Auxiliary driving data decomposition method and system for reconfigurable TPU computing device

Abstract

The invention discloses an auxiliary driving data decomposition method and system for a reconfigurable TPU computing device, and belongs to the technical field of electric digital data processing. The method comprises the steps of determining a preliminary decomposition parameter according to data attribute information of training aided driving data and operator combination information of an aided driving module deployed on a reconfigurable TPU computing device; the training auxiliary driving data are decomposed according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and the reconfigurable TPU computing device is dynamically reconstructed based on the preliminary decomposition parameters; inputting the preliminary decomposition result into an auxiliary driving module, and collecting resource state information of the reconfigurable TPU computing device; and performing iterative optimization on the initial decomposition parameter according to the resource state information to obtain a target decomposition parameter. According to the technical scheme, self-adaptive and efficient data decomposition accurately matched with data characteristics and reasoning requirements is achieved, computing power optimization and resource dynamic adaptation are achieved, and reasoning efficiency and precision are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electric digital data processing, and particularly relates to an assisted driving data decomposition method and system of a reconfigurable TPU computing device. BACKGROUND

[0002] The assisted driving technology is to obtain assisted driving data through a vehicle-mounted sensor, perform reasoning on the assisted driving data, and output corresponding control instructions based on the reasoning result, so as to realize accurate regulation and control of the vehicle driving state (such as adaptive cruise, emergency obstacle avoidance, etc.). The safety, real-time performance and decision accuracy of the assisted driving technology are highly dependent on the efficiency, stability and adaptability of data reasoning.

[0003] Before data reasoning, data processing is a crucial pre-step. In order to improve the computing efficiency, the prior art usually pre-processes and decomposes the input assisted driving data, for example, divides a high-resolution image into fixed-size tiles. However, this static and universal data decomposition strategy is difficult to adapt to the complex and changeable driving environment and data characteristics in the assisted driving scene, and may instead cause a series of efficiency loss and accuracy deviation problems. SUMMARY

[0004] The present application provides an assisted driving data decomposition method and system of a reconfigurable TPU computing device, aiming to realize adaptive and efficient data decomposition that is accurately matched with data characteristics and reasoning requirements based on a dynamic decomposition strategy driven by data characteristics and computing operators in cooperation, combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, and simultaneously construct a flexible hardware architecture that can cooperate with the decomposition parameters in real time, so as to achieve computing power optimization and dynamic resource adaptation, thereby significantly improving the reasoning efficiency and accuracy.

[0005] In a first aspect, the present application provides an assisted driving data decomposition method of a reconfigurable TPU computing device, which comprises: obtaining training assisted driving data, and determining preliminary decomposition parameters according to data attribute information of the training assisted driving data and operator combination information of an assisted driving module deployed on the reconfigurable TPU computing device; wherein the data attribute information comprises modality type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters comprise decomposition size, overlap ratio and precision compression ratio; 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; The preliminary decomposition parameters are optimized according to the resource state information, and the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet a preset optimization stop condition, so that target decomposition parameters are obtained, which are used for actually assisting in driving data decomposition according to the target decomposition parameters.

[0006] Optionally, the decomposing the training auxiliary driving data according to the preliminary decomposition parameters to obtain a preliminary decomposition result comprises: The decomposition moving step and the target decomposition size are determined according to the decomposition size and the overlap ratio, and the training auxiliary driving data is divided according to the decomposition moving step and the target decomposition size, so that a basic data block is obtained; The basic data block is reconfigured and encoded according to the precision compression ratio, so that a compressed data block is obtained; wherein the reconfigured and encoded comprises quantization encoding and sparsification encoding; Correspondingly, the inputting the preliminary decomposition result into the auxiliary driving module comprises: The compressed data block is input into the auxiliary driving module.

[0007] Optionally, after the reconfigured and encoded of the basic data block according to the precision compression ratio to obtain the compressed data block, the method further comprises: The calculation dependency relationship information between the compressed data blocks is determined according to the operator combination information; Correspondingly, the dynamic reconfiguration of the reconfigurable TPU computing device based on the preliminary decomposition parameters comprises: According to the attribute information of the compressed data block, the data bit width, the calculation mode and the local memory capacity of the processing unit in the reconfigurable TPU computing device are configured; The processing unit is allocated a compressed data block according to the calculation dependency relationship information, so that a mapping relationship between the processing unit and the compressed data block is obtained; The data communication demand information between the processing units is determined according to the calculation dependency relationship information and the mapping relationship, and the connection topology relationship and the buffer depth of the on-chip interconnection network between the processing units in the reconfigurable TPU computing device are configured according to the data communication demand information.

[0008] Optionally, after the data communication demand information between the processing units is determined according to the calculation dependency relationship information and the mapping relationship, the method further comprises: The parallel data stream arrangement information is generated according to the mapping relationship and the calculation dependency relationship information; The parallel data stream arrangement information is used to generate a synchronous processing time of the processing unit, and the processing unit is controlled to process the corresponding compressed data block according to the synchronous processing time in a model training process, so as to realize synchronous processing of the compressed data block.

[0009] Optionally, the preliminary decomposition parameter is determined according to the data attribute information of the training auxiliary driving data and the operator combination information of the auxiliary driving module deployed on the reconfigurable TPU computing device, and includes: The reference decomposition size, the reference overlap ratio and the reference precision compression ratio are allocated according to the modal type information of the training auxiliary driving data and the operator calculation complexity in the operator combination information; The decomposition size adjustment coefficient, the overlap ratio adjustment coefficient and the precision compression ratio adjustment coefficient are determined based on the perception priority information of the training auxiliary driving data and the continuous change rate information; The reference decomposition size is multiplied by the decomposition size adjustment coefficient, the reference overlap ratio is multiplied by the overlap ratio adjustment coefficient, and the reference precision compression ratio is multiplied by the precision compression ratio adjustment coefficient to obtain a basic decomposition parameter; The estimated performance information is calculated according to the basic decomposition parameter and the operator combination information, and the basic decomposition parameter is iteratively optimized according to the estimated performance information to obtain the preliminary decomposition parameter.

[0010] Optionally, the estimated performance information includes an estimated throughput and an estimated communication overhead. Correspondingly, the preliminary decomposition parameter is obtained by iteratively optimizing the basic decomposition parameter according to the estimated performance information, and includes: The estimated throughput and the estimated communication overhead are substituted into a joint optimization objective function to obtain an optimization objective difference value; wherein the joint optimization objective function takes maximizing the estimated throughput and minimizing the estimated communication overhead as the joint optimization objective; The gradient of the optimization objective difference value to the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameter is calculated respectively, and the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameter are adjusted according to the direction and size of the gradient; The above steps are repeatedly executed until the change rate of the optimization objective difference value is less than a preset change rate threshold, and the preliminary decomposition parameter is obtained.

[0011] Optionally, the resource state information includes the calculation utilization rate, the bandwidth utilization rate and the local memory occupancy rate of the processing unit in the reconfigurable TPU computing device. Correspondingly, the preliminary decomposition parameter is optimized according to the resource state information, and includes: In a case where the computing utilization exceeds a preset computing utilization threshold, a decomposition size optimization coefficient is calculated according to the computing utilization and the preset computing utilization threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain an optimized decomposition size; In a case where the bandwidth utilization exceeds a preset bandwidth utilization threshold, an overlap ratio optimization coefficient is calculated according to the bandwidth utilization and the preset bandwidth utilization threshold, and the overlap ratio optimization coefficient is multiplied by the overlap ratio to obtain an optimized overlap ratio; In a case where the local memory occupancy rate exceeds a preset occupancy threshold, an accuracy compression ratio optimization coefficient is calculated according to the local memory occupancy rate and the preset occupancy threshold, and the accuracy compression ratio optimization coefficient is multiplied by the accuracy compression ratio to obtain an optimized accuracy compression ratio.

[0012] In a second aspect, the present application provides an assisted driving data decomposition system of a reconfigurable TPU computing device, the system comprising: A preliminary scheme determination module is configured to obtain training assisted driving data, and determine preliminary decomposition parameters according to data attribute information of the training assisted driving data and operator combination information of an assisted driving module deployed on the reconfigurable TPU computing device; wherein the data attribute information comprises modality type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters comprise a decomposition size, an overlap ratio and an accuracy compression ratio; A TPU reconfiguration module is configured to decompose the training assisted driving data according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and dynamically reconfigure the reconfigurable TPU computing device based on the preliminary decomposition parameters; A decomposition result input module is configured to input the preliminary decomposition result to the assisted driving module, and collect resource state information of the reconfigurable TPU computing device; A decomposition parameter optimization module is configured to optimize the preliminary decomposition parameters according to the resource state information, and dynamically reconfigure the reconfigurable TPU computing device until the preliminary decomposition parameters meet a preset optimization stop condition, to obtain target decomposition parameters for decomposing actual assisted driving data according to the target decomposition parameters.

[0013] Optionally, the TPU reconfiguration module is specifically configured to: A decomposition moving step and a target decomposition size are determined according to the decomposition size and the overlap ratio, and the training assisted driving data is divided according to the decomposition moving step and the target decomposition size to obtain a basic data block; Reconstruct and encode the basic data block according to the precision compression ratio to obtain a compressed data block; wherein the reconstructing and encoding includes quantization encoding and sparsification encoding; Correspondingly, the decomposition result input module is specifically configured to: Input the compressed data block into the auxiliary driving module.

[0014] Optionally, the TPU reconstruction module is further configured to: Determine the calculation dependency relationship information between the compressed data blocks according to the operator combination information; According to the attribute information of the compressed data block, configure the data bit width, calculation mode and local memory capacity of the processing unit in the reconfigurable TPU computing device; According to the calculation dependency relationship information, allocate compressed data blocks to the processing units to obtain the mapping relationship between the processing units and the compressed data blocks; According to the calculation dependency relationship information and the mapping relationship, determine the data communication demand information between the processing units, and according to the data communication demand information, configure the connection topology relationship and buffer depth of the on-chip interconnection network between the processing units in the reconfigurable TPU computing device.

[0015] Optionally, the TPU reconstruction module is further configured to: According to the mapping relationship and the calculation dependency relationship information, generate parallel data stream arrangement information; According to the parallel data stream arrangement information, generate the synchronization processing time of the processing unit, which is used to control the processing unit to process the corresponding compressed data block according to the synchronization processing time in the model training process, so as to realize the synchronization processing of the compressed data block.

[0016] Optionally, the preliminary scheme determination module is specifically configured to: According to the modal type information of the training auxiliary driving data and the operator calculation complexity in the operator combination information, allocate a reference decomposition size, a reference overlap ratio and a reference precision compression ratio; Determine a decomposition size adjustment coefficient, an overlap ratio adjustment coefficient and a precision compression ratio adjustment coefficient based on the perception priority information of the training auxiliary driving data and the continuous change rate information; 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 precision compression ratio and the precision compression ratio adjustment coefficient to obtain a basic decomposition parameter; Calculate estimated performance information according to the basic decomposition parameters and the operator combination information, and iteratively optimize the basic decomposition parameters according to the estimated performance information to obtain preliminary decomposition parameters.

[0017] Optionally, the estimated performance information includes estimated throughput and estimated communication overhead. Correspondingly, the preliminary scheme determination module is specifically configured to: Substitute the estimated throughput and the estimated communication overhead into a joint optimization objective function to obtain an optimization objective difference value, wherein the joint optimization objective function takes maximizing the estimated throughput and minimizing the estimated communication overhead as joint optimization objectives. Calculate the gradient of the optimization objective difference value on the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameters respectively, and adjust the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameters according to the direction and size of the gradient. Repeat the above steps until the change rate of the optimization objective difference value is less than a preset change rate threshold to obtain preliminary decomposition parameters.

[0018] Optionally, the resource state information includes the calculation utilization rate, the bandwidth utilization rate and the local memory occupancy rate of the processing unit in the reconfigurable TPU computing device. Correspondingly, the decomposition parameter optimization module is specifically configured to: In the case that the calculation utilization rate exceeds a preset calculation utilization rate threshold, calculate a decomposition size optimization coefficient according to the calculation utilization rate and the preset calculation utilization rate threshold, and multiply the decomposition size optimization coefficient by the decomposition size to obtain an optimized decomposition size. In the case that the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, calculate an overlap ratio optimization coefficient according to the bandwidth utilization rate and the preset bandwidth utilization rate threshold, and multiply the overlap ratio optimization coefficient by the overlap ratio to obtain an optimized overlap ratio. In the case that the local memory occupancy rate exceeds a preset occupancy rate threshold, calculate a precision compression ratio optimization coefficient according to the local memory occupancy rate and the preset occupancy rate threshold, and multiply the precision compression ratio optimization coefficient by the precision compression ratio to obtain an optimized precision compression ratio.

[0019] In a third aspect, the present application provides an electronic device, which comprises a processor, a memory, and a program or instruction stored on the memory and executable on the processor, and the program or instruction is executed by the processor to implement the method of the first aspect.

[0020] In a fourth aspect, the present application provides a readable storage medium, which stores a program or instructions, and the program or instructions are executed by a processor to implement the method of the first aspect.

[0021] In the present application, training auxiliary driving data is acquired, and preliminary decomposition parameters are determined according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information includes modal type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters include decomposition size, overlap ratio and precision compression ratio; the training auxiliary driving data is decomposed according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters; the preliminary decomposition result is input into the auxiliary driving module, and resource state information of the reconfigurable TPU computing device is collected; the preliminary decomposition parameters are optimized according to the resource state information, and the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet a preset optimization stop condition, target decomposition parameters are obtained, and the actual auxiliary driving data is decomposed according to the target decomposition parameters. The above auxiliary driving data decomposition method of the reconfigurable TPU computing device is based on a dynamic decomposition strategy driven by data characteristics and computing operators, combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, realizes adaptive and efficient data decomposition accurately matched with data characteristics and inference requirements, and synchronously constructs a flexible hardware architecture that can cooperate with the decomposition parameters in real time, achieves power optimization and resource dynamic adaptation, and thus significantly improves inference efficiency and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of an auxiliary driving data decomposition method of a reconfigurable TPU computing device provided by the present application; Figure 2 is a flowchart of another auxiliary driving data decomposition method of a reconfigurable TPU computing device provided by the present application; Figure 3 is an example diagram of image data division provided by the present application; Figure 4 is a flowchart of another auxiliary driving data decomposition method of a reconfigurable TPU computing device provided by the present application; Figure 5 is a structural diagram of an auxiliary driving data decomposition system of a reconfigurable TPU computing device provided by the present application; Figure 6 is a structural diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0023] In order to make the purposes, technical solutions and advantages of the present application clearer, the specific embodiments of the present application are further described in detail below with reference to the drawings. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only parts related to the present application are shown in the drawings, but not all. Before discussing the example embodiments in more detail, it should be mentioned that some example embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be implemented in parallel, concurrently or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but can also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, etc.

[0024] The technical solutions in the embodiments of the present application will be described clearly below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.

[0025] The terms "first", "second", and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually a class, not limited to the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / ", generally indicates that the objects before and after are in an "or" relationship.

[0026] The auxiliary driving data decomposition method and system of the reconfigurable TPU computing device provided by the embodiments of the present application will be described in detail below with reference to the drawings, through specific embodiments and application scenarios.

[0027] First, the present application is applicable to the scene of pre-training and optimizing the data decomposition strategy used for auxiliary driving reasoning and the corresponding TPU hardware configuration parameters before the vehicle with auxiliary driving function is actually put into use. Based on the above use scenario, it can be understood that the execution subject of the present application can be a terminal device with computing capability, which is not limited here.

[0028] Figure 1is a flowchart of a method for assisting driving data decomposition of a reconfigurable TPU computing device provided by the present application. As shown in Figure 1 the specific steps include: S101, obtaining training assistance driving data, and determining preliminary decomposition parameters according to data attribute information of the training assistance driving data and operator combination information of an assistance driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information includes modality type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters include decomposition size, overlap ratio and precision compression ratio.

[0029] Wherein, the training assistance driving data can be environment perception data used for training the assistance driving module, and can include image data and at least one radar data; the assistance driving module can be a neural network model for realizing the assistance driving inference function. Specifically, the image data can be a color image, an infrared image or a grayscale image collected by a vehicle-mounted camera, containing visual information of environmental elements such as lane lines, traffic signal lights, pedestrians, non-motor vehicles, other vehicles and road signs; the radar data can be ranging and angle measurement data collected by a vehicle-mounted radar sensor for representing target distance, speed and azimuth, and can include millimeter wave radar data, laser radar data and ultrasonic radar data.

[0030] In one embodiment, the way of obtaining the training assistance driving data can adopt real environment perception data collected by real vehicle road test in real scene as the training assistance driving data, can also adopt target environment perception data screened in public data set as the training assistance driving data, and can also adopt simulation environment perception data generated by simulation platform as the training assistance driving data. Wherein, the real environment perception data is collected by image sensors, millimeter wave radar sensors, laser radar sensors and ultrasonic radar sensors on the real vehicle.

[0031] Wherein, the data attribute information of the training assistance driving data can be a core parameter set for representing data features of the training assistance driving data, and can include modality type information, perception priority information and continuous change rate information. Specifically, the modality type information can be a sensor modality category to which the training assistance driving data belongs, such as image modality, millimeter wave radar modality, laser radar modality and ultrasonic radar modality; the perception priority information can be an importance level of the training assistance driving data to the assistance driving inference; and the continuous change rate information can be a change amplitude of a target feature in unit time in the continuously collected training assistance driving data.

[0032] The reconfigurable TPU (Tensor Processing Unit) computing device can be a special neural network processor based on an array of configurable computing units and supporting dynamic reallocation of hardware resources.

[0033] The operator combination information of the ADAS module can be information describing the characteristics and interaction of each operator in the ADAS module, and can include the operator type, the total number of operators, and the execution dependency relationship between the operators.

[0034] The preliminary decomposition parameters can be a set of parameters preliminarily determined for decomposing the ADAS training data, and can include a decomposition size, an overlap ratio, and a precision compression ratio. Specifically, the decomposition size can be the spatial or temporal dimension size of the original data divided 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 block under the premise of maintaining the core features.

[0035] In one embodiment, according to the data attribute information of the ADAS training data and the operator combination information of the ADAS module deployed on the reconfigurable TPU computing device, the preliminary decomposition parameters can be determined by using a mapping rule library of the data attribute information, the operator combination information, and the preliminary decomposition parameters, querying the mapping rule library with the current data attribute information and the operator combination information as the query conditions, and obtaining the preliminary decomposition parameters. The mapping rule library of the data attribute information, the operator combination information, and the preliminary decomposition parameters can be a set of association rules of the preliminary decomposition parameters corresponding to different data attribute information and operator combination information, which are constructed based on a large amount of ADAS data and engineering experience. For example, when the data attribute information is image modal, high perception priority, and high continuous change rate, and the operator combination information is 60% of convolution operators, 25% of attention operators, and serial execution dependency, the mapped preliminary decomposition parameters are 400x300 pixels (decomposition size), 15% (overlap ratio), and 1 / 1 (precision compression ratio).

[0036] S102, decomposing the ADAS training 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.

[0037] The preliminary decomposition result can be a plurality of sub-data block sets obtained by segmenting and compressing the training assisted driving data according to the preliminary decomposition parameters. In an embodiment, the manner of obtaining the preliminary decomposition result by decomposing the training assisted driving data according to the preliminary decomposition parameters can include: dividing the training assisted driving data into first data blocks according to the decomposition size; expanding each first data block outward to obtain a second data block according to the overlap ratio; and compressing each second data block according to the precision compression ratio to obtain the preliminary decomposition result.

[0038] In an embodiment, the manner of dynamically reconfiguring the reconfigurable TPU computing device based on the preliminary decomposition parameters can include: establishing a mapping rule library of the preliminary decomposition parameters and the reconfigurable computing core allocation strategy, the on-chip storage unit partitioning strategy, and the interconnection bus bandwidth allocation strategy in advance; querying the mapping rule library with the current preliminary decomposition parameters as a query condition to obtain the corresponding reconfigurable computing core allocation strategy, the on-chip storage unit partitioning strategy, and the interconnection bus bandwidth allocation strategy; generating a reconfiguration instruction according to the reconfigurable computing core allocation strategy, the on-chip storage unit partitioning strategy, and the interconnection bus bandwidth allocation strategy; and issuing and verifying the reconfiguration instruction to complete the dynamic reconfiguration of the reconfigurable TPU computing device. The mapping rule library of the preliminary decomposition parameters and the reconfigurable computing core allocation strategy, the on-chip storage unit partitioning strategy, and the interconnection bus bandwidth allocation strategy can be constructed based on a large amount of hardware adaptation test data of the assisted driving data decomposition scene and the hardware performance parameters of the reconfigurable TPU computing device, and stores a set of associated rules of the reconfigurable computing core allocation strategy, the on-chip storage unit partitioning strategy, and the interconnection bus bandwidth allocation strategy corresponding to different preliminary decomposition parameters. For example, when the preliminary decomposition parameters are 400x300 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 arranged in a 2x4 array, the corresponding on-chip storage unit partitioning strategy is to divide 6 storage partitions each with a capacity of 16 MB, and the corresponding interconnection bus bandwidth allocation strategy is to allocate 8 Gbps exclusive bandwidth.

[0039] S103, inputting the preliminary decomposition result to the assisted driving module and collecting resource state information of the reconfigurable TPU computing device.

[0040] In an embodiment, the manner of inputting the preliminary decomposition result to the assisted driving module can include: determining the input priority of each sub-data block in the preliminary decomposition result according to the execution dependency relationship between the operators in the assisted driving module; and then transmitting each sub-data block in the preliminary decomposition result to the input layer of the assisted driving module in batches according to the hardware resource capacity of the reconfigurable TPU computing device for calculation and processing.

[0041] The resource state information of the reconfigurable TPU computing device can be information used to describe real-time hardware resource utilization of the reconfigurable TPU computing device in processing the preliminary decomposition result. The resource state information of the reconfigurable TPU computing device can include a computation utilization rate, a bandwidth utilization rate, and a local memory occupancy rate of a processing unit in the reconfigurable TPU computing device. Specifically, the processing unit can be a configurable computing core in the reconfigurable TPU computing device; the computation utilization rate can be a proportion of actual working time of the processing unit in a unit time; the bandwidth utilization rate can be a proportion of actual data transmission of an interconnection bus in a unit time to a maximum transmission capacity of the bus; and the local memory occupancy rate can be a proportion of used storage space in an on-chip storage unit to total storage space.

[0042] In an embodiment, the reconfigurable TPU computing device can collect the resource state information in real time by using a hardware performance monitoring unit (PMU) built in the reconfigurable TPU computing device to collect the computation utilization rate, the bandwidth utilization rate, and the local memory occupancy rate.

[0043] S104, optimizing the preliminary decomposition parameter according to the resource state information, and dynamically reconfiguring the reconfigurable TPU computing device until the preliminary decomposition parameter meets a preset optimization stop condition, to obtain a target decomposition parameter used for actually assisting driving data to be decomposed according to the target decomposition parameter.

[0044] The target decomposition parameter can be a decomposition parameter used for assisting driving inference in actual driving of a vehicle, that is, actual assisting driving data is collected by a vehicle-mounted sensor on the vehicle in actual driving of the vehicle, and the actual assisting driving data is decomposed according to the target decomposition parameter to obtain actual decomposition result, and finally the actual decomposition result is calculated and processed by an assisting driving module deployed on the reconfigurable TPU computing device.

[0045] In an embodiment, the reconfigurable TPU computing device can collect the resource state information in real time by using a hardware performance monitoring unit (PMU) built in the reconfigurable TPU computing device to collect the computation utilization rate, the bandwidth utilization rate, and the local memory occupancy rate.

[0042] In an embodiment, the reconfigurable TPU computing device can collect the resource state information in real time by using a hardware performance monitoring unit (PMU) built in the reconfigurable TPU computing device to collect the computation utilization rate, the bandwidth utilization rate, and the local memory occupancy rate.

[0043] S104, optimizing the preliminary decomposition parameter according to the resource state information, and dynamically reconfiguring the reconfigurable TPU computing device until the preliminary decomposition parameter meets a preset optimization stop condition, to obtain a target decomposition parameter used for actually assisting driving data to be decomposed according to the target decomposition parameter.

[0044] The target decomposition parameter can be a decomposition parameter used for assisting driving inference in actual driving of a vehicle, that is, actual assisting driving data is collected by a vehicle-mounted sensor on the vehicle in actual driving of the vehicle, and the actual assisting driving data is decomposed according to the target decomposition parameter to obtain actual decomposition result, and finally the actual decomposition result is calculated and processed by an assisting driving module deployed on the reconfigurable TPU computing device.

[0045] In an embodiment, the reconfigurable TPU computing device can collect the resource state information in real time by using a hardware performance monitoring unit (PMU) built in the reconfigurable TPU computing device to collect the computation utilization rate, the bandwidth utilization rate, and the local memory occupancy rate.

[0046] The preliminary decomposition parameter satisfies a preset optimization stop condition, which can be that the comprehensive evaluation score calculated according to the preliminary decomposition parameter after iterative optimization of resource state information is identified to exceed a preset evaluation score threshold for three consecutive times, can be that the gradient sizes of the decomposition size, the overlap ratio and the precision compression ratio are identified to be less than a preset gradient threshold, and can be that the number of iterative optimizations is identified to exceed a preset iteration number threshold. At this time, the preliminary decomposition parameter satisfying the preset optimization stop condition is the target decomposition parameter.

[0047] In an embodiment, the optimization of the preliminary decomposition parameter according to the resource state information comprises: in the case that the calculation utilization rate exceeds a preset calculation utilization rate threshold, calculating a decomposition size optimization coefficient according to the calculation utilization rate and the preset calculation utilization rate threshold, and multiplying the decomposition size optimization coefficient by the decomposition size to obtain the optimized decomposition size; in the case that the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, calculating an overlap ratio optimization coefficient according to 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; in the case that the local memory occupancy rate exceeds a preset occupancy rate threshold, calculating a precision compression ratio optimization coefficient according to 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 the optimized precision compression ratio.

[0048] The preset calculation utilization rate threshold can be a preset calculation utilization rate threshold value for judging whether the processing unit load in the reconfigurable TPU computing device is in a reasonable range, for example, 80%. Correspondingly, the calculation utilization rate exceeding the preset calculation utilization rate threshold value indicates that the processing unit load is too high, the processing pressure of the sub-data block corresponding to the current decomposition size exceeds the hardware carrying capacity, which is easy to cause the increase of data processing delay and affect the real-time performance of assisted driving inference. The decomposition size optimization coefficient can be an adjustment coefficient greater than 1 for reducing the number of sub-data blocks by increasing the decomposition size.

[0049] In an embodiment, the way of calculating the decomposition size optimization coefficient according to the calculation utilization rate and the preset calculation utilization rate threshold value, and multiplying the decomposition size optimization coefficient by the decomposition size to obtain the optimized decomposition size, can be to calculate the difference between the calculation utilization rate and the preset calculation utilization rate threshold value, divide the difference by the preset calculation utilization rate threshold value, add the division calculation result to 1 to obtain the decomposition size optimization coefficient, and multiply the decomposition size optimization coefficient by the current decomposition size to obtain the optimized decomposition size.

[0050] The preset bandwidth utilization threshold can be a preset bandwidth utilization critical value for judging whether the data transmission load of the interconnection bus in the reconfigurable TPU computing device is in a safe running interval, for example, 85%. Correspondingly, when the bandwidth utilization exceeds the preset bandwidth utilization threshold, it indicates that the data transmission load of the interconnection bus is saturated, the current overlap ratio causes repeated transmission in the overlap area of adjacent sub-data blocks, occupies too much bus bandwidth, causes data transmission congestion, and reduces the overall processing efficiency. The overlap ratio optimization coefficient can be an adjustment coefficient less than 1 for reducing the amount of repeated data transmission by reducing the overlap ratio.

[0051] In one embodiment, the manner of calculating the overlap ratio optimization coefficient according to the bandwidth utilization and the preset bandwidth utilization threshold, and multiplying the overlap ratio optimization coefficient by the overlap ratio to obtain the optimized overlap ratio can be as follows: calculating the difference between the bandwidth utilization and the preset bandwidth utilization threshold, dividing the difference by the preset bandwidth utilization threshold, subtracting the division calculation result from 1 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.

[0052] The preset occupancy threshold can be a preset occupancy critical value for judging whether the local memory resource occupancy in the reconfigurable TPU computing device is in a controllable range, for example, 85%. Correspondingly, when the local memory occupancy exceeds the preset occupancy threshold, it indicates that the available space of the local memory is insufficient, the storage demand of the sub-data block corresponding to the current precision compression ratio exceeds the upper limit of the storage resource, and data buffer overflow is easily caused, increasing the processing delay and energy consumption. The precision compression ratio optimization coefficient can be an adjustment coefficient less than 1 for reducing the storage space occupancy of the sub-data block by increasing the precision compression degree.

[0053] In one embodiment, the manner of calculating the precision compression ratio optimization coefficient according to the local memory occupancy and the preset occupancy threshold, and multiplying the precision compression ratio optimization coefficient by the precision compression ratio to obtain the optimized precision compression ratio can be as follows: calculating the difference between the local memory occupancy and the preset occupancy threshold, dividing the difference by the preset occupancy threshold, subtracting the division calculation result from 1 to obtain the precision compression ratio optimization coefficient, and multiplying the precision compression ratio optimization coefficient by the current precision compression ratio to obtain the optimized precision compression ratio.

[0054] Correspondingly, the preliminary decomposition parameter satisfying the preset optimization stop condition can be that the corresponding calculation utilization of the preliminary decomposition parameter does not exceed the preset calculation utilization threshold, and the corresponding bandwidth utilization does not exceed the preset bandwidth utilization threshold, and the corresponding local memory occupancy does not exceed the preset occupancy threshold.

[0055] The advantage of this scheme is that, by calculating and applying the optimization coefficient, the optimized decomposition parameters can be quickly dynamically adapted to the hardware resource state of the reconfigurable TPU computing device.

[0056] In the embodiments of the present application, training auxiliary driving data is obtained, and preliminary decomposition parameters are determined according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; the data attribute information includes modality type information, perception priority information, and continuous change rate information, and the preliminary decomposition parameters include decomposition size, overlap ratio, and precision compression ratio; the training auxiliary driving data is decomposed according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and the reconfigurable TPU computing device is dynamically reconfigured based on the preliminary decomposition parameters; the preliminary decomposition result is input into the auxiliary driving module, and resource state information of the reconfigurable TPU computing device is collected; the preliminary decomposition parameters are optimized according to the resource state information, and the reconfigurable TPU computing device is dynamically reconfigured until the preliminary decomposition parameters meet a preset optimization stop condition, target decomposition parameters are obtained, and are used for actual auxiliary driving data to be decomposed according to the target decomposition parameters. The auxiliary driving data decomposition method of the reconfigurable TPU computing device described above is based on a dynamic decomposition strategy driven by data characteristics and computing operators, combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, to realize adaptive and efficient data decomposition that accurately matches data characteristics and inference requirements, and simultaneously construct a flexible hardware architecture that can be real-time matched with the decomposition parameters, achieve power optimization and resource dynamic adaptation, and thus significantly improve inference efficiency and accuracy.

[0057] Figure 2 is another flowchart of an auxiliary driving data decomposition method of a reconfigurable TPU computing device provided by the present application. As shown in Figure 2 , the method specifically includes the following steps: S201, training auxiliary driving data is obtained, and preliminary decomposition parameters are determined according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; the data attribute information includes modality type information, perception priority information, and continuous change rate information, and the preliminary decomposition parameters include decomposition size, overlap ratio, and precision compression ratio.

[0058] S202, a decomposition moving step and a target decomposition size are determined according to the decomposition size and the overlap ratio, and the training auxiliary driving data is divided according to the decomposition moving step and the target decomposition size to obtain a basic data block.

[0059] The decomposition moving step can be a distance between starting positions of two adjacent basic data blocks when the training auxiliary driving data is divided by sliding.

[0060] In an embodiment, according to the manner of determining the decomposition moving step and the target decomposition size according to the decomposition size and the overlap ratio, the decomposition size, the overlap ratio and 0.5 can be multiplied to obtain the decomposition moving step, and the decomposition size and double the overlap ratio can be added to obtain the target decomposition size. For example, the preliminary decomposition parameter is 400*300 pixels (decomposition size) and 10% (overlap ratio), the corresponding decomposition moving step is 20 pixels (lateral) and 15 pixels (vertical), and the target decomposition size is 440*330 pixels.

[0061] The basic data block can be an original feature data block obtained by slidingly dividing the training auxiliary driving data according to the target decomposition size and the decomposition moving step, and keeping a preset overlap region.

[0062] In an embodiment, according to the manner of dividing the training auxiliary driving data according to the decomposition moving step and the target decomposition size to obtain the basic data block, the starting position of the training auxiliary driving data can be taken as the origin to construct a sliding window consistent with the target decomposition size, the sliding window can be sequentially slid along the spatial or temporal dimension of the training auxiliary driving data according to the decomposition moving step, a data set in the sliding window can be intercepted as a basic data block after each sliding, and all basic data blocks can be obtained until the sliding window covers all the training auxiliary driving data.

[0063] Figure 3 is an example of dividing image data provided by the present application. As shown in Figure 3 the size of the gray area is equal to the decomposition size (400*300 pixels) in the preliminary decomposition parameter, the dashed area is a sliding window, the size of the sliding window is equal to the target decomposition size (440*330 pixels), and the decomposition moving step is 20 pixels (lateral) and 15 pixels (vertical).

[0064] S203, reconstructively encoding the basic data block according to the precision compression ratio to obtain a compressed data block; wherein the reconstructively encoding includes quantization encoding and sparsification encoding.

[0065] The reconstruction coding can be a coding mode based on the precision compression ratio, and can include quantization coding and sparsification coding. Specifically, the quantization coding can be a coding mode of mapping high-precision original data (such as FP32 floating-point type) in the basic data block to low-precision data (such as FP16, INT8); and the sparsification coding can be a coding mode of eliminating redundant data (such as background noise, invalid point cloud) in the basic data block that is meaningless for the AD inference, and only retaining core effective data (such as target contour pixels, valid ranging data).

[0066] The compressed data block can be a low-precision and low-redundancy data block obtained by performing reconstruction coding on the basic data block. In an embodiment, the reconstruction coding on the basic data block according to the precision compression ratio to obtain the compressed data block can be performed in the following manner: determining a target quantization strategy according to the current precision compression ratio and a mapping relationship between the precision compression ratio and the quantization strategy (for example, the quantization strategy corresponding to the precision compression ratio of 1 / 2 is FP32 to FP16, and the quantization strategy corresponding to the precision compression ratio of 1 / 4 is FP32 to INT8), performing linear quantization coding on the basic data block according to the target quantization strategy to obtain quantized data, calculating a sparsification threshold (for example, data variance) corresponding to the quantized data, performing sparsification coding on the quantized data according to the sparsification threshold to retain core effective data, and performing variable-length coding on the core effective data to obtain the compressed data block.

[0067] S204, dynamically reconstructing the reconfigurable TPU computing device based on the preliminary decomposition parameter.

[0068] In an embodiment, after the reconstruction coding on the basic data block according to the precision compression ratio to obtain the compressed data block, the method further includes: determining the calculation dependency relationship information between the compressed data blocks according to the operator combination information.

[0069] The calculation dependency relationship information between the compressed data blocks can be information representing the processing order and the parallel / serial execution relationship of different compressed data blocks in the AD module operator execution process.

[0070] In an embodiment, the determination of the calculation dependency relationship information between the compressed data blocks according to the operator combination information can be performed in the following manner: matching each compressed data block to a suitable operator according to the attribute information of the compressed data block and the attribute information of each operator, determining the dependency type and the processing order of the compressed data blocks of different operators according to the execution dependency relationship between the operators, and inferring the parallel / serial execution relationship of the compressed data blocks according to the dependency type and the processing order.

[0071] Correspondingly, in an embodiment, the dynamic reconfiguration of the reconfigurable TPU computing device based on the preliminary decomposition parameters comprises: configuring the data bit width, the calculation mode and the local memory capacity of the processing unit in the reconfigurable TPU computing device according to the attribute information of the compressed data block; allocating the compressed data block to the processing unit according to the calculation dependency information to obtain the mapping relationship between the processing unit and the compressed data block; determining the data communication requirement information between the processing units according to the calculation dependency information and the mapping relationship, and configuring the connection topology relationship of the on-chip interconnection network and the buffer depth between the processing units in the reconfigurable TPU computing device according to the data communication requirement information.

[0072] The attribute information of the compressed data block can be information for representing the core features and processing requirements of the compressed data block, and can include data modal type, precision level, data size, perception priority and target operator type, etc.

[0073] The data bit width can be the number of binary bits of a single data that can be processed by the processing unit; 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 on-chip storage resource size allocated to the processing unit.

[0074] In an embodiment, the manner of configuring the data bit width, the calculation mode and the local memory capacity of the processing unit in the reconfigurable TPU computing device according to the attribute information of the compressed data block can adopt a mapping rule library of the attribute information of the compressed data block, the data bit width, the calculation mode and the local memory capacity, use the attribute information of the current compressed data block as a query condition to query the mapping rule library, and obtain the corresponding data bit width, calculation mode and local memory capacity. The mapping rule library of the attribute information of the compressed data block, the data bit width, the calculation mode and the local memory capacity can be a set of associated rules of the data bit width, the calculation mode and the local memory capacity corresponding to different attribute information of the compressed data block, which is constructed based on a large amount of hardware adaptation test data of the auxiliary driving data decomposition scene and the hardware performance parameters of the reconfigurable TPU computing device. For example, when the attribute information of the compressed data block is image modal (data modal 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.

[0075] The mapping relationship between the processing units and the compressed data blocks can represent one-to-one correspondence or one-to-many allocation relationship between the processing units and the compressed data blocks in the reconfigurable TPU computing device. In an embodiment, the manner of allocating the compressed data blocks to the processing units according to the computation dependency relationship information to obtain the mapping relationship between the processing units and the compressed data blocks can include identifying parallel dependency compressed data block groups (a set of compressed data blocks without sequence) and serial dependency links (a sequence of compressed data blocks with sequence) in the computation dependency relationship information, then allocating independent processing units (or processing unit groups) to each parallel dependency group according to the number of processing units of the reconfigurable TPU computing device, and sequentially allocating processing units (or multiplexing previous processing units) to each compressed data block in the serial dependency link according to the execution order, and finally recording the corresponding compressed data blocks of each processing unit to form a mapping relationship table.

[0076] The data communication requirement information between the processing units can represent the data communication process between the processing units with data interaction, and can include communication direction and communication data volume. In an embodiment, the manner of determining the data communication requirement information between the processing units according to the computation dependency relationship information and the mapping relationship can include extracting compressed data block pairs with data interaction from the computation dependency relationship information, then finding the corresponding processing units of each compressed data block pair through the mapping relationship and determining the communication direction and communication data volume between the two processing units, and finally summarizing to form the data communication requirement information.

[0077] The on-chip interconnection network can be a hardware network architecture connecting the processing units in the reconfigurable TPU computing device to realize data transmission between the processing units; the connection topology relationship can be the connection mode and link configuration of the processing units in the on-chip interconnection network; and the buffer depth can be the storage capacity of the buffer in each communication link in the on-chip interconnection network.

[0078] In an embodiment, the manner of configuring the connection topology relationship and the buffer depth of the on-chip interconnection network between the processing units in the reconfigurable TPU computing device according to the data communication requirement information can include first dividing the data communication requirement information into three levels of high, medium and low according to real-time requirements, configuring a point-to-point direct connection topology relationship and allocating a large-capacity buffer (for example, a buffer depth of 16 MB) for high-level data communication requirement information, configuring a shared crossbar topology relationship and allocating a medium-capacity buffer (for example, a buffer depth of 8 MB) for medium-level data communication requirement information, and configuring a bus-type topology relationship and allocating a small-capacity buffer (for example, a buffer depth of 4 MB) for low-level data communication requirement information, and issuing configuration instructions through the interconnection network controller of the reconfigurable TPU computing device to complete topology relationship switching and buffer capacity allocation.

[0079] In an embodiment, after the data communication requirement information between the processing units is determined according to the calculation dependency information and the mapping relationship, the method further comprises: generating parallel data flow arrangement information according to the mapping relationship and the calculation dependency information; and generating the synchronous processing time of the processing units according to the parallel data flow arrangement information, so as to control the processing units to process the corresponding compressed data blocks according to the synchronous processing time in the model training process, so as to realize the synchronous processing of the compressed data blocks.

[0080] The parallel data flow arrangement information can be structured information of a multi-level parallel group of processing units based on the calculation dependency relationship, and can comprise a parallel group division scheme of each level of processing units and a data flow direction between each processing unit in each parallel group of processing units.

[0081] In an embodiment, the parallel data flow arrangement information can be generated according to the mapping relationship and the calculation dependency information by dividing the parallel units without mutual dependency in the calculation dependency information (including independent compressed data blocks and non-disassembled serial dependency links) into parallel groups at different levels, and determining the processing units corresponding to each compressed data block in the parallel group and the data flow direction between each processing unit according to the mapping relationship, i.e., the parallel data flow arrangement information.

[0082] The synchronous processing time of the processing units can be a start time set for each processing unit so that the next processing unit can simultaneously receive the processing result data output by all previous processing units. Accordingly, the processing units can start processing the corresponding compressed data blocks at the corresponding synchronous processing time in the model training process, and if a processing unit completes processing in advance, the processing unit enters a synchronous waiting state until all compressed data blocks in the parallel group are processed, and then the processing units of the next level of parallel groups are triggered to start simultaneously.

[0083] In an embodiment, the synchronous processing time of the processing units can be generated according to the parallel data flow arrangement information by calculating the estimated calculation processing time of each processing unit according to the attribute information of each processing unit and the attribute information of the corresponding compressed data block in the parallel data flow arrangement information, and determining the synchronous processing time of the processing units according to the estimated calculation processing time of each processing unit and the data flow direction between each processing unit.

[0084] The advantage of this arrangement is that the precise planning of the synchronous processing time realizes the progress alignment of the processing units in the parallel group, and improves the utilization rate of parallel computing resources.

[0085] The advantage of this scheme is that, based on the hardware configuration of the compressed data block attribute information and the calculation dependency information, the operation capacity of the processing unit, the storage resource and the data processing demand are accurately matched, the resource waste is avoided, the interconnection network topology and the buffer depth are dynamically configured according to the communication demand, the efficiency and real-time performance of data transmission between processing units are guaranteed, and communication congestion is reduced.

[0086] S205, input the compressed data block to the auxiliary driving module, and collect the resource state information of the reconfigurable TPU computing device.

[0087] In one embodiment, the way of inputting the compressed data block to the auxiliary driving module can adopt to determine the input priority of each compressed data block according to the execution dependency relationship between each operator in the auxiliary driving module, and then transmit the compressed data block to the input layer of the auxiliary driving module in batches according to the hardware resource capacity of the reconfigurable TPU computing device for calculation processing.

[0088] S206, optimizing the preliminary decomposition parameters according to the resource state information, and dynamically reconfiguring the reconfigurable TPU computing device until the preliminary decomposition parameters meet the preset optimization stop condition, obtaining the target decomposition parameters for actually decomposing the auxiliary driving data according to the target decomposition parameters.

[0089] The advantage of this scheme is that it can ensure that the basic data block not only completely covers the original training auxiliary driving data, but also retains the spatio-temporal continuity of the data through the overlapping area, avoids feature rupture caused by division, and realizes accurate compression of data by using the reconstruction method combining quantization coding and sparsification coding, reduces the storage and transmission pressure on the premise of meeting the accuracy requirement.

[0090] Figure 4 is a flowchart of another auxiliary driving data decomposition method of a reconfigurable TPU computing device provided by the present application. As shown in Figure 4 , the specific steps include: S401, obtaining training auxiliary driving data.

[0091] S402, assigning a basic decomposition size, a basic overlap ratio and a basic accuracy compression ratio according to the modal type information of the training auxiliary driving data and the operator calculation complexity in the operator combination information.

[0092] Wherein, the operator calculation complexity can be a quantitative index representing the amount of calculation, storage access times and data transmission amount required for a single operator or operator combination in the auxiliary driving module to execute an operation once, which can be FLOP (floating point operation times).

[0093] The reference resolution size can be a resolution size preset according to the modal type information of the training assisted driving data and the operator calculation complexity; the reference overlap ratio can be a overlap ratio preset according to the modal type information of the training assisted driving data and the operator calculation complexity; and the reference precision compression ratio can be a precision compression ratio preset according to the modal type information of the training assisted driving data and the operator calculation complexity.

[0094] In an embodiment, according to the way of assigning the reference resolution size, the reference overlap ratio and the reference precision compression ratio according to the modal type information of the training assisted driving data and the operator calculation complexity in the operator combination information, a mapping rule library of the modal type information, the operator calculation complexity, the reference resolution size, the reference overlap ratio and the reference precision compression ratio can be pre-established, the mapping rule library is queried to obtain the reference resolution size, the reference overlap ratio and the reference precision compression ratio by taking the current modal type information and the operator calculation complexity as a query condition. The mapping rule library of the modal type information, the operator calculation complexity, the reference resolution size, the reference overlap ratio and the reference precision compression ratio can be a set of association rules of the reference resolution size, the reference overlap ratio and the reference precision compression ratio corresponding to different modal type information and operator calculation complexity, which is constructed based on a large amount of measured data of assisted driving scenes, average operation efficiency test results of operators on TPU and engineering practical experience, for example: when the modal type information is image modal and the operator calculation complexity is FLOP, the reference resolution size is 600x450 pixels, the reference overlap ratio is 12%, and the reference precision compression ratio is 1 / 2.

[0095] S403, determining a resolution size adjustment coefficient, an overlap ratio adjustment coefficient and a precision compression ratio adjustment coefficient based on the perception priority information and the continuous change rate information of the training assisted driving data.

[0096] The resolution size adjustment coefficient can be a coefficient for correcting the reference resolution size determined based on the perception priority information and the continuous change rate information of the training assisted driving data; the overlap ratio adjustment coefficient can be a coefficient for correcting the reference overlap ratio determined based on the perception priority information and the continuous change rate information of the training assisted driving data; and the precision compression ratio adjustment coefficient can be a coefficient for correcting the reference precision compression ratio determined based on the perception priority information and the continuous change rate information of the training assisted driving data.

[0097] In an embodiment, based on the manner of determining the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient according to the perception priority information and the continuous change rate information of the training assisted driving data, a mapping rule library of the perception priority information, the continuous change rate information, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient can be pre-established, and the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient can be obtained by querying the mapping rule library with the current perception priority information and the continuous change rate information as the query condition. The mapping rule library of the perception priority information, the continuous change rate information, the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient can be a set of associated rules of the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient corresponding to different perception priority information and continuous change rate information, which are constructed based on the perception importance demand of the assisted driving data and the influence degree of the data continuous change on the processing precision, for example: when the perception priority information is high perception priority and the continuous change rate information is fast (the inter-frame feature change amount is greater than or equal to 30%), the decomposition size adjustment coefficient is 0.8, the overlap ratio adjustment coefficient is 1.2, and the precision compression ratio adjustment coefficient is 1.2.

[0098] S404, multiplying the reference decomposition size and the decomposition size adjustment coefficient, multiplying the reference overlap ratio and the overlap ratio adjustment coefficient, and multiplying the reference precision compression ratio and the precision compression ratio adjustment coefficient to obtain a basic decomposition parameter.

[0099] The basic decomposition parameter includes a basic decomposition size, a basic overlap ratio, and a basic precision compression ratio.

[0100] S405, calculating estimated performance information according to the basic decomposition parameter and the operator combination information, and iteratively optimizing the basic decomposition parameter according to the estimated performance information to obtain a preliminary decomposition parameter. The preliminary decomposition parameter includes a decomposition size, an overlap ratio, and a precision compression ratio.

[0101] The estimated performance information can be a performance index of the auxiliary driving module processing the training auxiliary driving data, and can include an estimated processing delay and an estimated data transmission bandwidth. In an embodiment, the estimated performance information can be calculated according to the base decomposition parameters and the operator combination information, and the calculation method can include calculating the number of data blocks, the average single data block calculation amount, and the average single data block memory access amount according to the operator calculation complexity of each operator in the operator combination information, the base decomposition size, and the base precision compression ratio, obtaining the average value of the peak computing power of each processing unit of the reconfigurable TPU computing device, dividing the average single data block calculation amount by the average value and multiplying the result by the number of data blocks to obtain the estimated processing delay, and multiplying the average single data block memory access amount by the number of data blocks and dividing the result by the estimated processing delay to obtain the estimated data transmission bandwidth.

[0102] In an embodiment, the base decomposition parameters can be iteratively optimized according to the estimated performance information to obtain preliminary decomposition parameters, and the optimization method can include increasing the base precision compression ratio by a first preset step size and reducing the base decomposition size by a second preset step size when the estimated processing delay exceeds a preset processing delay threshold, and reducing the base overlap ratio by a third preset step size when the estimated data transmission bandwidth exceeds a preset data transmission bandwidth threshold.

[0103] The estimated performance information includes an estimated throughput and an estimated communication overhead. Accordingly, in an embodiment, the base decomposition parameters can be iteratively optimized according to the estimated performance information to obtain preliminary decomposition parameters, and the optimization method can include substituting the estimated throughput and the estimated communication overhead into a joint optimization objective function to obtain an optimization objective difference value, wherein the joint optimization objective function has a joint optimization objective of maximizing the estimated throughput and minimizing the estimated communication overhead, calculating the gradient of the optimization objective difference value with respect to the decomposition size, the overlap ratio, and the precision compression ratio in the base decomposition parameters, and adjusting the decomposition size, the overlap ratio, and the precision compression ratio in the base decomposition parameters according to the direction and size of the gradient, and repeating the above steps until the change rate of the optimization objective difference value is less than a preset change rate threshold to obtain the preliminary decomposition parameters.

[0104] The estimated throughput can be the total amount of data that can be processed per unit time, and the estimated communication overhead can be an additional data transmission cost generated due to data blocking, precision conversion, and intermediate result exchange, which 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, and has a joint optimization objective of maximizing the estimated throughput and minimizing the estimated communication overhead. Specifically, the joint optimization objective function can be: ; wherein, and The preset weight coefficient is a preset weight coefficient.

[0105] The gradient can be a partial derivative of the optimization target difference value with respect to the basic decomposition size, the basic overlap ratio, or the basic precision compression ratio. In an embodiment, the manner of calculating the gradient of the optimization target difference value with respect to the decomposition size, the overlap ratio, and the precision compression ratio in the basic decomposition parameter can be to calculate the gradient of the optimization target difference value with respect to the basic decomposition size, the basic overlap ratio, and the basic precision compression ratio in the basic decomposition parameter by using a finite difference method. The finite difference method can be an approximate method for calculating the partial derivative of a target function with respect to a parameter in numerical calculation, and the core idea is to estimate the gradient by using the change in the value of the target function caused by a small change in the value of the parameter.

[0106] In an embodiment, the manner of adjusting the decomposition size, the overlap ratio, and the precision compression ratio in the basic decomposition parameter according to the direction and size of the gradient can be to obtain a preset basic learning rate, multiply the preset basic learning rate by the size of the gradient corresponding to the basic decomposition size, add the multiplication result to the basic decomposition size to obtain an optimized basic decomposition size in the case where the direction of the gradient is positive, subtract the multiplication result from the basic decomposition size to obtain an optimized basic decomposition size in the case where the direction of the gradient is negative, and the like.

[0107] The change rate of the optimization target difference value can be an average change value of the optimization target difference value in a plurality of continuous iterations, and the preset change rate threshold can be a small positive number as a convergence determination condition.

[0108] The change rate of the optimization target difference value being less than the preset change rate threshold indicates that the optimization target difference value has tended to be stable, the small adjustment of the parameter has no significant effect on the performance improvement, that is, the optimization process has reached or is close to a local optimal solution under the current constraint, and the iteration can be terminated.

[0109] The advantage of this scheme is that the iterative optimization with the joint optimization target of maximizing the estimated throughput and minimizing the estimated communication overhead can accurately balance the calculation efficiency and the data transmission cost, and avoid the performance imbalance caused by single-index optimization.

[0110] S406, decompose the training auxiliary driving data according to the preliminary decomposition parameter to obtain a preliminary decomposition result, and dynamically reconfigure the reconfigurable TPU computing device based on the preliminary decomposition parameter.

[0111] S407, input the preliminary decomposition result to the auxiliary driving module, and collect resource state information of the reconfigurable TPU computing device.

[0112] S408, optimizing the preliminary decomposition parameter according to the resource state information, and dynamically reconstructing the reconfigurable TPU computing device until the preliminary decomposition parameter meets a preset optimization stop condition, to obtain a target decomposition parameter for actual auxiliary driving data to be decomposed according to the target decomposition parameter.

[0113] The advantage of this scheme is that the accurate matching of the decomposition parameter and the data characteristics of the training auxiliary driving data is realized through the double-layer parameter determination logic of the reference parameter and the adjustment coefficient, and the performance defects of the basic decomposition parameter are avoided in advance through the performance simulation and iterative optimization, thereby reducing the trial-and-error cost of subsequent hardware reconstruction.

[0114] Figure 5 is a structural schematic diagram of an auxiliary driving data decomposition system of a reconfigurable TPU computing device provided by the present application. As shown in Figure 5 The system comprises: A preliminary scheme determination module 510 is configured to obtain training auxiliary driving data, and determine a preliminary decomposition parameter according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information comprises modality type information, perception priority information and continuous change rate information, and the preliminary decomposition parameter comprises a decomposition size, an overlap ratio and a precision compression ratio. A TPU reconstruction module 520 is configured to decompose the training auxiliary driving data according to the preliminary decomposition parameter to obtain a preliminary decomposition result, and dynamically reconstruct the reconfigurable TPU computing device based on the preliminary decomposition parameter. A decomposition result input module 530 is configured to input the preliminary decomposition result to the auxiliary driving module, and collect resource state information of the reconfigurable TPU computing device. A decomposition parameter optimization module 540 is configured to optimize the preliminary decomposition parameter according to the resource state information, and dynamically reconstruct the reconfigurable TPU computing device until the preliminary decomposition parameter meets a preset optimization stop condition, to obtain a target decomposition parameter for actual auxiliary driving data to be decomposed according to the target decomposition parameter.

[0115] Optionally, the TPU reconstruction module 520 is specifically configured to: determine a decomposition moving step and a target decomposition size according to the decomposition size and the overlap ratio, and divide the training auxiliary driving data according to the decomposition moving step and the target decomposition size to obtain a basic data block; reconstruct and encode the basic data block according to the precision compression ratio to obtain a compressed data block; wherein the reconstruction and encoding comprises quantization encoding and sparsification encoding. Correspondingly, the decomposition result input module 530 is specifically configured to: input the compressed data block to the auxiliary driving module.

[0116] Optionally, the TPU reconstruction module 520 is further configured to: determine the calculation dependency relationship information between the compressed data blocks according to the operator combination information; configure the data bit width, the calculation mode, and the local memory capacity of the processing unit in the reconfigurable TPU computing device according to the attribute information of the compressed data block; allocate the compressed data blocks to the processing unit according to the calculation dependency relationship information, to obtain the mapping relationship between the processing unit and the compressed data block; determine the data communication requirement information between the processing units according to the calculation dependency relationship information and the mapping relationship, and configure the connection topology relationship and the buffer depth of the on-chip interconnection network between the processing units in the reconfigurable TPU computing device according to the data communication requirement information.

[0117] Optionally, the TPU reconstruction module 520 is further configured to: generate parallel data stream arrangement information according to the mapping relationship and the calculation dependency relationship information; generate the synchronization processing time of the processing unit according to the parallel data stream arrangement information, so as to control the processing unit to process the corresponding compressed data block according to the synchronization processing time in the model training process, to realize the synchronization processing of the compressed data block.

[0118] Optionally, the preliminary scheme determination module 510 is specifically configured to: allocate the decomposition size, the overlap ratio, and the precision compression ratio according to the modal type information of the training auxiliary driving data and the operator calculation complexity in the operator combination information; determine the decomposition size adjustment coefficient, the overlap ratio adjustment coefficient, and the precision compression ratio adjustment coefficient based on the perception priority information of the training auxiliary driving data and the continuous change rate information; multiply the decomposition size adjustment coefficient by the reference decomposition size, multiply the overlap ratio adjustment coefficient by the reference overlap ratio, and multiply the precision compression ratio adjustment coefficient by the reference precision compression ratio, to obtain the basic decomposition parameter; calculate the estimated performance information according to the basic decomposition parameter and the operator combination information, and iteratively optimize the basic decomposition parameter according to the estimated performance information, to obtain the preliminary decomposition parameter.

[0119] Optionally, the estimated performance information comprises an estimated throughput and an estimated communication overhead. Correspondingly, the preliminary scheme determination module 510 is specifically configured to: substitute the estimated throughput and the estimated communication overhead into a joint optimization objective function to obtain an optimization objective difference value, wherein the joint optimization objective function takes maximizing the estimated throughput and minimizing the estimated communication overhead as joint optimization objectives; respectively calculate gradients of the optimization objective difference value with respect to a decomposition size, an overlap ratio and an accuracy compression ratio in the basic decomposition parameter, and adjust the decomposition size, the overlap ratio and the accuracy compression ratio in the basic decomposition parameter according to directions and sizes of the gradients; repeat the above steps until a change rate of the optimization objective difference value is less than a preset change rate threshold, and obtain a preliminary decomposition parameter.

[0120] Optionally, the resource state information comprises a computation utilization rate, a bandwidth utilization rate and a local memory occupancy rate of a processing unit in the reconfigurable TPU computing device. Correspondingly, the decomposition parameter optimization module 540 is specifically configured to: in a case where the computation utilization rate exceeds a preset computation utilization rate threshold, calculate a decomposition size optimization coefficient according to the computation utilization rate and the preset computation utilization rate threshold, and multiply the decomposition size optimization coefficient by the decomposition size to obtain an optimized decomposition size; in a case where the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, calculate an overlap ratio optimization coefficient according to the bandwidth utilization rate and the preset bandwidth utilization rate threshold, and multiply the overlap ratio optimization coefficient by the overlap ratio to obtain an optimized overlap ratio; in a case where the local memory occupancy rate exceeds a preset occupancy rate threshold, calculate an accuracy compression ratio optimization coefficient according to the local memory occupancy rate and the preset occupancy rate threshold, and multiply the accuracy compression ratio optimization coefficient by the accuracy compression ratio to obtain an optimized accuracy compression ratio.

[0121] In the embodiment of the present application, the preliminary scheme determination module is configured to obtain training auxiliary driving data, and determine preliminary decomposition parameters according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information comprises modal type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters comprise decomposition size, overlap ratio and precision compression ratio; the TPU reconfiguration module is configured to decompose the training auxiliary driving data according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and dynamically reconfigure the reconfigurable TPU computing device based on the preliminary decomposition parameters; the decomposition result input module is configured to input the preliminary decomposition result to the auxiliary driving module, and collect resource state information of the reconfigurable TPU computing device; the decomposition parameter optimization module is configured to optimize the preliminary decomposition parameters according to the resource state information, and dynamically reconfigure the reconfigurable TPU computing device until the preliminary decomposition parameters meet a preset optimization stop condition, to obtain target decomposition parameters for decomposing actual auxiliary driving data according to the target decomposition parameters. The auxiliary driving data decomposition system of the reconfigurable TPU computing device described above, based on a dynamic decomposition strategy driven by data characteristics and computing operators, combined with the hardware dynamic reconfiguration capability of the reconfigurable TPU, realizes adaptive and efficient data decomposition accurately matched with data characteristics and inference requirements, and simultaneously constructs a flexible hardware architecture that can be real-time matched with the decomposition parameters, achieves power optimization and resource dynamic adaptation, thereby significantly improving inference efficiency and accuracy.

[0122] The auxiliary driving data decomposition system of the reconfigurable TPU computing device provided by the embodiments of the present application can realize the processes realized by the above-mentioned embodiments. To avoid repetition, details are not described here.

[0123] Figure 6 is a structural schematic diagram of an electronic device provided by the present application. As shown in Figure 6 The embodiments of the present application further provide an electronic device 600, which comprises a processor 601, a memory 602, a program or instruction stored in the memory 602 and executable on the processor 601. The program or instruction is executed by the processor 601 to realize the processes of the above-mentioned auxiliary driving data decomposition method embodiments of the reconfigurable TPU computing device and achieve the same technical effects. To avoid repetition, details are not described here.

[0124] It should be noted that the electronic device in the embodiments of the present application includes the above-mentioned mobile electronic device and non-mobile electronic device.

[0125] The embodiment of the application further provides a readable storage medium, which stores a program or instructions, and the program or instructions are executed by a processor to realize each process of the auxiliary driving data decomposition method of the reconfigurable TPU computing device and achieve the same technical effects. To avoid repetition, details are not described herein.

[0126] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0127] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiments of the application is not limited to the order of functions shown or discussed, and can also include functions performed in a substantially simultaneous manner or in reverse order, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0128] From the above description of the embodiments, those skilled in the art can clearly understand that the above embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the application can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a plurality of instructions for making a terminal (which can be a mobile phone, computer, server, or network equipment, etc.) execute the method described in each embodiment of the application.

[0129] The embodiments of the present application are described above with reference to the drawings, but the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative, but not restrictive, and those of ordinary skill in the art can make many forms without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, which all belong to the protection of the present application.

[0130] The above are only the preferred embodiments of the present application and the technical principles used. The present application is not limited to the specific embodiments described herein, and various obvious changes, readjustments and substitutions made by those skilled in the art without departing from the scope of the present application do not fall within the scope of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and more other equivalent embodiments can be included without departing from the concept of the present application, and the scope of the present application is determined by the scope of the claims.

Claims

1. An assisted driving data decomposition method of a reconfigurable TPU computing device, characterized in that, The method comprises: acquiring training auxiliary driving data, and determining preliminary decomposition parameters according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information comprises modal type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters comprise decomposition size, overlap ratio and precision compression ratio; decomposing the training auxiliary driving data according to the preliminary decomposition parameters to obtain preliminary decomposition results, and dynamically reconfiguring the reconfigurable TPU computing device based on the preliminary decomposition parameters; inputting the preliminary decomposition results into the auxiliary driving module, and collecting resource state information of the reconfigurable TPU computing device; optimizing the preliminary decomposition parameters according to the resource state information, and dynamically reconfiguring the reconfigurable TPU computing device until the preliminary decomposition parameters meet a preset optimization stop condition, to obtain target decomposition parameters for decomposing actual auxiliary driving data according to the target decomposition parameters. 2.The method of claim 1, wherein, The method further comprises: determining decomposition moving step and target decomposition size according to the decomposition size and the overlap ratio, and dividing the training auxiliary driving data according to the decomposition moving step and the target decomposition size to obtain basic data blocks; reconfiguring and encoding the basic data blocks according to the precision compression ratio to obtain compressed data blocks; wherein the reconfiguring and encoding comprises quantization encoding and sparsification encoding; correspondingly, the method further comprises: inputting the compressed data blocks into the auxiliary driving module. 3.The method of claim 2, wherein, After the reconfiguring and encoding of the basic data blocks according to the precision compression ratio to obtain compressed data blocks, the method further comprises: determining calculation dependency relationship information between the compressed data blocks according to the operator combination information; correspondingly, the method further comprises: configuring data bit width, calculation mode and local memory capacity of processing units in the reconfigurable TPU computing device according to attribute information of the compressed data blocks; allocating compressed data blocks to the processing units according to the calculation dependency relationship information to obtain mapping relationship between the processing units and the compressed data blocks; determining data communication demand information between the processing units according to the calculation dependency relationship information and the mapping relationship, and configuring connection topology relationship and buffer depth of an on-chip interconnection network between the processing units in the reconfigurable TPU computing device according to the data communication demand information. 4.The method of claim 3, wherein, After the determination of the data communication demand information between the processing units according to the calculation dependency relationship information and the mapping relationship, the method further comprises: generating parallel data stream arrangement information according to the mapping relationship and the calculation dependency relationship information; and determining the data communication demand information between the processing units according to the calculation dependency relationship information and the mapping relationship. According to the parallel data stream arrangement information, a synchronous processing time of the processing unit is generated, and the processing unit is controlled to process the corresponding compressed data block according to the synchronous processing time in a model training process, so as to realize synchronous processing of the compressed data block. 5.The method of claim 1, wherein, The preliminary decomposition parameter is determined according to the data attribute information of the training auxiliary driving data and the operator combination information of the auxiliary driving module deployed on the reconfigurable TPU computing device, and includes: According to the modal type information of the training auxiliary driving data and the operator calculation complexity in the operator combination information, a base decomposition size, a base overlap ratio and a base precision compression ratio are allocated; Based on the perception priority information of the training auxiliary driving data and the continuous change rate information, a decomposition size adjustment coefficient, an overlap ratio adjustment coefficient and a precision compression ratio adjustment coefficient are determined; The base decomposition size and the decomposition size adjustment coefficient are multiplied, the base overlap ratio and the overlap ratio adjustment coefficient are multiplied, and the base precision compression ratio and the precision compression ratio adjustment coefficient are multiplied to obtain a basic decomposition parameter; According to the basic decomposition parameter and the operator combination information, estimated performance information is calculated, and the basic decomposition parameter is iteratively optimized according to the estimated performance information to obtain a preliminary decomposition parameter. 6.The method of claim 5, wherein, The estimated performance information includes estimated throughput and estimated communication overhead; Correspondingly, the basic decomposition parameter is iteratively optimized according to the estimated performance information to obtain a preliminary decomposition parameter, including: The estimated throughput and the estimated communication overhead are substituted into a joint optimization objective function to obtain an optimization target difference value; wherein the joint optimization objective function maximizes the estimated throughput and minimizes the estimated communication overhead as a joint optimization objective; The gradients of the optimization target difference value with respect to the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameter are calculated respectively, and the decomposition size, the overlap ratio and the precision compression ratio in the basic decomposition parameter are adjusted according to the directions and sizes of the gradients; The above steps are repeatedly executed until the change rate of the optimization target difference value is less than a preset change rate threshold, and a preliminary decomposition parameter is obtained.

7. The method of claim 1, wherein the method further comprises: The resource state information includes the calculation utilization rate, the bandwidth utilization rate and the local memory occupancy rate of the processing unit in the reconfigurable TPU computing device; Correspondingly, the preliminary decomposition parameter is optimized according to the resource state information, including: In the case that the calculation utilization rate exceeds a preset calculation utilization rate threshold, a decomposition size optimization coefficient is calculated according to the calculation utilization rate and the preset calculation utilization rate threshold, and the decomposition size optimization coefficient is multiplied by the decomposition size to obtain an optimized decomposition size; In the case that the bandwidth utilization rate exceeds a preset bandwidth utilization rate threshold, an overlap ratio optimization coefficient is calculated according to 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 an optimized overlap ratio; In a case where the local memory occupancy rate exceeds a preset occupancy threshold, an accuracy compression ratio optimization coefficient is calculated according to the local memory occupancy rate and the preset occupancy threshold, and the accuracy compression ratio optimization coefficient is multiplied by the accuracy compression ratio to obtain an optimized accuracy compression ratio.

8. An assisted driving data decomposition system of a reconfigurable TPU computing device, characterized in that, The system comprises: A preliminary scheme determination module is configured to acquire training auxiliary driving data, and determine preliminary decomposition parameters according to data attribute information of the training auxiliary driving data and operator combination information of an auxiliary driving module deployed on a reconfigurable TPU computing device; wherein the data attribute information comprises modality type information, perception priority information and continuous change rate information, and the preliminary decomposition parameters comprise decomposition size, overlap ratio and accuracy compression ratio; A TPU reconfiguration module is configured to decompose the training auxiliary driving data according to the preliminary decomposition parameters to obtain a preliminary decomposition result, and dynamically reconfigure the reconfigurable TPU computing device based on the preliminary decomposition parameters; A decomposition result input module is configured to input the preliminary decomposition result to the auxiliary driving module, and collect resource state information of the reconfigurable TPU computing device; A decomposition parameter optimization module is configured to optimize the preliminary decomposition parameters according to the resource state information, and dynamically reconfigure the reconfigurable TPU computing device until the preliminary decomposition parameters meet a preset optimization stop condition, to obtain target decomposition parameters for decomposing actual auxiliary driving data according to the target decomposition parameters.

9. An electronic device, comprising: A processor, a memory and a program or instructions stored on the memory and executable on the processor are included, and the program or instructions are executed by the processor to implement the auxiliary driving data decomposition method of the reconfigurable TPU computing device according to any one of claims 1-7.

10. A readable storage medium, characterized by, A readable storage medium stores a program or instructions, and the program or instructions are executed by a processor to implement the auxiliary driving data decomposition method of the reconfigurable TPU computing device according to any one of claims 1-7.

Images