that meets the condition and has smaller frame delay data to update the new current parameter vector; when neither the forward discrete trial parameter point nor the reverse discrete trial parameter point satisfies the above conditions, keep the current parameter vector unchanged and process the next parameter search direction; S57. The direction reconstruction step is implemented as the first improved construction process: The parameter changes of all adopted and updated discrete trial parameter points relative to the reference parameter vector are accumulated in the direction search evaluation record table to form a cumulative change vector. The cumulative change vector is then signed in three dimensions and constrained to [missing information]. Obtain the reconstructed principal direction; replace the reconstructed principal direction with the first parameter search direction, and sequentially perform component elimination processing on the second and third parameter search directions to remove non-zero components in the same direction as the determined direction; when the second or third parameter search direction becomes a zero vector, from In the middle, a direction that is different from the determined direction and is non-zero is selected and replaced according to a fixed enumeration order; S58. The execution direction acceptance locking step is used as the second improved construction process: In the direction search evaluation record table, the number of times the first parameter search direction, the second parameter search direction, and the third parameter search direction are used to update the current parameter vector is counted. When the update count of any parameter search direction is zero, the parameter search direction is marked as a locked elimination direction and no longer constructs a positive trial parameter vector and a negative trial parameter vector in the next round of direction successive linear trial. S59. Output the first parameter search direction, the second parameter search direction, and the third parameter search direction as parameter search directions, and use the reference parameter vector and the updated current parameter vector as direction search inputs.

[0024] In this embodiment, S6 includes: S61. Read the reference parameter vector, the updated current parameter vector, and the first parameter search direction, the second parameter search direction, and the third parameter search direction. Establish an iteration control table, record the frame delay data and image quality evaluation data corresponding to the current parameter vector in the iteration control table, and write the current parameter vector into the initial position of the refined rendering parameter set. S62. Perform a single-direction iterative search in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction. Set a corresponding direction search direction step size for each parameter search direction. The direction search direction step size is the maximum value of the interval between two adjacent values ​​in the preset value set of rendering resolution parameter, the preset value set of pixel sampling number parameter, and the preset value set of detail level parameter. S63. Starting from the current parameter vector, construct a forward iterative parameter vector along the current parameter search direction. Map the forward iterative parameter vector to the nearest value in the corresponding preset value set in the three parameter dimensions to form a forward discrete iterative parameter point. Input the forward discrete iterative parameter point into the 3D engine for trial rendering to obtain the corresponding frame latency data and image quality evaluation data. When the forward discrete iterative parameter point satisfies the condition that "the frame latency data is less than the frame latency data corresponding to the current parameter vector and the image quality evaluation data is not less than the image quality evaluation data corresponding to the current parameter vector", update the forward discrete iterative parameter point to a new current parameter vector and write it into the refined rendering parameter set. S64. When the forward discrete iteration parameter point does not meet the update condition, construct the reverse iteration parameter vector and map it to the reverse discrete iteration parameter point, and perform the same trial rendering and judgment steps as the forward discrete iteration parameter point; when the reverse discrete iteration parameter point meets the update condition, update the reverse discrete iteration parameter point to the new current parameter vector and write it into the refined rendering parameter set. S65. When neither the forward discrete iteration parameter point nor the reverse discrete iteration parameter point satisfies the update condition, the step size of the current parameter search direction is replaced with the next smaller adjacent value interval, and S73 to S74 are re-executed after the step size replacement; when the first parameter search direction, the second parameter search direction and the third parameter search direction have all completed the above processing and the current parameter vector has not been updated in the three directions, the refined rendering parameter set is output as the refined rendering parameter set.

[0025] In this embodiment, S7 includes: S71. Read the refined rendering parameter set, select the current parameter vector obtained from the last update from the refined rendering parameter set as the target rendering parameter vector for the current rendering cycle; write the rendering resolution parameter in the target rendering parameter vector into the frame buffer size control unit, write the pixel sampling number parameter into the pixel sampling control unit, write the detail level parameter into the model detail selection control unit, and complete parameter latching. S72. After completing parameter latching, start the complete rendering process of the 3D engine, including the geometry processing stage, rasterization stage, pixel shading stage and frame buffer output stage; record the first timestamp at the beginning of the rendering process, record the second timestamp when the frame buffer output stage is completed, and calculate the difference between the second timestamp and the first timestamp as the frame latency data of the current rendering cycle. S73. During the execution of the complete rendering process of the 3D engine, the hardware processor utilization sampling value and the video memory usage sampling value are collected at a fixed sampling period through the hardware monitoring interface. The arithmetic average of the hardware processor utilization sampling value and the video memory usage sampling value are calculated to obtain the hardware processor utilization data and video memory usage data of the current rendering cycle. S74. After the frame buffer output stage is completed, perform image quality evaluation calculation on the rendered image frame of the current rendering cycle: calculate the pixel-by-pixel difference between the rendered image frame of the current rendering cycle and the rendered image frame of the previous rendering cycle, and count the proportion of pixels whose absolute difference value is less than a preset pixel difference threshold as the first quality component; perform edge detection operation on the rendered image frame of the current rendering cycle, and count the proportion of pixels whose edge response value is greater than a preset edge threshold as the second quality component; add the first quality component and the second quality component to obtain the image quality evaluation data of the current rendering cycle; S75. Establish a correspondence between the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle and the target rendering parameter vector, write it into the rendering cycle record table, and output the rendered image frame and its corresponding data of the current rendering cycle.

[0026] In this embodiment, S8 includes: S81. Read the frame latency data, hardware processor utilization data, video memory usage data and image quality evaluation data of the current rendering cycle, and read the target rendering parameter vector and the start timestamp of the current rendering cycle corresponding to the current rendering cycle. S82. Establish a rendering cycle state cache table. Write the current rendering cycle start timestamp, frame latency data, hardware processor utilization data, video memory usage data, image quality evaluation data, and target rendering parameter vector into the rendering cycle state cache table in chronological order to form a complete rendering cycle record entry. S83. Perform a sequential index update operation on the rendering cycle state cache table, mark the index of the current rendering cycle record entry as the latest index position, and mark the previous rendering cycle record entry as the previous rendering cycle record entry. S84. At the start of the next rendering cycle, read frame latency data and image quality evaluation data from the record entries of the previous rendering cycle in a fixed field order, and read hardware processor utilization data and video memory usage data from the record entries of the current rendering cycle, and write the above data into the input data buffer of the next rendering cycle. S85. The frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data in the input data buffer are used as the data source for generating the rendering state feature vector when S1 is executed in the next rendering cycle.

[0027] Example: To verify the proposed adaptive deep learning-based 3D engine rendering control method, this invention was deployed in a large-scale 3D visualization system. This system is used for real-time display of various types of 3D scenes, including building models, road models, vegetation models, water body models, and dynamic object models, with multiple layers of texture resources overlaid. During operation, the system needs to continuously switch between different viewpoints and display modes, resulting in continuous changes in the size of visible objects, the number of triangles, texture call frequency, and hardware resource consumption. Existing technologies, using fixed parameters or simple threshold adjustments, cannot simultaneously balance frame latency and image quality, easily leading to rendering stuttering and loss of detail.

[0028] In this scenario, the rendering state feature construction process is executed first. The scene management structure of the 3D engine obtains the set of mesh objects participating in the current rendering cycle. The number of triangles for each object in the set is counted and accumulated to obtain the total number of triangles. Simultaneously, the number of objects participating in rendering and the number of objects undergoing detail level switching are also counted to form scene geometric complexity data. The texture management structure obtains the currently loaded texture set. The number of pixels corresponding to the resolution of each texture is accumulated, and the number of texture sampling calls is counted to form texture complexity data. The hardware monitoring interface collects hardware processor utilization and video memory usage samples multiple times within the current rendering cycle, and performs arithmetic averaging on the samples to obtain hardware processor utilization and video memory usage data. Simultaneously, frame latency data and image quality evaluation data from previous rendering cycles are read from the cache structure. The above data are written into an array structure according to a preset field order to form a rendering state feature vector.

[0029] After the rendering state feature vector is generated, it is input into an adaptive deep learning prediction model. The prediction model outputs predicted values ​​for rendering resolution parameters, pixel sampling times parameters, and level of detail parameters, and maps these predicted values ​​to corresponding discrete levels in a preset value set to obtain an initial rendering parameter set. Then, a Hooke-Jeeves exploration search is performed on the initial rendering parameter set. Specifically, this involves reading the initial rendering parameter set and constructing a budget constraint vector by combining processor utilization data, memory usage data, scene geometric complexity data, and texture complexity data; performing joint feasible region constraint correction on the parameters to obtain the initial exploration point; constructing an exploration step size set based on the interval between adjacent levels in the preset value set of each parameter; and using the initial exploration point as a reference, performing forward and backward exploration in the three dimensions of rendering resolution parameters, pixel sampling times parameters, and level of detail parameters respectively to generate a set of trial parameter points.

[0030] After the set of trial parameter points is generated, a trial rendering execution list is established according to the generation order. Before trial rendering, the camera pose matrix, light source parameter set, animation timestamp, and visible object set for the current rendering cycle are written to a read-only buffer to form a fixed rendering input state. Then, the trial parameter points are written to the 3D engine control unit one by one, the complete rendering process is executed, and trial rendered image frames are output. Frame latency data is calculated by recording the rendering start and end timestamps. Simultaneously, image quality evaluation is performed on the trial rendered image frames, obtaining image quality evaluation data through pixel-by-pixel interpolation and edge detection. Based on the frame latency data and image quality evaluation data, the trial parameter points are filtered to obtain a candidate parameter point set.

[0031] A Rosenbroke direction search is performed based on the candidate parameter point set. First, candidate parameter points that satisfy frame latency and image quality constraints are selected to determine reference candidate parameter points. Then, quality-priority candidate parameter points and latency-priority candidate parameter points are determined separately. Based on these candidate parameter points, parameter vectors are constructed and difference vectors are calculated. Multiple parameter search directions are generated through sign and component processing. For each search direction, a discrete step size is constructed, and forward and reverse probing is performed. The current parameter vector is updated by comparing frame latency data with image quality evaluation data. Subsequently, a unidirectional iterative search is performed along each parameter search direction, continuously updating the current parameter vector. When no update is available, step size reduction is performed until a refined rendering parameter set is obtained.

[0032] After obtaining the refined rendering parameter set, the latest updated current parameter vector is selected as the target rendering parameter vector, and it is written into the frame buffer size control unit, pixel sampling control unit, and model detail selection control unit, respectively. The complete 3D engine rendering process is then executed to generate the rendered image frame for the current rendering cycle. Simultaneously, processor utilization and video memory usage data are collected, and image quality evaluation calculations are performed on the rendered image frame. A correspondence is established between frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data and the target rendering parameter vector, and this is written into the rendering cycle record table, serving as input data for calculations in the next rendering cycle.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A 3D engine rendering control method based on adaptive deep learning, characterized in that, include: S1. Collect scene data for the current rendering cycle and generate a rendering state feature vector; S2. Input the rendering state feature vector into the adaptive deep learning prediction network and output the initial rendering parameter set. S3. Use the initial set of rendering parameters as the initial point for Hooke-Jeeves exploration and search, and generate trial parameter points; S4. Input the trial parameter points into the 3D engine for trial rendering, obtain the frame latency data and image quality evaluation data corresponding to the trial parameter points, and determine the candidate parameter point set. S5. Use the set of candidate parameter points as input for the Rosenbrock direction search to determine the parameter search direction; S6. Perform iterative search on the candidate parameter point set along the parameter search direction to obtain the refined rendering parameter set; S7. Load the refined rendering parameter set into the 3D engine rendering process, generate the rendering image frame of the current rendering cycle, and obtain the frame latency data, hardware processor utilization data, video memory usage data and image quality evaluation data of the current rendering cycle. S8. Use the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle as input data for the next rendering cycle.

2. The 3D engine rendering control method based on adaptive deep learning according to claim 1, characterized in that, S3 includes: S31. Read the initial rendering parameter set and form a discretized initial rendering parameter set; S32. Constructing the initial point construction process for the Hooke-Jeeves exploration search based on the discretized initial rendering parameter set. The initial point construction process includes a joint feasible region constraint correction step: Read the hardware processor utilization data, video memory usage data, scene geometry complexity data, and texture complexity data from the rendering state feature vector; The hardware processor budget coefficient is determined in the preset hardware processor budget mapping table based on the hardware processor utilization data, and the video memory budget coefficient is determined in the preset video memory budget mapping table based on the video memory usage data. The geometric budget coefficients are determined in the preset geometric budget mapping table based on the scene geometric complexity data, and the texture budget coefficients are determined in the preset texture budget mapping table based on the texture complexity data. According to a fixed combination rule, the hardware processor budget coefficient, video memory budget coefficient, geometric budget coefficient, and texture budget coefficient are written into the budget constraint vector, and the pixel sampling budget threshold and detail level budget threshold are calculated based on the budget constraint vector. The discretized initial rendering parameter set is substituted into the pixel sampling budget threshold and the level of detail budget threshold for consistency verification; if the consistency verification fails, the parameters are adjusted downwards in a fixed order: First, select the next lower value from the preset set of pixel sampling number parameter values ​​to replace the pixel sampling number parameter. Then, select the next lower value from the preset set of rendering resolution parameter values ​​to replace the rendering resolution parameter. Finally, select the next lower value from the preset set of detail level parameter values ​​to replace the detail level parameter. This process continues until the consistency check is successful. The initial point for the Hooke-Jeeves exploration search is formed by combining the rendering resolution parameters, pixel sampling count parameters, and detail level parameters after the consistency check is successful. S33. Constructing the exploration step size construction process for Hooke–Jeeves exploratory search: Calculate the fixed interval between two adjacent values ​​in the preset set of rendering resolution parameters, and define the fixed interval as the rendering resolution step size; Calculate the fixed interval between two adjacent values ​​in the preset set of pixel sampling number parameters, and define the fixed interval as the pixel sampling number step size; Calculate the fixed interval between two adjacent values ​​in the preset set of parameters at the level of detail, and define the fixed interval as the step size at the level of detail. Based on the scene geometric complexity data and texture complexity data, determine the step size level number in the preset step size level mapping table, and multiply the integer multiple coefficient corresponding to the step size level number by the rendering resolution step size, pixel sampling number step size and detail level step size respectively to obtain the initial exploration step size set; S34. Perform the exploration and movement steps of Hooke–Jeeves' exploration search: Using the initial point of the Hooke–Jeeves exploration search as the current reference point, single-dimensional exploration movement is performed sequentially according to the rendering resolution parameters, pixel sampling number parameters, and detail level parameters. For each parameter dimension, perform one forward exploration move and one reverse exploration move. The forward exploration move is to keep the other parameters unchanged at the current reference point and move the current parameter only in the direction of increasing its preset value set by one initial exploration step. The reverse exploration move is to keep the other parameters unchanged at the current reference point and move the current parameter only in the direction of decreasing its preset value set by one initial exploration step. The joint feasible region constraint correction step is performed on both the exploration movement parameter points obtained from each forward exploration movement and the exploration movement parameter points obtained from each reverse exploration movement. Write the parameter points output by the joint feasible region constraint correction step into the set of trial parameter points in the order of generation. S35. Perform evaluation calculations on the set of trial parameter points and determine the exploration result points; S36. After the exploration result points are determined, execute the asymmetric coupling trial step to generate exploration result points; S37. Execute the Hooke-Jeeves exploration search pattern movement step: Construct pattern movement parameter points using the current reference point and exploration result point. The direction of change of each parameter level of the pattern movement parameter point is consistent with the direction of change of the parameter level from the current reference point to the exploration result point, and the magnitude of the change of the parameter level is equal to the magnitude of the change of the parameter level from the current reference point to the exploration result point. After processing the pattern movement parameter points through the joint feasible region constraint correction step, input them into the 3D engine for trial rendering to obtain the frame latency data and image quality evaluation data corresponding to the pattern movement parameter points, and calculate the pattern movement parameter evaluation value according to the preset evaluation function. When the pattern movement parameter evaluation value is less than the exploration result evaluation value, update the pattern movement parameter point to a new current reference point and return to the exploration movement step of the Hooke-Jeeves exploration search; otherwise, update the exploration result point to a new current reference point and return to the exploration movement step of the Hooke-Jeeves exploration search. S38. When no update point with a reduced evaluation value is obtained in a consecutive round, the initial exploration step size set is updated according to the fixed demotion rule: the step size level number is replaced with the next smaller step size level number, and the initial exploration step size set is rebuilt; when the step size level number reaches the preset minimum step size level number, all the generated trial parameter points are output as trial parameter points.

3. The 3D engine rendering control method based on adaptive deep learning according to claim 2, characterized in that, The step of performing asymmetric coupling probing after the exploration result point is determined includes: When the exploration result point is adjusted up in terms of rendering resolution parameters relative to the current reference point, a lower value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, forming a coupled exploration parameter point; When the exploration result point is downgraded in terms of rendering resolution parameters relative to the current baseline point, a higher value is selected from the preset set of pixel sampling number parameters to replace the pixel sampling number parameters, thus forming a coupled exploration parameter point. After the coupling trial parameter points are processed by the joint feasible region constraint correction step, they are input into the 3D engine for trial rendering. The frame delay data and image quality evaluation data corresponding to the coupling trial parameter points are obtained, and the evaluation value of the coupling trial parameters is calculated according to the preset evaluation function. When the evaluation value of the coupling trial parameters is less than the evaluation value of the exploration result, the coupling trial parameter points are updated to exploration result points.

4. The 3D engine rendering control method based on adaptive deep learning according to claim 1, characterized in that, S4 includes: S41. Read the set of trial parameter points and establish a trial rendering execution list in the order of generation. At the same time, write the camera pose matrix, light source parameter set, animation timestamp and visible object set of the current rendering cycle into the read-only buffer to form a fixed rendering input state. S42. Write each trial parameter point into the 3D engine control register unit in sequence according to the trial rendering execution list. Execute the complete rendering process and output the trial rendering image frame while keeping the fixed rendering input state unchanged. At the same time, calculate the corresponding frame delay data by recording the timestamp. S43. After outputting the trial rendered image frame, perform differential stability evaluation on the trial rendered image frame: Perform a pixel-by-pixel difference calculation between the trial rendered image frame and the baseline image frame cached at the beginning of the current rendering cycle; The percentage of pixels whose absolute pixel difference is greater than a preset difference threshold is used to obtain the image perturbation ratio. Cross-validate the image perturbation ratio with the pixel sampling number parameter. When the pixel sampling number parameter increases but the image perturbation ratio does not decrease, mark the trial parameter point as a structural invalid point. S44. Establish a tiered screening mechanism: The first filtering condition is that the frame delay data is less than or equal to the frame delay data corresponding to the current reference point; The second layer of filtering criteria is that the image perturbation ratio is less than or equal to the image perturbation ratio corresponding to the current reference point; Only trial parameter points that simultaneously meet the first-level and second-level screening conditions are included in the candidate parameter point set. S45. When the candidate parameter point set is empty, select the trial parameter point with the smallest frame delay data as the unique element of the candidate parameter point set; when the candidate parameter point set is not empty, output the candidate parameter point set in ascending order of frame delay data.

5. The 3D engine rendering control method based on adaptive deep learning according to claim 1, characterized in that, S5 includes: S51. Based on the candidate parameter point set and the previous rendering cycle data, feasible candidate parameter points are selected according to the dual constraints of frame latency and image quality, and reference candidate parameter points are determined according to the priority rules. S52. Construct a reference parameter vector based on the reference candidate parameter points, and construct corresponding parameter vectors by selecting quality-priority candidate parameter points and delay-priority candidate parameter points respectively. S53. Based on the difference between the quality-priority parameter vector and the reference parameter vector, and the difference between the delay-priority parameter vector and the reference parameter vector, the first parameter search direction, the second parameter search direction, and the third parameter search direction are constructed by taking the sign, eliminating components, and enumerating directions. S54. Based on the preset value set of rendering resolution parameters, pixel sampling number parameters, and detail level parameters, determine the directional step size of each parameter search direction in the corresponding dimension and form a set of discrete step sizes. S55. Perform Rosenbrock direction search by successive linear attempts to form forward discrete trial parameter points and reverse discrete trial parameter points. S56. Input the forward discrete trial parameter points and the reverse discrete trial parameter points into the 3D engine for trial rendering and record the corresponding data. When the frame delay and image quality constraints are met, select the discrete trial parameter point with smaller frame delay to update the current parameter vector. Otherwise, keep the current parameter vector unchanged and process the next parameter search direction. S57. Execution Direction Reconstruction: Accumulate the parameter changes of all adopted and updated discrete trial parameter points relative to the reference parameter vector in the direction search evaluation record table to form a cumulative change vector. Sign the cumulative change vector in each of the three dimensions and limit it to [missing information]. Obtain the reconstructed principal direction; replace the reconstructed principal direction with the first parameter search direction, and sequentially perform component elimination processing on the second and third parameter search directions to remove non-zero components in the same direction as the determined direction; when the second or third parameter search direction becomes a zero vector, from In the middle, a direction that is different from the determined direction and is non-zero is selected and replaced according to a fixed enumeration order; S58. Execution direction acceptance lock: In the direction search evaluation record table, count the number of times the first parameter search direction, the second parameter search direction, and the third parameter search direction are used to update the current parameter vector. When the update count of any parameter search direction is zero, mark the parameter search direction as a locked elimination direction and do not construct a forward trial parameter vector and a reverse trial parameter vector for the parameter search direction in the next round of direction successive linear trial. S59. Output the first parameter search direction, the second parameter search direction, and the third parameter search direction as parameter search directions, and combine the reference parameter vector with the updated current parameter vector.

6. The 3D engine rendering control method based on adaptive deep learning according to claim 5, characterized in that, The Rosenbrock direction search includes successive linear exploration of the direction, including: Using the reference parameter vector as the current parameter vector, process them one by one in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction; For each parameter search direction, construct a forward trial parameter vector and a reverse trial parameter vector. The forward trial parameter vector is the parameter vector obtained by moving the current parameter vector along the parameter search direction by one step size corresponding to the parameter search direction. The reverse trial parameter vector is the parameter vector obtained by moving the current parameter vector in the opposite direction along the parameter search direction by one step size corresponding to the parameter search direction. The forward and reverse trial parameter vectors are mapped to the nearest values ​​of the preset sets of rendering resolution parameters, pixel sampling times parameters, and detail level parameters, respectively, to form forward discrete trial parameter points and reverse discrete trial parameter points.

7. The 3D engine rendering control method based on adaptive deep learning according to claim 1, characterized in that, S6 includes: S61. Read the reference parameter vector, the updated current parameter vector, and the first parameter search direction, the second parameter search direction, and the third parameter search direction. Establish an iterative control table that records the frame delay data and image quality evaluation data corresponding to the current parameter vector, and write the current parameter vector into the initial position of the refined rendering parameter set. S62. Perform a single-direction iterative search in the order of the first parameter search direction, the second parameter search direction, and the third parameter search direction, and set the corresponding direction search direction step size for each parameter search direction. S63. Starting from the current parameter vector, construct forward discrete iterative parameter points along the current parameter search direction and perform trial rendering. When the forward discrete iterative parameter points satisfy the conditions of reduced frame latency and no reduction in image quality evaluation data, update the current parameter vector and write it into the refined rendering parameter set. S64. When the forward discrete iteration parameter points do not meet the update conditions, construct the reverse discrete iteration parameter points and perform trial rendering and judgment. When the reverse discrete iteration parameter points meet the update conditions, update the current parameter vector and write it into the refined rendering parameter set. S65. When neither the forward discrete iteration parameter point nor the reverse discrete iteration parameter point satisfies the update condition, the step size of the current parameter search direction is replaced with the next smaller adjacent value interval and the iteration search is re-executed until no update occurs in the three directions, and then the refined rendering parameter set is output.

8. The 3D engine rendering control method based on adaptive deep learning according to claim 1, characterized in that, S7 includes: S71. Read the refined rendering parameter set and select the current parameter vector obtained from the last update as the target rendering parameter vector for the current rendering cycle. Write each parameter into the corresponding control unit and complete parameter latching. S72. After completing parameter latching, start the complete rendering process of the 3D engine and calculate the frame latency data of the current rendering cycle by recording the rendering start timestamp and end timestamp. S73. During the execution of the complete rendering process of the 3D engine, the hardware processor utilization sampling value and the video memory usage sampling value are collected through the hardware monitoring interface and the arithmetic average is performed to obtain the hardware processor utilization data and video memory usage data of the current rendering cycle. S74. After the frame buffer output stage is completed, perform pixel-by-pixel difference calculation and edge detection operation on the rendered image frame of the current rendering cycle, and obtain the image quality evaluation data of the current rendering cycle based on the first quality component and the second quality component. S75. Establish a correspondence between the frame latency data, hardware processor utilization data, video memory usage data, and image quality evaluation data of the current rendering cycle and the target rendering parameter vector, write it into the rendering cycle record table, and output the rendered image frame and its corresponding data of the current rendering cycle.

Images