An assisted driving data decomposition method and system for a reconfigurable tpu computing device

Abstract

The application discloses an assisted driving data decomposition method and system of a reconfigurable TPU computing device, and belongs to the technical field of electric digital data processing. The method comprises the following steps: determining preliminary decomposition parameters according to data attribute information of training assisted driving data and operator combination information of an assisted driving module deployed on the reconfigurable TPU computing device; decomposing the training assisted driving data according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and dynamically reconfiguring the reconfigurable TPU computing device based on the preliminary decomposition parameters; inputting the preliminary decomposition result into the assisted driving module, and collecting resource state information of the reconfigurable TPU computing device; and iteratively optimizing the preliminary decomposition parameters according to the resource state information to obtain target decomposition parameters. According to the technical scheme, adaptive and efficient data decomposition that is accurately matched with data characteristics and reasoning requirements is realized, power optimization and resource dynamic adaptation are achieved, and reasoning efficiency and accuracy are significantly improved.

Description

Technical Field

[0001] This application belongs to the field of electronic digital data processing technology, specifically relating to a method and system for decomposing assisted driving data of a reconfigurable TPU computing device. Background Technology

[0002] Assisted driving technology involves acquiring data from onboard sensors, processing that data for inference, and then outputting corresponding control commands based on the inference results. This enables precise control of the vehicle's driving status (such as adaptive cruise control and emergency obstacle avoidance). The safety, real-time performance, and decision-making accuracy of assisted driving technology highly depend on the efficiency, stability, and adaptability of data inference.

[0003] Data processing is a crucial prerequisite before data inference. To improve computational efficiency, existing technologies typically preprocess and decompose the input assisted driving data, such as segmenting high-resolution images into fixed-size tiles. However, this static and general data decomposition strategy is difficult to adapt to the complex and ever-changing driving environments and data characteristics in assisted driving scenarios, and may instead lead to a series of efficiency losses and accuracy deviations. Summary of the Invention

[0004] This application provides a method and system for decomposing assisted driving data using a reconfigurable TPU computing device. The purpose is to achieve adaptive and efficient data decomposition that is precisely matched with data characteristics and inference requirements by combining a dynamic decomposition strategy driven by data features and computing operators with the hardware dynamic reconfiguration capability of the reconfigurable TPU. At the same time, a flexible hardware architecture that can cooperate with the decomposition parameters in real time is constructed to achieve computing power optimization and dynamic resource adaptation, thereby significantly improving inference efficiency and accuracy.

[0005] In a first aspect, this application provides a method for decomposing assisted driving data using a reconfigurable TPU computing device, the method comprising:

[0006] Acquire training assisted driving data, and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio.

[0007] The training assisted driving data is decomposed according to the preliminary decomposition parameters to obtain preliminary decomposition results, and the reconfigurable TPU computing device is dynamically reconstructed based on the preliminary decomposition parameters.

[0008] The preliminary decomposition results are input into the driver assistance module, and the resource status information of the reconfigurable TPU computing device is collected.

[0009] The initial decomposition parameters are optimized based on the resource status information, and the reconfigurable TPU computing device is dynamically reconfigured until the initial decomposition parameters meet the preset optimization stop condition to obtain the target decomposition parameters, which are used to decompose actual assisted driving data according to the target decomposition parameters.

[0010] Optionally, the step of decomposing the training assisted driving data according to the preliminary decomposition parameters to obtain the preliminary decomposition result includes:

[0011] The decomposition movement step and target decomposition size are determined based on the decomposition size and the overlap ratio, and the training assisted driving data is divided according to the decomposition movement step and the target decomposition size to obtain basic data blocks.

[0012] The basic data block is reconstructed and encoded according to the stated precision compression ratio to obtain a compressed data block; wherein, the reconstruction encoding includes quantization encoding and sparsification encoding;

[0013] Accordingly, inputting the preliminary decomposition results into the driver assistance module includes:

[0014] The compressed data block is input into the driver assistance module.

[0015] Optionally, after reconstructing and encoding the basic data block according to the precision compression ratio to obtain a compressed data block, the method further includes:

[0016] The computational dependency information between the compressed data blocks is determined based on the operator combination information;

[0017] Accordingly, the dynamic reconfiguration of the reconfigurable TPU computing device based on the preliminary decomposition parameters includes:

[0018] Based on the attribute information of the compressed data block, configure the data bit width, computing mode, and local memory capacity of the processing unit in the reconfigurable TPU computing device;

[0019] Based on the computational dependency information, compressed data blocks are allocated to the processing unit to obtain the mapping relationship between the processing unit and the compressed data blocks;

[0020] Based on the computational dependency information and the mapping relationship, the data communication requirement information between the processing units is determined, and the connection topology and buffer depth of the on-chip interconnect network between the processing units in the reconfigurable TPU computing device are configured based on the data communication requirement information.

[0021] Optionally, after determining the data communication requirement information between the processing units based on the computational dependency information and the mapping relationship, the method further includes:

[0022] Parallel data stream orchestration information is generated based on the mapping relationship and the computational dependency information;

[0023] The synchronous processing time of the processing unit is generated based on the parallel data stream orchestration information, and is used to control the processing unit to process the corresponding compressed data block according to the synchronous processing time during model training, so as to realize the synchronous processing of the compressed data block.

[0024] Optionally, determining the preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device includes:

[0025] Based on the modality type information of the training assisted driving data and the operator computational complexity in the operator combination information, the baseline decomposition size, the baseline overlap ratio, and the baseline precision compression ratio are allocated;

[0026] Based on the perception priority information and the continuous rate of change information of the training assisted driving data, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the accuracy compression ratio adjustment coefficient are determined.

[0027] Multiply the baseline decomposition size and the decomposition size adjustment factor, multiply the baseline overlap ratio and the overlap ratio adjustment factor, and multiply the baseline precision compression ratio and the precision compression ratio adjustment factor to obtain the basic decomposition parameters;

[0028] The estimated performance information is calculated based on the basic decomposition parameters and the operator combination information, and the basic decomposition parameters are iteratively optimized based on the estimated performance information to obtain the preliminary decomposition parameters.

[0029] Optionally, the estimated performance information includes estimated throughput and estimated communication overhead;

[0030] Accordingly, the iterative optimization of the basic decomposition parameters based on the estimated performance information to obtain preliminary decomposition parameters includes:

[0031] Substituting the estimated throughput and the estimated communication overhead into the joint optimization objective function yields the optimization objective difference; wherein the joint optimization objective function aims to maximize the estimated throughput and minimize the estimated communication overhead.

[0032] Calculate the gradient of the optimization target difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters, and adjust the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradient.

[0033] Repeat the above steps until the rate of change of the optimization target difference is less than the preset rate of change threshold, and obtain the preliminary decomposition parameters.

[0034] Optionally, the resource status information includes the computing utilization rate, bandwidth utilization rate, and local memory occupancy rate of the processing units in the reconfigurable TPU computing device.

[0035] Accordingly, optimizing the preliminary decomposition parameters based on the resource status information includes:

[0036] If the computational utilization rate exceeds a preset computational utilization rate threshold, the decomposition size optimization coefficient is calculated based on the computational utilization rate and the preset computational utilization rate threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain the optimized decomposition size;

[0037] When the bandwidth utilization exceeds a preset bandwidth utilization threshold, an overlap ratio optimization coefficient is calculated based on the bandwidth utilization and the preset bandwidth utilization threshold, and the overlap ratio optimization coefficient is multiplied by the overlap ratio to obtain the optimized overlap ratio.

[0038] If the local memory occupancy rate exceeds a preset occupancy rate threshold, a precision compression ratio optimization coefficient is calculated based on the local memory occupancy rate and the preset occupancy rate threshold, and the precision compression ratio optimization coefficient is multiplied by the precision compression ratio to obtain the optimized precision compression ratio.

[0039] Secondly, this application provides an assisted driving data decomposition system for a reconfigurable TPU computing device, the system comprising:

[0040] The preliminary scheme determination module is used to acquire training assisted driving data and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio.

[0041] The TPU reconstruction module is used to decompose the training assisted driving data according to the preliminary decomposition parameters to obtain preliminary decomposition results, and to dynamically reconstruct the reconfigurable TPU computing device based on the preliminary decomposition parameters.

[0042] The decomposition result input module is used to input the preliminary decomposition results to the assisted driving module and collect the resource status information of the reconfigurable TPU computing device;

[0043] The decomposition parameter optimization module is used to optimize the preliminary decomposition parameters based on the resource status information and dynamically reconfigure the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop conditions to obtain the target decomposition parameters, which are used to decompose the actual assisted driving data according to the target decomposition parameters.

[0044] Optionally, the TPU reconfiguration module is specifically used for:

[0045] The decomposition movement step and target decomposition size are determined based on the decomposition size and the overlap ratio, and the training assisted driving data is divided according to the decomposition movement step and the target decomposition size to obtain basic data blocks.

[0046] The basic data block is reconstructed and encoded according to the stated precision compression ratio to obtain a compressed data block; wherein, the reconstruction encoding includes quantization encoding and sparsification encoding;

[0047] Accordingly, the decomposition result input module is specifically used for:

[0048] The compressed data block is input into the driver assistance module.

[0049] Optionally, the TPU reconfiguration module is further configured to:

[0050] The computational dependency information between the compressed data blocks is determined based on the operator combination information;

[0051] Based on the attribute information of the compressed data block, configure the data bit width, computing mode, and local memory capacity of the processing unit in the reconfigurable TPU computing device;

[0052] Based on the computational dependency information, compressed data blocks are allocated to the processing unit to obtain the mapping relationship between the processing unit and the compressed data blocks;

[0053] Based on the computational dependency information and the mapping relationship, the data communication requirement information between the processing units is determined, and the connection topology and buffer depth of the on-chip interconnect network between the processing units in the reconfigurable TPU computing device are configured based on the data communication requirement information.

[0054] Optionally, the TPU reconfiguration module is further configured to:

[0055] Parallel data stream orchestration information is generated based on the mapping relationship and the computational dependency information;

[0056] The synchronous processing time of the processing unit is generated based on the parallel data stream orchestration information, and is used to control the processing unit to process the corresponding compressed data block according to the synchronous processing time during model training, so as to realize the synchronous processing of the compressed data block.

[0057] Optionally, the preliminary scheme determination module is specifically used for:

[0058] Based on the modality type information of the training assisted driving data and the operator computational complexity in the operator combination information, the baseline decomposition size, the baseline overlap ratio, and the baseline precision compression ratio are allocated;

[0059] Based on the perception priority information and the continuous rate of change information of the training assisted driving data, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the accuracy compression ratio adjustment coefficient are determined.

[0060] Multiply the baseline decomposition size and the decomposition size adjustment factor, multiply the baseline overlap ratio and the overlap ratio adjustment factor, and multiply the baseline precision compression ratio and the precision compression ratio adjustment factor to obtain the basic decomposition parameters;

[0061] The estimated performance information is calculated based on the basic decomposition parameters and the operator combination information, and the basic decomposition parameters are iteratively optimized based on the estimated performance information to obtain the preliminary decomposition parameters.

[0062] Optionally, the estimated performance information includes estimated throughput and estimated communication overhead;

[0063] Accordingly, the preliminary scheme determination module is specifically used for:

[0064] Substituting the estimated throughput and the estimated communication overhead into the joint optimization objective function yields the optimization objective difference; wherein the joint optimization objective function aims to maximize the estimated throughput and minimize the estimated communication overhead.

[0065] Calculate the gradient of the optimization target difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters, and adjust the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradient.

[0066] Repeat the above steps until the rate of change of the optimization target difference is less than the preset rate of change threshold, and obtain the preliminary decomposition parameters.

[0067] Optionally, the resource status information includes the computing utilization rate, bandwidth utilization rate, and local memory occupancy rate of the processing units in the reconfigurable TPU computing device.

[0068] Accordingly, the decomposition parameter optimization module is specifically used for:

[0069] If the computational utilization rate exceeds a preset computational utilization rate threshold, the decomposition size optimization coefficient is calculated based on the computational utilization rate and the preset computational utilization rate threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain the optimized decomposition size;

[0070] When the bandwidth utilization exceeds a preset bandwidth utilization threshold, an overlap ratio optimization coefficient is calculated based on the bandwidth utilization and the preset bandwidth utilization threshold, and the overlap ratio optimization coefficient is multiplied by the overlap ratio to obtain the optimized overlap ratio.

[0071] If the local memory occupancy rate exceeds a preset occupancy rate threshold, a precision compression ratio optimization coefficient is calculated based on the local memory occupancy rate and the preset occupancy rate threshold, and the precision compression ratio optimization coefficient is multiplied by the precision compression ratio to obtain the optimized precision compression ratio.

[0072] Thirdly, this application provides an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method described in the first aspect.

[0073] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the method described in the first aspect.

[0074] In this application, training assisted driving data is acquired, and preliminary decomposition parameters are determined based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on a reconfigurable TPU computing device. The data attribute information includes modality type information, perception priority information, and continuous rate of change information; the preliminary decomposition parameters include decomposition size, overlap ratio, and precision compression ratio. The training assisted driving data is decomposed according to the preliminary decomposition parameters to obtain preliminary decomposition results, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters. The preliminary decomposition results are input to the assisted driving module, and the resource status information of the reconfigurable TPU computing device is collected. The preliminary decomposition parameters are optimized based on the resource status information, and the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet a preset optimization stop condition, resulting in target decomposition parameters, which are used to decompose actual assisted driving data according to the target decomposition parameters. The aforementioned assisted driving data decomposition method using a reconfigurable TPU computing device is based on a dynamic decomposition strategy driven by data features and computing operators. Combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, it achieves adaptive and efficient data decomposition that precisely matches data characteristics and inference requirements. Simultaneously, it constructs a flexible hardware architecture that can work in real time with the decomposition parameters, achieving computing power optimization and dynamic resource adaptation, thereby significantly improving inference efficiency and accuracy. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating a method for decomposing assisted driving data using a reconfigurable TPU computing device, as provided in this application.

[0076] Figure 2 This is a flowchart illustrating another method for decomposing assisted driving data using a reconfigurable TPU computing device provided in this application;

[0077] Figure 3 This is an example diagram of the partitioning of image data provided in this application;

[0078] Figure 4 This is a flowchart illustrating another method for decomposing assisted driving data using a reconfigurable TPU computing device provided in this application;

[0079] Figure 5 This is a schematic diagram of the structure of an assisted driving data decomposition system for a reconfigurable TPU computing device provided in this application;

[0080] Figure 6 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0082] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0083] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0084] The following description, in conjunction with the accompanying drawings, details the assisted driving data decomposition method and system for the reconfigurable TPU computing device provided in this application, through specific embodiments and application scenarios.

[0085] First, this application applies to scenarios where, before vehicles with assisted driving functions are actually put into use, the data decomposition strategy used during assisted driving inference and the corresponding TPU hardware configuration parameters are pre-trained and optimized. Based on the above application scenario, it is understood that the executing entity of this application can be a terminal device with computing capabilities, and no specific limitation is made here.

[0086] Figure 1This is a flowchart illustrating a method for decomposing assisted driving data using a reconfigurable TPU computing device, as provided in this application. Figure 1 As shown, the specific steps include the following:

[0087] S101, acquire training assisted driving data, and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio.

[0088] The training data for assisted driving can be environmental perception data used to train the assisted driving module, and may include image data and at least one type of radar data. The assisted driving module can be a neural network model that realizes assisted driving inference functions. Specifically, the image data can be color images, infrared images, or grayscale images captured by the vehicle-mounted camera, containing visual information of environmental elements such as lane lines, traffic lights, pedestrians, non-motorized vehicles, other vehicles, and road signs. The radar data can be ranging and angle measurement data collected by the vehicle-mounted radar sensor to characterize target distance, speed, and azimuth, and may include millimeter-wave radar data, lidar data, and ultrasonic radar data.

[0089] In one embodiment, the method for acquiring training data for assisted driving can be as follows: collecting real-world environmental perception data through real-vehicle road tests in real-world scenarios; selecting target environmental perception data from publicly available datasets; or generating simulated environmental perception data through a simulation platform. Specifically, collecting real-world environmental perception data through real-vehicle road tests involves using onboard sensors such as image sensors, millimeter-wave radar sensors, lidar sensors, and ultrasonic radar sensors to collect real-world environmental perception data from the vehicle itself.

[0090] The data attribute information of the training assisted driving data can be a set of core parameters used to characterize the data features of the training assisted driving data, which may include modality type information, perception priority information, and continuous rate of change information. Specifically, modality type information can be the sensor modality category to which the training assisted driving data belongs, such as image modality, millimeter-wave radar modality, lidar modality, ultrasonic radar modality, etc.; perception priority information can be the importance level of the training assisted driving data to assisted driving inference; and continuous rate of change information can be the magnitude of change of target features in continuously collected training assisted driving data per unit time.

[0091] Among them, the reconfigurable TPU (Tensor Processing Unit) computing device can be a dedicated neural network processor based on a configurable computing unit array that supports dynamic reallocation of hardware resources.

[0092] The operator combination information of the driver assistance module can be information used to describe the characteristics and interaction relationships of each operator in the driver assistance module, which may include the operator type, the total number of operators, and the execution dependencies between operators.

[0093] The preliminary decomposition parameters can be a pre-determined set of parameters used to decompose the training assisted driving data, and may include decomposition size, overlap ratio, and precision compression ratio. Specifically, the decomposition size can be the spatial or temporal dimension of dividing the original data into sub-data blocks; the overlap ratio can be the proportion of the overlapping area between adjacent sub-data blocks to a single sub-data block; and the precision compression ratio can be the precision compression coefficient of the sub-data blocks while preserving the core features.

[0094] In one embodiment, the method of determining preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device can be achieved by pre-establishing a mapping rule base for data attribute information, operator combination information, and preliminary decomposition parameters. The preliminary decomposition parameters are obtained by querying the mapping rule base using the current data attribute information and operator combination information as query conditions. The mapping rule base for data attribute information, operator combination information, and preliminary decomposition parameters can be a set of association rules built based on a large amount of assisted driving data and engineering experience, storing preliminary decomposition parameters corresponding to different data attribute information and operator combination information. For example, when the data attribute information is image modality, high perception priority, and high continuous change rate, and the operator combination information is 60% convolution operator, 25% attention operator, and serial execution dependency, the mapped preliminary decomposition parameters are 400×300 pixels (decomposition size), 15% (overlap ratio), and 1 / 1 (precision compression ratio).

[0095] S102, the training assisted driving data is decomposed according to the preliminary decomposition parameters to obtain preliminary decomposition results, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters.

[0096] The preliminary decomposition result can be a set of several sub-data blocks obtained by dividing and compressing the training assisted driving data according to the preliminary decomposition parameters. In one embodiment, the method of decomposing the training assisted driving data according to the preliminary decomposition parameters to obtain the preliminary decomposition result can be as follows: divide the training assisted driving data into first data blocks according to the decomposition size, expand each first data block outward according to the overlap ratio to obtain second data blocks, and compress each second data block according to the precision compression ratio to obtain the preliminary decomposition result.

[0097] In one embodiment, the dynamic reconfiguration of the reconfigurable TPU computing device based on preliminary decomposition parameters can be achieved by pre-establishing a mapping rule base between the preliminary decomposition parameters and the reconfigurable computing core allocation strategy, on-chip storage unit partitioning strategy, and interconnect bus bandwidth allocation strategy. The current preliminary decomposition parameters are used as query conditions to query the mapping rule base to obtain the corresponding reconfigurable computing core allocation strategy, on-chip storage unit partitioning strategy, and interconnect bus bandwidth allocation strategy. A reconfiguration command is generated based on the reconfigurable computing core allocation strategy, on-chip storage unit partitioning strategy, and interconnect bus bandwidth allocation strategy. The reconfiguration command is then issued and the response is verified to complete the dynamic reconfiguration of the reconfigurable TPU computing device. The mapping rule base between the initial decomposition parameters and the reconfigurable computing core allocation strategy, on-chip storage unit partitioning strategy, and interconnect bus bandwidth allocation strategy can be constructed based on hardware adaptation test data of a large number of assisted driving data decomposition scenarios and hardware performance parameters of the reconfigurable TPU computing device. It stores the set of association rules for the reconfigurable computing core allocation strategy, on-chip storage unit partitioning strategy, and interconnect bus bandwidth allocation strategy corresponding to different initial decomposition parameters. For example, when the initial decomposition parameters are 400×300 pixels (decomposition size), 15% (overlap ratio), and 1 / 1 (precision compression ratio), the corresponding reconfigurable computing core allocation strategy is to allocate 8 high-precision computing cores and deploy them in a 2×4 array. The corresponding on-chip storage unit partitioning strategy is to divide the storage into 6 storage partitions, each with a capacity of 16MB. The corresponding interconnect bus bandwidth allocation strategy is to allocate 8Gbps of dedicated bandwidth.

[0098] S103, the preliminary decomposition results are input to the driver assistance module, and the resource status information of the reconfigurable TPU computing device is collected.

[0099] In one embodiment, the method of inputting the preliminary decomposition results to the driver assistance module can be to determine the input priority of each sub-data block in the preliminary decomposition results based on the execution dependencies between the operators in the driver assistance module, and then transmit each sub-data block in the preliminary decomposition results to the input layer of the driver assistance module in batches for calculation processing according to the hardware resource capacity of the reconfigurable TPU computing device.

[0100] The resource status information of the reconfigurable TPU computing device can be information describing the real-time hardware resource utilization of the reconfigurable TPU computing device during the processing of preliminary decomposition results. This resource status information can include the computational utilization, bandwidth utilization, and local memory occupancy of the processing units within the reconfigurable TPU computing device. Specifically, a processing unit can be a configurable computing core within the reconfigurable TPU computing device; computational utilization can be the percentage of actual working time of a processing unit per unit of time; bandwidth utilization can be the percentage of actual data transfer volume of the interconnect bus per unit of time relative to the bus's maximum transmission capacity; and local memory occupancy can be the percentage of used storage space in on-chip storage units relative to the total storage space.

[0101] In one embodiment, the resource status information of the reconfigurable TPU computing device can be collected in real time by using the hardware performance monitoring unit (PMU) built into the reconfigurable TPU computing device to collect computing utilization, bandwidth utilization, and local memory occupancy.

[0102] S104, optimize the preliminary decomposition parameters according to the resource status information, and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop condition to obtain the target decomposition parameters, which are used to decompose the actual assisted driving data according to the target decomposition parameters.

[0103] The target decomposition parameters can be the decomposition parameters used during the assisted driving inference process in the actual driving process of the vehicle. That is, during the actual driving process of the vehicle, the actual assisted driving data is collected by the on-board sensors on the vehicle, and the actual assisted driving data is decomposed according to the target decomposition parameters to obtain the actual decomposition results. Finally, the assisted driving module deployed on the reconfigurable TPU computing device calculates and processes the actual decomposition results.

[0104] In one embodiment, optimizing the initial decomposition parameters based on resource status information and dynamically reconfiguring the reconfigurable TPU computing device can be achieved by substituting the current computing utilization, bandwidth utilization, and local memory occupancy into a pre-built comprehensive evaluation score calculation function to obtain a comprehensive evaluation score. Based on the comprehensive evaluation score, the gradient direction and magnitude of the decomposition size, overlap ratio, and precision compression ratio in the initial decomposition parameters are determined as the parameter optimization direction and magnitude. The decomposition size, overlap ratio, and precision compression ratio are then adjusted according to the parameter optimization direction and magnitude to obtain new initial decomposition parameters. The training assisted driving data is decomposed according to the initial decomposition parameters to obtain the initial decomposition result. The reconfigurable TPU computing device is then dynamically reconfigured based on the initial decomposition parameters. This process is repeated until the initial decomposition parameters meet a preset optimization stop condition, resulting in the target decomposition parameters.

[0105] The preliminary decomposition parameters satisfying the preset optimization stopping conditions can be determined by: the comprehensive evaluation score calculated based on the resource status information corresponding to the iteratively optimized preliminary decomposition parameters exceeding a preset evaluation score threshold for three consecutive iterations; the gradient magnitudes of the decomposition size, overlap ratio, and precision compression ratio all being less than preset gradient thresholds; or the number of iterations exceeding a preset iteration number threshold. In this case, the preliminary decomposition parameters satisfying the preset optimization stopping conditions are the target decomposition parameters.

[0106] In one embodiment, optimizing the preliminary decomposition parameters based on the resource status information includes: when the computing utilization rate exceeds a preset computing utilization rate threshold, calculating a decomposition size optimization coefficient based on the computing utilization rate and the preset computing utilization rate threshold, and multiplying the decomposition size optimization coefficient by the decomposition size to obtain an optimized decomposition size; when the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, calculating an overlap ratio optimization coefficient based on the bandwidth utilization rate and the preset bandwidth utilization rate threshold, and multiplying the overlap ratio optimization coefficient by the overlap ratio to obtain an optimized overlap ratio; when the local memory occupancy rate exceeds a preset occupancy rate threshold, calculating a precision compression ratio optimization coefficient based on the local memory occupancy rate and the preset occupancy rate threshold, and multiplying the precision compression ratio optimization coefficient by the precision compression ratio to obtain an optimized precision compression ratio.

[0107] The preset computing utilization threshold can be a pre-defined critical value used to determine whether the load on the processing units in the reconfigurable TPU computing device is within a reasonable range, such as 80%. Correspondingly, if the computing utilization exceeds the preset threshold, it indicates that the processing unit load is too high, the processing pressure of the sub-data blocks corresponding to the current decomposition size exceeds the hardware's carrying capacity, easily leading to increased data processing latency and affecting the real-time performance of assisted driving reasoning. The decomposition size optimization coefficient can be an adjustment coefficient greater than 1 used to reduce the number of sub-data blocks by increasing the decomposition size.

[0108] In one embodiment, the method of calculating the decomposition size optimization coefficient based on the computational utilization rate and a preset computational utilization rate threshold, and multiplying the decomposition size optimization coefficient by the decomposition size to obtain the optimized decomposition size, can be achieved by calculating the difference between the computational utilization rate and the preset computational utilization rate threshold, dividing the difference by the preset computational utilization rate threshold, adding the division result to 1 to obtain the decomposition size optimization coefficient, and multiplying the decomposition size optimization coefficient by the current decomposition size to obtain the optimized decomposition size.

[0109] The preset bandwidth utilization threshold can be a pre-defined critical value used to determine whether the data transmission load of the interconnect bus in the reconfigurable TPU computing device is within a safe operating range, such as 85%. Correspondingly, if the bandwidth utilization exceeds the preset threshold, it indicates that the data transmission load of the interconnect bus is saturated. The current overlap ratio causes repeated transmission in the overlapping areas of adjacent sub-data blocks, consuming excessive bus bandwidth, causing data transmission congestion, and reducing overall processing efficiency. The overlap ratio optimization coefficient can be an adjustment coefficient less than 1 used to reduce the amount of repeated data transmission by lowering the overlap ratio.

[0110] In one embodiment, the method of calculating the overlap ratio optimization coefficient based on the bandwidth utilization rate and the preset bandwidth utilization rate threshold, and multiplying the overlap ratio optimization coefficient by the overlap ratio to obtain the optimized overlap ratio can be achieved by calculating the difference between the bandwidth utilization rate and the preset bandwidth utilization rate threshold, dividing the difference by the preset bandwidth utilization rate threshold, and using the 1-subtraction method to calculate the result to obtain the overlap ratio optimization coefficient, and multiplying the overlap ratio optimization coefficient by the current overlap ratio to obtain the optimized overlap ratio.

[0111] The preset occupancy threshold can be a pre-set critical value used to determine whether the local memory resource usage in the reconfigurable TPU computing device is within a controllable range, such as 85%. Correspondingly, if the local memory occupancy exceeds the preset threshold, it indicates insufficient available space in the local memory, and the storage requirement of the sub-data block corresponding to the current precision compression ratio exceeds the storage resource limit, easily leading to data buffer overflow and increasing processing latency and energy consumption. The precision compression ratio optimization coefficient can be an adjustment coefficient less than 1 used to reduce the storage space usage of sub-data blocks by increasing the precision compression level.

[0112] In one embodiment, the method of calculating the precision compression ratio optimization coefficient based on the local memory occupancy rate and a preset occupancy rate threshold, and multiplying the precision compression ratio optimization coefficient by the precision compression ratio to obtain the optimized precision compression ratio, can be achieved by calculating the difference between the local memory occupancy rate and the preset occupancy rate threshold, dividing the difference by the preset occupancy rate threshold, and using the 1-subtraction method to calculate the result to obtain the precision compression ratio optimization coefficient. The precision compression ratio optimization coefficient is then multiplied by the current precision compression ratio to obtain the optimized precision compression ratio.

[0113] Accordingly, the preliminary decomposition parameters can meet the preset optimization stopping conditions if the computational utilization rate corresponding to the preliminary decomposition parameters does not exceed the preset computational utilization rate threshold, the corresponding bandwidth utilization rate does not exceed the preset bandwidth utilization rate threshold, and the corresponding local memory occupancy rate does not exceed the preset occupancy rate threshold.

[0114] The advantage of this approach is that by calculating and applying optimization coefficients, the optimized decomposition parameters can be quickly and dynamically adapted to the hardware resource status of the reconfigurable TPU computing device.

[0115] In this embodiment, training assisted driving data is acquired, and preliminary decomposition parameters are determined based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device. The data attribute information includes modality type information, perception priority information, and continuous rate of change information; the preliminary decomposition parameters include decomposition size, overlap ratio, and accuracy compression ratio. The training assisted driving data is decomposed according to the preliminary decomposition parameters to obtain preliminary decomposition results, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters. The preliminary decomposition results are input to the assisted driving module, and the resource status information of the reconfigurable TPU computing device is collected. The preliminary decomposition parameters are optimized based on the resource status information, and the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet a preset optimization stop condition, resulting in target decomposition parameters, which are used to decompose actual assisted driving data according to the target decomposition parameters. The aforementioned assisted driving data decomposition method using a reconfigurable TPU computing device is based on a dynamic decomposition strategy driven by data features and computing operators. Combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, it achieves adaptive and efficient data decomposition that precisely matches data characteristics and inference requirements. Simultaneously, it constructs a flexible hardware architecture that can work in real time with the decomposition parameters, achieving computing power optimization and dynamic resource adaptation, thereby significantly improving inference efficiency and accuracy.

[0116] Figure 2 This is a flowchart illustrating another method for decomposing assisted driving data using a reconfigurable TPU computing device provided in this application. Figure 2 As shown, the specific steps include the following:

[0117] S201, acquire training assisted driving data, and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio.

[0118] S202, determine the decomposition movement step size and target decomposition size according to the decomposition size and the overlap ratio, and divide the training assisted driving data according to the decomposition movement step size and the target decomposition size to obtain basic data blocks.

[0119] The decomposition step size can be the distance between the starting positions of two adjacent basic data blocks when sliding the training assisted driving data; the target decomposition size can be the actual size of the final determined single basic data block.

[0120] In one embodiment, the decomposition movement step size and target decomposition size are determined based on the decomposition size and overlap ratio. This can be achieved by multiplying the decomposition size, overlap ratio, and 0.5 to obtain the decomposition movement step size, and then adding the decomposition size to twice the overlap ratio to obtain the target decomposition size. For example, with initial decomposition parameters of 400×300 pixels (decomposition size) and 10% (overlap ratio), the corresponding decomposition movement step sizes are 20 pixels (horizontal) and 15 pixels (vertical), resulting in a target decomposition size of 440×330 pixels.

[0121] Among them, the basic data block can be the original feature data block that retains the preset overlapping area after the training assisted driving data is divided according to the target decomposition size and the decomposition movement step size.

[0122] In one embodiment, the method of dividing the training assisted driving data into basic data blocks according to the decomposition movement step size and the target decomposition size can be achieved by constructing a sliding window with the starting position of the training assisted driving data as the origin and the same as the target decomposition size. The window is then slid along the spatial or temporal dimensions of the training assisted driving data according to the decomposition movement step size. After each slide, the dataset within the sliding window is taken as a basic data block until the sliding window covers all the training assisted driving data, thus obtaining all basic data blocks.

[0123] Figure 3 This is an example diagram of the segmentation of image data provided in this application. For example... Figure 3 As shown, the gray area is equal to the decomposition size (400×300 pixels) in the initial decomposition parameters. The dashed area is the sliding window, which is equal to the target decomposition size (440×330 pixels). The decomposition movement step is 20 pixels (horizontal) and 15 pixels (vertical).

[0124] S203, the basic data block is reconstructed and encoded according to the precision compression ratio to obtain a compressed data block; wherein, the reconstruction encoding includes quantization encoding and sparsification encoding.

[0125] Reconstruction coding can be a feature-preserving compression method that performs compression processing on the original data of the basic data block based on the precision compression ratio. It can include quantization coding and sparsity coding. Specifically, quantization coding can be a coding method that maps high-precision original data (such as FP32 floating-point type) in the basic data block to low-precision data (such as FP16, INT8); sparsity coding can be a coding method that removes redundant data (such as background noise, invalid point cloud) in the basic data block that is meaningless for assisted driving reasoning, and retains only the core effective data (such as target contour pixels, effective ranging data).

[0126] The compressed data block can be a low-precision, low-redundancy data block obtained by reconstructing and encoding a basic data block. In one embodiment, the method of reconstructing and encoding the basic data block according to the precision compression ratio to obtain the compressed data block can be as follows: determine the target quantization strategy based on the current precision compression ratio and a pre-built mapping relationship between the precision compression ratio and the quantization strategy (e.g., the quantization strategy corresponding to a 1 / 2 precision compression ratio is FP32 to FP16, and the quantization strategy corresponding to a 1 / 4 precision compression ratio is FP32 to INT8). Perform linear quantization encoding on the basic data block according to the target quantization strategy to obtain quantized data. Calculate the sparsity threshold (e.g., data variance) corresponding to the quantized data, and perform sparsity encoding on the quantized data according to the sparsity threshold to retain the core effective data. Perform variable-length encoding and storage on the core effective data to obtain the compressed data block.

[0127] S204, Dynamically reconfigure the reconfigurable TPU computing device based on the preliminary decomposition parameters.

[0128] In one embodiment, after reconstructing and encoding the basic data block according to the precision compression ratio to obtain a compressed data block, the method further includes: determining the computational dependency information between the compressed data blocks according to the operator combination information.

[0129] Among them, the computational dependency information between compressed data blocks can be information that characterizes the sequential processing order and parallel / serial execution relationship of different compressed data blocks in the operator execution process of the driver assistance module.

[0130] In one embodiment, the method of determining the computational dependency information between compressed data blocks based on operator combination information can be as follows: matching each compressed data block to a suitable operator based on the attribute information of the compressed data block and the attribute information of each operator; deriving the dependency type and processing order between compressed data blocks bound to different operators based on the execution dependency relationship between each operator; and inferring the parallel / serial execution relationship between compressed data blocks based on the dependency type and processing order.

[0131] Accordingly, in one embodiment, the dynamic reconfiguration of the reconfigurable TPU computing device based on the preliminary decomposition parameters includes: configuring the data bit width, computing mode, and local memory capacity of the processing units in the reconfigurable TPU computing device according to the attribute information of the compressed data blocks; allocating compressed data blocks to the processing units according to the computing dependency information to obtain the mapping relationship between the processing units and the compressed data blocks; determining the data communication requirement information between the processing units according to the computing dependency information and the mapping relationship, and configuring the connection topology and buffer depth of the on-chip interconnect network between the processing units in the reconfigurable TPU computing device according to the data communication requirement information.

[0132] The attribute information of the compressed data block can be information used to characterize the core features and processing requirements of the compressed data block, including data modality type, precision level, data size, perception priority, and target operator type.

[0133] Among them, the data bit width can be the number of binary bits of a single data that the processing unit can process; the calculation mode can be the hardware operation configuration mode of the processing unit for different operator types; and the local memory capacity can be the size of the on-chip storage resources allocated to the processing unit.

[0134] In one embodiment, the data bit width, computing mode, and local memory capacity of the processing unit in the reconfigurable TPU computing device are configured based on the attribute information of the compressed data block. This can be achieved by pre-establishing a mapping rule base for the attribute information, data bit width, computing mode, and local memory capacity of the compressed data block, and using the attribute information of the current compressed data block as the query condition to query the mapping rule base to obtain the corresponding data bit width, computing mode, and local memory capacity. The mapping rule base for the attribute information, data bit width, calculation mode, and local memory capacity of compressed data blocks can be a set of association rules that store the data bit width, calculation mode, and local memory capacity corresponding to the attribute information of different compressed data blocks. This is based on hardware adaptation test data of a large number of assisted driving data decomposition scenarios and hardware performance parameters of reconfigurable TPU computing devices. For example, when the attribute information of compressed data blocks is image modality (data modality type), FP16 (precision level), 16MB (data size), high perception priority (perception priority), and convolution operator type (target operator type), the mapped data bit width is 16 bits, the mapped calculation mode is convolution calculation mode, and the mapped local memory capacity is 20MB.

[0135] The mapping relationship between processing units and compressed data blocks can represent a one-to-one or one-to-many allocation relationship between each processing unit and compressed data block in the reconfigurable TPU computing device. In one embodiment, the method of allocating compressed data blocks to processing units based on computational dependency information to obtain the mapping relationship between processing units and compressed data blocks can be achieved by identifying parallel dependency compressed data block groups (a set of compressed data blocks without sequential order) and serial dependency links (a sequence of compressed data blocks with sequential order) in the computational dependency information. Then, based on the number of processing units in the reconfigurable TPU computing device, an independent processing unit (or processing unit group) is allocated to each parallel dependency group. For serial dependency links, processing units are allocated sequentially to each compressed data block in the link according to the execution order (or previous processing units are reused). Finally, the compressed data blocks corresponding to each processing unit are recorded to form a mapping relationship table.

[0136] The data communication requirement information between processing units can be information characterizing the data communication process between processing units that have data interaction, and may include the communication direction and the amount of communication data. In one embodiment, the method of determining the data communication requirement information between processing units based on computational dependency information and mapping relationship can be as follows: extracting compressed data block pairs that have data interaction from the computational dependency information, then finding the processing units corresponding to each compressed data block pair through the mapping relationship, determining the communication direction and the amount of communication data between the two processing units, and summarizing them to form the data communication requirement information.

[0137] The on-chip interconnect network can be a hardware network architecture in a reconfigurable TPU computing device that connects various processing units and enables data transmission between processing units; the connection topology can be the connection method and link configuration of each processing unit in the on-chip interconnect network; and the buffer depth can be the storage capacity of each communication link buffer in the on-chip interconnect network.

[0138] In one embodiment, the connection topology and buffer depth of the on-chip interconnection network between processing units in the reconfigurable TPU computing device can be configured according to data communication requirements. This can be achieved by first dividing the data communication requirements into three levels: high, medium, and low, based on real-time requirements. For high-level data communication requirements, a point-to-point direct connection topology is configured and a large-capacity buffer (e.g., buffer depth of 16MB) is allocated. For medium-level data communication requirements, a shared crossbar topology is configured and a medium-capacity buffer (e.g., buffer depth of 8MB) is allocated. For low-level data communication requirements, a bus topology is configured and a small-capacity buffer (e.g., buffer depth of 4MB) is allocated. The topology switching and buffer capacity allocation are completed by issuing configuration commands through the interconnection network controller of the reconfigurable TPU computing device.

[0139] In one embodiment, after determining the data communication requirement information between the processing units based on the computational dependency information and the mapping relationship, the method further includes: generating parallel data stream orchestration information based on the mapping relationship and the computational dependency information; generating a synchronization processing time for the processing units based on the parallel data stream orchestration information, used to control the processing units to process corresponding compressed data blocks according to the synchronization processing time during model training, so as to achieve synchronous processing of the compressed data blocks.

[0140] Among them, the parallel data flow orchestration information can be structured information of multi-level parallel groups of processing units divided based on computational dependencies. It can include the partitioning scheme of parallel groups of processing units at each level and the data flow between processing units in each parallel group.

[0141] In one embodiment, the method of generating parallel data stream orchestration information based on mapping relationships and computational dependency information can be achieved by dividing parallel units (including independent compressed data blocks and indivisible serial dependency links) that have no mutual dependencies in the computational dependency information into parallel groups at different levels. Based on the mapping relationship, the processing unit corresponding to each compressed data block in the parallel group and the data flow direction between each processing unit are determined, which is the parallel data stream orchestration information.

[0142] The synchronous processing time of a processing unit can be the start time set for each processing unit to ensure that the next processing unit can simultaneously receive the processing results data output by all the preceding processing units. Accordingly, during model training, the processing units are controlled to process the corresponding compressed data blocks according to the synchronous processing time. This can be achieved by having a processing unit start processing the corresponding compressed data block at the corresponding synchronous processing time. If a processing unit finishes early, it enters a synchronous waiting state until all compressed data blocks in the parallel group containing the corresponding compressed data block have been processed, and then the processing units of the next level of parallel group are triggered to start synchronously according to the data flow.

[0143] In one embodiment, the method of generating the synchronization processing time of the processing unit based on the parallel data stream orchestration information can be achieved by calculating the estimated processing time of each processing unit based on the attribute information of each processing unit in the parallel data stream orchestration information and the attribute information of the corresponding compressed data block, and determining the synchronization processing time of the processing unit based on the estimated processing time of each processing unit and the data flow direction between each processing unit.

[0144] The advantage of this scheme is that the precise planning of synchronous processing time enables the progress alignment of processing units within the parallel group, thereby improving the utilization rate of parallel computing resources.

[0145] The advantage of this scheme is that, based on the hardware configuration of compressed data block attribute information and computational dependency information, the computing power, storage resources and data processing requirements of the processing unit are precisely matched, avoiding resource waste. At the same time, the interconnection network topology and buffer depth are dynamically configured according to communication requirements to ensure the efficiency and real-time performance of data transmission between processing units and reduce communication congestion.

[0146] S205, the compressed data block is input to the driver assistance module, and the resource status information of the reconfigurable TPU computing device is collected.

[0147] In one embodiment, the compressed data blocks are input to the driver assistance module by determining the input priority of each compressed data block based on the execution dependencies between operators in the driver assistance module, and then transmitting the compressed data blocks in batches to the input layer of the driver assistance module for computation processing according to the hardware resource capacity of the reconfigurable TPU computing device.

[0148] S206, optimize the preliminary decomposition parameters according to the resource status information, and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop condition to obtain the target decomposition parameters, which are used to decompose the actual assisted driving data according to the target decomposition parameters.

[0149] The advantage of this scheme is that it ensures that the basic data blocks fully cover the original training data for assisted driving, while preserving the spatiotemporal continuity of the data through overlapping areas, avoiding feature breaks caused by partitioning. At the same time, it adopts a reconstruction method that combines quantization coding and sparse coding to achieve accurate data compression, reducing storage and transmission pressure while meeting accuracy requirements.

[0150] Figure 4 This is a flowchart illustrating another method for decomposing assisted driving data using a reconfigurable TPU computing device provided in this application. Figure 4 As shown, the specific steps include the following:

[0151] S401, acquire training data for assisted driving.

[0152] S402, based on the modal type information of the training assisted driving data and the operator computational complexity in the operator combination information, allocate the benchmark decomposition size, benchmark overlap ratio, and benchmark precision compression ratio.

[0153] Among them, operator computational complexity can be a quantitative indicator that represents the amount of computation, number of memory accesses and data transfer required for a single operator or combination of operators to perform an operation in the driver assistance module. Specifically, it can be FLOP (floating-point operations).

[0154] The baseline decomposition size can be a pre-set decomposition size based on the modal type information and operator computational complexity of the training assisted driving data; the baseline overlap ratio can be a pre-set overlap ratio based on the modal type information and operator computational complexity of the training assisted driving data; and the baseline precision compression ratio can be a pre-set precision compression ratio based on the modal type information and operator computational complexity of the training assisted driving data.

[0155] In one embodiment, the baseline decomposition size, baseline overlap ratio, and baseline precision compression ratio are assigned based on the modality type information and operator computational complexity in the operator combination information of the training assisted driving data. This can be achieved by pre-establishing a mapping rule base for modality type information, operator computational complexity, baseline decomposition size, baseline overlap ratio, and baseline precision compression ratio. The current modality type information and operator computational complexity are used as query conditions to retrieve these values ​​from the mapping rule base. The mapping rule base for modality type information, operator computational complexity, baseline decomposition size, baseline overlap ratio, and baseline precision compression ratio can be a set of association rules built based on a large amount of measured data from assisted driving scenarios, average computational efficiency test results of operators on TPUs, and engineering practice experience. This set stores the baseline decomposition size, baseline overlap ratio, and baseline precision compression ratio corresponding to different modality type information and operator computational complexity. For example, when the modality type information is image modality and the operator computational complexity is... When using FLOP, the baseline resolution size is 600×450 pixels, the baseline overlap ratio is 12%, and the baseline precision compression ratio is 1 / 2.

[0156] S403, based on the perception priority information of the training assisted driving data and the continuous rate of change information, determine the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the accuracy compression ratio adjustment coefficient.

[0157] Among them, the decomposition size adjustment coefficient can be a coefficient determined based on the perception priority information and continuous rate of change information of the training assisted driving data to correct the benchmark decomposition size; the overlap ratio adjustment coefficient can be a coefficient determined based on the perception priority information and continuous rate of change information of the training assisted driving data to correct the benchmark overlap ratio; and the precision compression ratio adjustment coefficient can be a coefficient determined based on the perception priority information and continuous rate of change information of the training assisted driving data to correct the benchmark precision compression ratio.

[0158] In one embodiment, the method of determining the decomposition size adjustment coefficient, overlap ratio adjustment coefficient, and accuracy compression ratio adjustment coefficient based on the perception priority information and continuous rate of change information of the training assisted driving data can be achieved by pre-establishing a mapping rule base for perception priority information, continuous rate of change information, decomposition size adjustment coefficient, overlap ratio adjustment coefficient, and accuracy compression ratio adjustment coefficient. The current perception priority information and continuous rate of change information are used as query conditions to query the mapping rule base to obtain the decomposition size adjustment coefficient, overlap ratio adjustment coefficient, and accuracy compression ratio adjustment coefficient. The mapping rule base for perception priority information, continuous rate of change information, decomposition size adjustment coefficient, overlap ratio adjustment coefficient, and accuracy compression ratio adjustment coefficient can be a set of association rules built based on the perception importance requirements of assisted driving data and the impact of continuous data changes on processing accuracy. This set stores the decomposition size adjustment coefficient, overlap ratio adjustment coefficient, and accuracy compression ratio adjustment coefficient corresponding to different perception priority information and continuous rate of change information. For example, when the perception priority information is high perception priority and the continuous rate of change information is fast (inter-frame feature change ≥ 30%), the decomposition size adjustment coefficient is 0.8, the overlap ratio adjustment coefficient is 1.2, and the accuracy compression ratio adjustment coefficient is 1.2.

[0159] S404, multiply the reference decomposition size and the decomposition size adjustment coefficient, multiply the reference overlap ratio and the overlap ratio adjustment coefficient, and multiply the reference accuracy compression ratio and the accuracy compression ratio adjustment coefficient to obtain the basic decomposition parameters.

[0160] The basic decomposition parameters include the basic decomposition size, the basic overlap ratio, and the basic accuracy compression ratio.

[0161] S405, calculate the estimated performance information based on the basic decomposition parameters and the operator combination information, and iteratively optimize the basic decomposition parameters based on the estimated performance information to obtain preliminary decomposition parameters; wherein, the preliminary decomposition parameters include decomposition size, overlap ratio and precision compression ratio.

[0162] The estimated performance information can be a performance indicator of the driver assistance module processing training driver assistance data, calculated based on the basic decomposition parameters and operator combination information. This can include estimated processing latency and estimated data transmission bandwidth. In one embodiment, the method for calculating the estimated performance information based on the basic decomposition parameters and operator combination information can be as follows: The number of data blocks, average computational cost per data block, and average memory access cost per data block are calculated based on the operator computational complexity of each operator in the operator combination information, the basic decomposition size, and the basic precision compression ratio. The average peak computing power of each processing unit of the reconfigurable TPU computing device is then obtained. The average computational cost per data block is divided by this average value and multiplied by the number of data blocks to obtain the estimated processing latency. Finally, the average memory access cost per data block is multiplied by the number of data blocks and divided by the estimated processing latency to obtain the estimated data transmission bandwidth.

[0163] In one embodiment, the method of iteratively optimizing the basic decomposition parameters based on the estimated performance information to obtain the preliminary decomposition parameters can be achieved by increasing the basic accuracy compression ratio by a first preset step size and decreasing the basic decomposition size by a second preset step size when the estimated processing latency exceeds a preset processing latency threshold, and decreasing the basic overlap ratio by a third preset step size when the estimated data transmission bandwidth exceeds a preset data transmission bandwidth threshold.

[0164] The estimated performance information includes estimated throughput and estimated communication overhead. Accordingly, in one embodiment, the iterative optimization of the basic decomposition parameters based on the estimated performance information to obtain preliminary decomposition parameters includes: substituting the estimated throughput and estimated communication overhead into a joint optimization objective function to obtain an optimization objective difference; wherein the joint optimization objective function aims to maximize the estimated throughput and minimize the estimated communication overhead; calculating the gradients of the optimization objective difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters, and adjusting the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradients; repeating the above steps until the rate of change of the optimization objective difference is less than a preset rate of change threshold, thereby obtaining the preliminary decomposition parameters.

[0165] The estimated throughput can be the total amount of data that can be processed per unit time; the estimated communication overhead can be the additional data transmission cost incurred due to data partitioning, precision conversion, and intermediate result exchange, and can be represented by the total data transmission delay. The joint optimization objective function can be a scalar function that balances the estimated throughput and the estimated communication overhead, with the joint optimization objective being to maximize the estimated throughput and minimize the estimated communication overhead. Specifically, the joint optimization objective function can be:

[0166] ;in, and These are pre-set weighting coefficients.

[0167] The gradient can be the partial derivative of the optimization objective difference with respect to the base decomposition size, base overlap ratio, or base precision compression ratio, respectively. In one embodiment, the gradients of the optimization objective difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the base decomposition parameters can be calculated using the finite difference method. The finite difference method can be an approximate method for numerically calculating the partial derivative of the objective function with respect to a certain parameter; its core idea is to estimate the gradient by utilizing the change in the objective function value caused by a small change in the parameter value.

[0168] In one embodiment, adjusting the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradient can be achieved by obtaining a preset basic learning rate, multiplying the preset basic learning rate by the magnitude of the gradient corresponding to the basic decomposition size, adding the multiplication result to the basic decomposition size when the gradient direction is positive to obtain the optimized basic decomposition size, and subtracting the multiplication result from the basic decomposition size when the gradient direction is negative to obtain the optimized basic decomposition size, and so on for the basic overlap ratio and basic precision compression ratio.

[0169] The rate of change of the target difference can be the average change of the target difference over multiple iterations; the preset rate of change threshold can be a small positive number as a convergence criterion.

[0170] If the rate of change of the target difference is less than the preset rate of change threshold, it means that the target difference has stabilized and small adjustments to the parameters have no significant effect on performance improvement. In other words, the optimization process has reached or is close to the local optimum under the current constraints, and the iteration can be terminated.

[0171] The advantage of this approach is that iterative optimization, with the joint optimization objective of maximizing the estimated throughput and minimizing the estimated communication overhead, can accurately balance computational efficiency and data transmission costs, avoiding performance imbalances caused by optimizing a single metric.

[0172] S406, the training assisted driving data is decomposed according to the preliminary decomposition parameters to obtain preliminary decomposition results, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters.

[0173] S407, the preliminary decomposition results are input to the driver assistance module, and the resource status information of the reconfigurable TPU computing device is collected.

[0174] S408, optimize the preliminary decomposition parameters according to the resource status information, and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop condition to obtain the target decomposition parameters, which are used to decompose the actual assisted driving data according to the target decomposition parameters.

[0175] The advantage of this approach is that by using a two-layer parameter determination logic consisting of baseline parameters and adjustment coefficients, it achieves a precise match between the decomposed parameters and the data characteristics of the training assisted driving data. Furthermore, through performance prediction simulation and iterative optimization, it avoids performance defects of the basic decomposed parameters in advance, reducing the trial-and-error costs of subsequent hardware reconstruction.

[0176] Figure 5 This is a schematic diagram of the structure of an assisted driving data decomposition system for a reconfigurable TPU computing device provided in this application. Figure 5 As shown, the system includes:

[0177] The preliminary scheme determination module 510 is used to acquire training assisted driving data and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio.

[0178] TPU reconstruction module 520 is used to decompose the training assisted driving data according to the preliminary decomposition parameters to obtain preliminary decomposition results, and to dynamically reconstruct the reconfigurable TPU computing device based on the preliminary decomposition parameters.

[0179] The decomposition result input module 530 is used to input the preliminary decomposition result to the assisted driving module and collect the resource status information of the reconfigurable TPU computing device;

[0180] The decomposition parameter optimization module 540 is used to optimize the preliminary decomposition parameters according to the resource status information and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop condition to obtain the target decomposition parameters, which are used to decompose the actual assisted driving data according to the target decomposition parameters.

[0181] Optionally, the TPU reconfiguration module 520 is specifically used for:

[0182] The decomposition movement step and target decomposition size are determined based on the decomposition size and the overlap ratio, and the training assisted driving data is divided according to the decomposition movement step and the target decomposition size to obtain basic data blocks.

[0183] The basic data block is reconstructed and encoded according to the stated precision compression ratio to obtain a compressed data block; wherein, the reconstruction encoding includes quantization encoding and sparsification encoding;

[0184] Accordingly, the decomposition result input module 530 is specifically used for:

[0185] The compressed data block is input into the driver assistance module.

[0186] Optionally, the TPU reconfiguration module 520 is further configured to:

[0187] The computational dependency information between the compressed data blocks is determined based on the operator combination information;

[0188] Based on the attribute information of the compressed data block, configure the data bit width, computing mode, and local memory capacity of the processing unit in the reconfigurable TPU computing device;

[0189] Based on the computational dependency information, compressed data blocks are allocated to the processing unit to obtain the mapping relationship between the processing unit and the compressed data blocks;

[0190] Based on the computational dependency information and the mapping relationship, the data communication requirement information between the processing units is determined, and the connection topology and buffer depth of the on-chip interconnect network between the processing units in the reconfigurable TPU computing device are configured based on the data communication requirement information.

[0191] Optionally, the TPU reconfiguration module 520 is further configured to:

[0192] Parallel data stream orchestration information is generated based on the mapping relationship and the computational dependency information;

[0193] The synchronous processing time of the processing unit is generated based on the parallel data stream orchestration information, and is used to control the processing unit to process the corresponding compressed data block according to the synchronous processing time during model training, so as to realize the synchronous processing of the compressed data block.

[0194] Optionally, the preliminary scheme determination module 510 is specifically used for:

[0195] Based on the modality type information of the training assisted driving data and the operator computational complexity in the operator combination information, the baseline decomposition size, the baseline overlap ratio, and the baseline precision compression ratio are allocated;

[0196] Based on the perception priority information and the continuous rate of change information of the training assisted driving data, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the accuracy compression ratio adjustment coefficient are determined.

[0197] Multiply the baseline decomposition size and the decomposition size adjustment factor, multiply the baseline overlap ratio and the overlap ratio adjustment factor, and multiply the baseline precision compression ratio and the precision compression ratio adjustment factor to obtain the basic decomposition parameters;

[0198] The estimated performance information is calculated based on the basic decomposition parameters and the operator combination information, and the basic decomposition parameters are iteratively optimized based on the estimated performance information to obtain the preliminary decomposition parameters.

[0199] Optionally, the estimated performance information includes estimated throughput and estimated communication overhead;

[0200] Accordingly, the preliminary scheme determination module 510 is specifically used for:

[0201] Substituting the estimated throughput and the estimated communication overhead into the joint optimization objective function yields the optimization objective difference; wherein the joint optimization objective function aims to maximize the estimated throughput and minimize the estimated communication overhead.

[0202] Calculate the gradient of the optimization target difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters, and adjust the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradient.

[0203] Repeat the above steps until the rate of change of the optimization target difference is less than the preset rate of change threshold, and obtain the preliminary decomposition parameters.

[0204] Optionally, the resource status information includes the computing utilization rate, bandwidth utilization rate, and local memory occupancy rate of the processing units in the reconfigurable TPU computing device.

[0205] Accordingly, the decomposition parameter optimization module 540 is specifically used for:

[0206] If the computational utilization rate exceeds a preset computational utilization rate threshold, the decomposition size optimization coefficient is calculated based on the computational utilization rate and the preset computational utilization rate threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain the optimized decomposition size;

[0207] When the bandwidth utilization exceeds a preset bandwidth utilization threshold, an overlap ratio optimization coefficient is calculated based on the bandwidth utilization and the preset bandwidth utilization threshold, and the overlap ratio optimization coefficient is multiplied by the overlap ratio to obtain the optimized overlap ratio.

[0208] If the local memory occupancy rate exceeds a preset occupancy rate threshold, a precision compression ratio optimization coefficient is calculated based on the local memory occupancy rate and the preset occupancy rate threshold, and the precision compression ratio optimization coefficient is multiplied by the precision compression ratio to obtain the optimized precision compression ratio.

[0209] In this embodiment, a preliminary scheme determination module is used to acquire training assisted driving data and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device. The data attribute information includes modal type information, perception priority information, and continuous rate of change information; the preliminary decomposition parameters include decomposition size, overlap ratio, and accuracy compression ratio. A TPU reconstruction module is used to decompose the training assisted driving data according to the preliminary decomposition parameters to obtain preliminary decomposition results, and dynamically reconstruct the reconfigurable TPU computing device based on the preliminary decomposition parameters. A decomposition result input module is used to input the preliminary decomposition results to the assisted driving module and collect the resource status information of the reconfigurable TPU computing device. A decomposition parameter optimization module is used to optimize the preliminary decomposition parameters according to the resource status information and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameters meet a preset optimization stop condition, obtaining target decomposition parameters for decomposing actual assisted driving data according to the target decomposition parameters. The aforementioned assisted driving data decomposition system of the reconfigurable TPU computing device is based on a dynamic decomposition strategy driven by data features and computing operators. Combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, it achieves adaptive and efficient data decomposition that is precisely matched with data characteristics and inference requirements. Simultaneously, it constructs a flexible hardware architecture that can cooperate with the decomposition parameters in real time, achieving computing power optimization and dynamic resource adaptation, thereby significantly improving inference efficiency and accuracy.

[0210] The assisted driving data decomposition system of the reconfigurable TPU computing device provided in this application embodiment can realize the various processes implemented in the above embodiments. To avoid repetition, it will not be described again here.

[0211] Figure 6 This is a schematic diagram of the structure of an electronic device provided in this application. For example... Figure 6 As shown, this application embodiment also provides an electronic device 600, including a processor 601, a memory 602, and a program or instructions stored in the memory 602 and executable on the processor 601. When the program or instructions are executed by the processor 601, they implement the various processes of the above-described embodiment of the assisted driving data decomposition method of the reconfigurable TPU computing device and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0212] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0213] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiments of the assisted driving data decomposition method for the reconfigurable TPU computing device and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0214] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0215] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0216] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0217] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0218] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. A method for decomposing assisted driving data using a reconfigurable TPU computing device, characterized in that, The method includes: Acquire training assisted driving data, and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio. The decomposition movement step size and target decomposition size are determined based on the decomposition size and the overlap ratio. The training assisted driving data is then divided according to the decomposition movement step size and the target decomposition size to obtain basic data blocks. The basic data blocks are reconstructed and encoded according to the precision compression ratio to obtain compressed data blocks. The reconstruction encoding includes quantization encoding and sparsification encoding. The reconstructable TPU computing device is dynamically reconstructed based on the preliminary decomposition parameters. The compressed data block is input to the driver assistance module, and the resource status information of the reconfigurable TPU computing device is collected; wherein, the resource status information includes the computing utilization rate, bandwidth utilization rate, and local memory occupancy rate of the processing unit in the reconfigurable TPU computing device; When the computational utilization rate exceeds a preset computational utilization rate threshold, a decomposition size optimization coefficient is calculated based on the computational utilization rate and the preset computational utilization rate threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain the optimized decomposition size; when the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, an overlap ratio optimization coefficient is calculated based on the bandwidth utilization rate and the preset bandwidth utilization rate threshold, and the overlap ratio optimization coefficient is multiplied by the overlap ratio to obtain the optimized overlap ratio; when the local memory occupancy rate exceeds a preset occupancy rate threshold, a precision compression ratio optimization coefficient is calculated based on the local memory occupancy rate and the preset occupancy rate threshold, and the precision compression ratio optimization coefficient is multiplied by the precision compression ratio to obtain the optimized precision compression ratio; the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet the preset optimization stop condition to obtain target decomposition parameters, which are used to decompose actual assisted driving data according to the target decomposition parameters.

2. The method for decomposing assisted driving data using a reconfigurable TPU computing device according to claim 1, characterized in that, After reconstructing and encoding the basic data block according to the stated precision compression ratio to obtain a compressed data block, the method further includes: The computational dependency information between the compressed data blocks is determined based on the operator combination information; Accordingly, the dynamic reconfiguration of the reconfigurable TPU computing device based on the preliminary decomposition parameters includes: Based on the attribute information of the compressed data block, configure the data bit width, computing mode, and local memory capacity of the processing unit in the reconfigurable TPU computing device; Based on the computational dependency information, compressed data blocks are allocated to the processing unit to obtain the mapping relationship between the processing unit and the compressed data blocks; Based on the computational dependency information and the mapping relationship, the data communication requirement information between the processing units is determined, and the connection topology and buffer depth of the on-chip interconnect network between the processing units in the reconfigurable TPU computing device are configured based on the data communication requirement information.

3. The method for decomposing assisted driving data using a reconfigurable TPU computing device according to claim 2, characterized in that, After determining the data communication requirement information between the processing units based on the computational dependency information and the mapping relationship, the method further includes: Parallel data stream orchestration information is generated based on the mapping relationship and the computational dependency information; The synchronous processing time of the processing unit is generated based on the parallel data stream orchestration information, and is used to control the processing unit to process the corresponding compressed data block according to the synchronous processing time during model training, so as to realize the synchronous processing of the compressed data block.

4. The method for decomposing assisted driving data using a reconfigurable TPU computing device according to claim 1, characterized in that, The step of determining the preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device includes: Based on the modality type information of the training assisted driving data and the operator computational complexity in the operator combination information, the baseline decomposition size, the baseline overlap ratio, and the baseline precision compression ratio are allocated; Based on the perception priority information and the continuous rate of change information of the training assisted driving data, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the accuracy compression ratio adjustment coefficient are determined. Multiply the baseline decomposition size and the decomposition size adjustment factor, multiply the baseline overlap ratio and the overlap ratio adjustment factor, and multiply the baseline precision compression ratio and the precision compression ratio adjustment factor to obtain the basic decomposition parameters; The estimated performance information is calculated based on the basic decomposition parameters and the operator combination information, and the basic decomposition parameters are iteratively optimized based on the estimated performance information to obtain the preliminary decomposition parameters.

5. The method for decomposing assisted driving data using a reconfigurable TPU computing device according to claim 4, characterized in that, The estimated performance information includes estimated throughput and estimated communication overhead; Accordingly, the iterative optimization of the basic decomposition parameters based on the estimated performance information to obtain preliminary decomposition parameters includes: Substituting the estimated throughput and the estimated communication overhead into the joint optimization objective function yields the optimization objective difference; wherein the joint optimization objective function aims to maximize the estimated throughput and minimize the estimated communication overhead. Calculate the gradient of the optimization target difference with respect to the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters, and adjust the decomposition size, overlap ratio, and precision compression ratio in the basic decomposition parameters according to the direction and magnitude of the gradient. Repeat the above steps until the rate of change of the optimization target difference is less than the preset rate of change threshold, and obtain the preliminary decomposition parameters.

6. A driver assistance data decomposition system for a reconfigurable TPU computing device, characterized in that, The system includes: The preliminary scheme determination module is used to acquire training assisted driving data and determine preliminary decomposition parameters based on the data attribute information of the training assisted driving data and the operator combination information of the assisted driving module deployed on the reconfigurable TPU computing device; wherein, the data attribute information includes modal type information, perception priority information and continuous rate of change information, and the preliminary decomposition parameters include decomposition size, overlap ratio and accuracy compression ratio. The TPU reconstruction module is used to determine the decomposition movement step size and the target decomposition size based on the decomposition size and the overlap ratio, and to divide the training assisted driving data into basic data blocks based on the decomposition movement step size and the target decomposition size; to reconstruct and encode the basic data blocks according to the precision compression ratio to obtain compressed data blocks; wherein, the reconstruction encoding includes quantization encoding and sparsification encoding; and to dynamically reconstruct the reconfigurable TPU computing device based on the preliminary decomposition parameters. The decomposition result input module is used to input the compressed data block to the assisted driving module and collect the resource status information of the reconfigurable TPU computing device; wherein, the resource status information includes the computing utilization rate, bandwidth utilization rate and local memory occupancy rate of the processing unit in the reconfigurable TPU computing device; The decomposition parameter optimization module is used to: calculate a decomposition size optimization coefficient based on the computing utilization rate and the preset computing utilization rate threshold when the computing utilization rate exceeds a preset computing utilization rate threshold, and multiply the decomposition size optimization coefficient by the decomposition size to obtain the optimized decomposition size; calculate an overlap ratio optimization coefficient based on the bandwidth utilization rate and the preset bandwidth utilization rate threshold when the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, and multiply the overlap ratio by the overlap ratio to obtain the optimized overlap ratio; calculate a precision compression ratio optimization coefficient based on the local memory utilization rate and the preset occupancy rate threshold when the local memory occupancy rate exceeds a preset occupancy rate threshold, and multiply the precision compression ratio by the precision compression ratio to obtain the optimized precision compression ratio; and dynamically reconfigure the reconfigurable TPU computing device until the initial decomposition parameters meet a preset optimization stop condition to obtain target decomposition parameters, which are used to decompose actual assisted driving data according to the target decomposition parameters.

7. An electronic device, characterized in that, The device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the assisted driving data decomposition method of the reconfigurable TPU computing device as described in any one of claims 1-5.

8. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the assisted driving data decomposition method of the reconfigurable TPU computing device as described in any one of claims 1-5.

Images