A reconfigurable operator combination determination method and device for assisted driving

Abstract

The application discloses a reconfigurable operator combination determination method and device for assisted driving, and belongs to the technical field of electric digital data processing. The method comprises the following steps: acquiring key feature information of preset road condition data of a vehicle and hardware performance information of a tensor processing unit of the vehicle; selecting candidate operators based on the key feature information and the hardware performance information and determining perturbation indexes of the candidate operators; pairing the candidate operators two by two to obtain basic operator pairs and determining matching scores of the basic operator pairs according to the perturbation indexes of the candidate operators in the basic operator pairs; selecting a preset number of basic operator pairs as candidate operator pairs in the order from high to low of the matching scores and determining a reconfigurable operator combination to be applied to the vehicle according to the candidate operator pairs. According to the technical scheme, the road condition data features and the TPU hardware characteristics are accurately matched, and the influence of the quantized operator cooperative disturbance is quantified, so that reasonable screening of the reconfigurable operator combination is realized, and the real-time performance of the assisted driving control is improved.

Description

Technical Field

[0001] This application belongs to the field of electronic digital data processing technology, specifically relating to a method and apparatus for determining a combination of reconfigurable operators for assisted driving. Background Technology

[0002] In assisted driving technology, road condition data can be acquired through onboard sensors, and then the model can infer the road condition data and output corresponding control commands to achieve precise control of the vehicle's driving status (such as adaptive cruise control and emergency obstacle avoidance). The smoothness of this process depends on the efficient operation of the model inference. The core factor affecting the efficiency of model inference is the rationality of the combination of reconfigurable operators: operators are the basic units of model computation, and their combination affects core elements such as computing power allocation and data transmission efficiency. A reasonable combination of reconfigurable operators can improve inference speed and stability, while an unreasonable combination of reconfigurable operators will reduce efficiency and ultimately affect the effectiveness of assisted driving.

[0003] Currently, with the development of multi-sensor fusion applications, road condition data exhibits multi-source heterogeneous characteristics, and the onboard tensor processing units (TPUs) equipped in different vehicle models have significant differences in hardware architecture. However, existing operator selection schemes are mostly limited to single performance indicators, which not only fails to take into account the multi-source heterogeneous nature and differentiated processing needs of current road condition data, but also ignores the compatibility between operators and TPU hardware characteristics. This can easily lead to reconfigurable operator combinations failing to accurately meet the specific computing needs and real-time processing standards of current assisted driving scenarios. Summary of the Invention

[0004] This application provides a method and apparatus for determining reconfigurable operator combinations for assisted driving. The purpose is to improve the real-time performance of assisted driving control by accurately matching road condition data characteristics with TPU hardware features and quantifying the impact of cooperative perturbations of operators to achieve reasonable screening of reconfigurable operator combinations.

[0005] In a first aspect, this application provides a method for determining a combination of reconfigurable operators for assisted driving, the method comprising:

[0006] The system acquires key feature information of preset road condition data of the vehicle, as well as hardware performance information of the vehicle's tensor processing unit; wherein the preset road condition data includes image data and at least one type of radar data.

[0007] Candidate operators are selected from a preset operator pool based on the key feature information and the hardware performance information, and the perturbation index of the candidate operators is determined; wherein, the perturbation index includes cache perturbation index, bandwidth perturbation index and synchronization perturbation index;

[0008] The candidate operators are paired up to obtain basic operator pairs, and the matching score of the basic operator pairs is determined based on the perturbation exponents of the two candidate operators in the basic operator pairs.

[0009] A preset number of basic operator pairs are selected as candidate operator pairs according to the matching scores from high to low, and the reconfigurable operator combination to be applied to the vehicle is determined based on the candidate operator pairs for assisted driving when the vehicle is driving.

[0010] Optionally, determining the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pair includes:

[0011] The occurrence count of each operator in all candidate operator pairs is counted, and the initial score of each operator is determined based on the occurrence count and the matching score.

[0012] Arbitrarily combine the operators to obtain candidate reconfigurable operator combinations, and determine the estimated cooperative efficiency of the candidate reconfigurable operator combinations on the tensor processing unit based on the hardware performance information;

[0013] Based on the initial score and the estimated collaborative efficiency, the total score of the candidate reconfigurable operator combination is determined;

[0014] The combination of reconfigurable operators to be applied to the vehicle is determined from the candidate combinations of reconfigurable operators based on the total score.

[0015] Optionally, the hardware performance information includes the peak computing power of the computing unit, memory bandwidth, and cache capacity;

[0016] Accordingly, determining the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit based on the hardware performance information includes:

[0017] The computing unit utilization weight is determined based on the peak computing power of the computing unit, and the memory access overhead weight is determined based on the memory bandwidth and the cache capacity.

[0018] A collaborative efficiency calculation function is constructed based on the utilization weight of the computing unit and the memory access overhead weight.

[0019] Substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the cooperative efficiency calculation function, the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit is obtained.

[0020] Optionally, the step of selecting candidate operators from a preset operator pool based on the key feature information and the hardware performance information includes:

[0021] Based on the key feature information, the compatibility information of each operator in the preset operator pool with the preset traffic data is determined; wherein, the key feature information includes data type information, real-time requirement information, and resource requirement information;

[0022] Based on the hardware performance information, the fitness reference information of each operator in the preset operator pool for the tensor processing unit is determined; wherein, the fitness reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption;

[0023] Based on the compatibility information and the adaptability reference information, candidate operators are selected from the preset operator pool.

[0024] Optionally, determining the adaptation reference information of each operator in the preset operator pool to the tensor processing unit based on the hardware performance information includes:

[0025] Construct a flow resistance topology diagram based on the hardware performance information;

[0026] Computational fluid data is generated based on the data type information and the attribute information of each operator in the preset operator pool;

[0027] The predicted flow velocity of the computational fluid data is determined based on the flow resistance topology diagram as the data processing efficiency.

[0028] Optionally, determining the predicted flow velocity of the computational fluid data as data processing efficiency based on the flow resistance topology map includes:

[0029] The fluid properties of the computational fluid dynamics data are mapped to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and a baseline initial pressure value is assigned to the remaining nodes of the flow resistance topology graph.

[0030] Based on the pressure distribution of the flow resistance topology, the instantaneous flow rate of the computational fluid data flowing through each side is calculated, and the pressure values ​​of all nodes in the flow resistance topology are updated based on the instantaneous flow rate.

[0031] Repeat the above steps until the variation norm of the pressure distribution in the flow resistance topology is less than the preset convergence threshold.

[0032] The instantaneous flow rate of the output edge of the flow resistance topology is determined as the predicted flow velocity to obtain the data processing efficiency.

[0033] Optionally, determining the matching score of the basic operator pair based on the perturbation exponents of the two candidate operators in the basic operator pair includes:

[0034] The buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the component perturbation cost, and the component perturbation cost is weighted and summed to obtain the total perturbation cost.

[0035] The absolute difference values ​​of the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are calculated to obtain the individual complementary gain values. The individual complementary gain values ​​that exceed the preset complementary gain threshold are summed to obtain the total complementary gain value.

[0036] The first calculation result is obtained by subtracting the total disturbance cost from the preset base score, and the second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The matching score of the basic operator pair is obtained by summing the first calculation result and the second calculation result.

[0037] Secondly, this application provides a reconfigurable operator combination determination device for assisted driving, the device comprising:

[0038] The information acquisition module is used to acquire key feature information of the vehicle's preset road condition data, as well as the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data.

[0039] The candidate determination module is used to select candidate operators from a preset operator pool based on the key feature information and the hardware performance information, and to determine the perturbation index of the candidate operators; wherein, the perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index;

[0040] The score determination module is used to pair the candidate operators one by one to obtain basic operator pairs, and to determine the matching score of the basic operator pairs based on the perturbation indexes of the two candidate operators in the basic operator pairs;

[0041] The combination determination module is used to select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs, for assisted driving when the vehicle is driving.

[0042] Optionally, the combination determination module is specifically used for:

[0043] The occurrence count of each operator in all candidate operator pairs is counted, and the initial score of each operator is determined based on the occurrence count and the matching score.

[0044] Arbitrarily combine the operators to obtain candidate reconfigurable operator combinations, and determine the estimated cooperative efficiency of the candidate reconfigurable operator combinations on the tensor processing unit based on the hardware performance information;

[0045] Based on the initial score and the estimated collaborative efficiency, the total score of the candidate reconfigurable operator combination is determined;

[0046] The combination of reconfigurable operators to be applied to the vehicle is determined from the candidate combinations of reconfigurable operators based on the total score.

[0047] Optionally, the hardware performance information includes the peak computing power of the computing unit, memory bandwidth, and cache capacity;

[0048] Accordingly, the combination determination module is specifically used for:

[0049] The computing unit utilization weight is determined based on the peak computing power of the computing unit, and the memory access overhead weight is determined based on the memory bandwidth and the cache capacity.

[0050] A collaborative efficiency calculation function is constructed based on the utilization weight of the computing unit and the memory access overhead weight.

[0051] Substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the cooperative efficiency calculation function, the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit is obtained.

[0052] Optionally, the candidate determination module is specifically used for:

[0053] Based on the key feature information, the compatibility information of each operator in the preset operator pool with the preset traffic data is determined; wherein, the key feature information includes data type information, real-time requirement information, and resource requirement information;

[0054] Based on the hardware performance information, the fitness reference information of each operator in the preset operator pool for the tensor processing unit is determined; wherein, the fitness reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption;

[0055] Based on the compatibility information and the adaptability reference information, candidate operators are selected from the preset operator pool.

[0056] Optionally, the candidate determination module is specifically used for:

[0057] Construct a flow resistance topology diagram based on the hardware performance information;

[0058] Computational fluid data is generated based on the data type information and the attribute information of each operator in the preset operator pool;

[0059] The predicted flow velocity of the computational fluid data is determined based on the flow resistance topology diagram as the data processing efficiency.

[0060] Optionally, the candidate determination module is specifically used for:

[0061] The fluid properties of the computational fluid dynamics data are mapped to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and a baseline initial pressure value is assigned to the remaining nodes of the flow resistance topology graph.

[0062] Based on the pressure distribution of the flow resistance topology, the instantaneous flow rate of the computational fluid data flowing through each side is calculated, and the pressure values ​​of all nodes in the flow resistance topology are updated based on the instantaneous flow rate.

[0063] Repeat the above steps until the variation norm of the pressure distribution in the flow resistance topology is less than the preset convergence threshold.

[0064] The instantaneous flow rate of the output edge of the flow resistance topology is determined as the predicted flow velocity to obtain the data processing efficiency.

[0065] Optionally, the score determination module is specifically used for:

[0066] The buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the component perturbation cost, and the component perturbation cost is weighted and summed to obtain the total perturbation cost.

[0067] The absolute difference values ​​of the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are calculated to obtain the individual complementary gain values. The individual complementary gain values ​​that exceed the preset complementary gain threshold are summed to obtain the total complementary gain value.

[0068] The first calculation result is obtained by subtracting the total disturbance cost from the preset base score, and the second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The matching score of the basic operator pair is obtained by summing the first calculation result and the second calculation result.

[0069] Thirdly, this application provides an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method described in the first aspect.

[0070] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the method described in the first aspect.

[0071] In this application, key feature information of preset road condition data of a vehicle and hardware performance information of the vehicle's tensor processing unit are obtained. The preset road condition data includes image data and at least one type of radar data. Based on the key feature information and the hardware performance information, candidate operators are selected from a preset operator pool, and the perturbation index of the candidate operators is determined. The perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index. The candidate operators are paired to obtain basic operator pairs, and the matching score of the basic operator pairs is determined based on the perturbation indices of the two candidate operators. A preset number of basic operator pairs are selected as candidate operator pairs according to the matching scores from high to low, and the reconfigurable operator combination to be applied to the vehicle is determined based on the candidate operator pairs for assisted driving during vehicle operation. The above-described method for determining the reconfigurable operator combination for assisted driving achieves reasonable screening of reconfigurable operator combinations by accurately matching road condition data features with TPU hardware characteristics and quantifying the impact of operator-coordinated perturbations, thereby improving the real-time performance of assisted driving control. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating a method for determining a combination of reconfigurable operators for assisted driving, provided in an embodiment of this application.

[0073] Figure 2 This is a flowchart illustrating another method for determining the combination of reconfigurable operators for assisted driving, provided in an embodiment of this application.

[0074] Figure 3 This is a flowchart illustrating another method for determining the combination of reconfigurable operators for assisted driving, provided in an embodiment of this application.

[0075] Figure 4 This is an example diagram of a flow resistance topology provided in an embodiment of this application;

[0076] Figure 5 This is a schematic diagram of a reconfigurable operator combination determination device for assisted driving provided in an embodiment of this application;

[0077] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0079] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0080] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0081] The method and apparatus for determining the combination of reconfigurable operators for assisted driving provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0082] First, this application applies to scenarios where a reasoning and prediction model for assisted driving decisions is pre-configured before a vehicle with assisted driving functions is actually put into use. Based on the above application scenario, it is understood that the executing entity of this application can be a terminal device with computing capabilities, and no specific limitation is made here.

[0083] Figure 1 This is a flowchart illustrating a method for determining the combination of reconfigurable operators for assisted driving, provided in an embodiment of this application. Figure 1As shown, the specific steps include the following:

[0084] S101, acquire key feature information of the vehicle's preset road condition data, as well as the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data.

[0085] The preset road condition data can be pre-set environmental perception data used to assist driving decisions, and may include image data and at least one type of radar data. Specifically, the image data can be color images, infrared images, or grayscale images captured by the vehicle-mounted camera, containing visual information of environmental elements such as lane lines, traffic lights, pedestrians, non-motorized vehicles, other vehicles, and road signs; the radar data can be ranging and angle measurement data collected by the vehicle-mounted radar sensor to characterize target distance, speed, and azimuth, and may include millimeter-wave radar data, lidar data, and ultrasonic radar data.

[0086] Among them, the key feature information of the preset traffic condition data can be the core attribute information that reflects the preset traffic condition data processing needs and affects the selection of operators, and can include data type information, real-time requirement information and resource requirement information.

[0087] In one embodiment, the key feature information of the vehicle's preset road condition data can be determined by using the performance information of the vehicle's onboard camera and onboard radar sensor, or by combining the vehicle's preset driving scenario type, the feature statistics of historical road condition data (big data), and the accuracy requirements of the data for assisted driving decision-making.

[0088] Among them, the Tensor Processing Unit (TPU) can be a dedicated hardware acceleration unit on the vehicle for processing tensor operations. It supports parallel computing, has configurable operation logic, and can efficiently handle tensor-intensive tasks such as image recognition and point cloud segmentation in assisted driving.

[0089] Among them, the hardware performance information of the tensor processing unit can be core parameter information that reflects the computing power and resource limitations of the TPU.

[0090] In one embodiment, the hardware performance information of the vehicle's tensor processing unit can be obtained by reading the vehicle's tensor processing unit's factory documentation or user manual.

[0091] S102, based on the key feature information and the hardware performance information, candidate operators are selected from the preset operator pool, and the perturbation index of the candidate operators is determined; wherein, the perturbation index includes cache perturbation index, bandwidth perturbation index and synchronization perturbation index.

[0092] Here, the operator can be the basic computational unit in the inference and prediction model of assisted driving decision-making. Correspondingly, the preset operator pool can be a pre-built set of operators that stores multiple TPUs (e.g., including convolution operators, pooling operators, sparse computation operators, temporal fusion operators, etc.); the candidate operators can be operators selected from the preset operator pool that match the key feature information of the preset road condition data and the TPU hardware performance information.

[0093] In one embodiment, the method of selecting candidate operators from a preset operator pool based on key feature information and hardware performance information can be achieved by obtaining the attribute information of each operator in the preset operator pool, establishing screening constraints based on the key feature information and hardware performance information, and determining the operators whose attribute information meets the screening constraints as candidate operators. The attribute information can be information inherent to the operator during its design and implementation, used to describe its function, computational characteristics, resource requirements, and performance, and may include functional identification information, computational complexity information, and storage access information, etc.

[0094] The disturbance index of a candidate operator is a quantitative indicator used to measure the degree of interference caused by the candidate operator on hardware resources and task scheduling synchronization process when the candidate operator runs on the TPU. It can include cache disturbance index, bandwidth disturbance index, and synchronization disturbance index. Specifically, the cache disturbance index is a quantitative value of the decrease in cache hit rate caused by the candidate operator's occupation and replacement of the TPU cache; the bandwidth disturbance index is a quantitative value of the degree of bandwidth contention caused by the candidate operator's occupation of the TPU's data input / output bandwidth; and the synchronization disturbance index is a quantitative value of the synchronization waiting time caused by the mismatch between the scheduling rhythm of the candidate operator and other operators.

[0095] In one embodiment, the method of determining the perturbation index of candidate operators based on key feature information and hardware performance information can be achieved by pre-establishing a mapping table between key feature information, hardware performance information, attribute information and perturbation index, and querying the mapping table with the attribute information of the currently determined candidate operators, the currently acquired key feature information and hardware performance information to obtain the perturbation index corresponding to each candidate operator.

[0096] S103, pair the candidate operators to obtain basic operator pairs, and determine the matching score of the basic operator pairs based on the perturbation indexes of the two candidate operators in the basic operator pairs.

[0097] Among them, the basic operator pair can be a binary operator combination formed by combining any two different candidate operators from all candidate operators.

[0098] In one embodiment, the method of pairing candidate operators to obtain basic operator pairs can be achieved by ordering the candidate operator set. For each candidate operator, candidate operators with an index greater than the candidate operator index are selected in sequence and paired to obtain basic operator pairs without repetition.

[0099] Among them, the matching score of the basic operator pair can be a quantitative indicator used to measure the degree of cooperative adaptation between the two candidate operators in the basic operator pair. The higher the matching score, the better the compatibility, the higher the resource utilization, and the better the processing efficiency when the two candidate operators run in cooperation.

[0100] In one embodiment, the matching score of a basic operator pair can be determined based on the perturbation indices of the two candidate operators in the basic operator pair by converting the perturbation indices of the two candidate operators in the basic operator pair into perturbation index vectors [cache perturbation index, bandwidth perturbation index, synchronization perturbation index], calculating the similarity between the two perturbation index vectors, and converting the similarity into a percentage score to obtain the matching score.

[0101] Optionally, determining the matching score of the basic operator pair based on the perturbation exponents of the two candidate operators in the basic operator pair includes:

[0102] The buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the component perturbation cost, and the component perturbation cost is weighted and summed to obtain the total perturbation cost.

[0103] The absolute difference values ​​of the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are calculated to obtain the individual complementary gain values. The individual complementary gain values ​​that exceed the preset complementary gain threshold are summed to obtain the total complementary gain value.

[0104] The first calculation result is obtained by subtracting the total disturbance cost from the preset base score, and the second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The matching score of the basic operator pair is obtained by summing the first calculation result and the second calculation result.

[0105] In one embodiment, the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the individual perturbation costs. The weighted summation of these individual perturbation costs yields the total perturbation cost, calculated using the following formula:

[0106] ; ; ;

[0107] ;

[0108] in, and These are the cache perturbation exponents of the two candidate operators in the basic operator pair, respectively. This is the cost of cache item perturbation; and These are the bandwidth perturbation indices of the two candidate operators in the basic operator pair. This is the cost of bandwidth component disturbance; and These are the synchronization perturbation indices of the two candidate operators in the basic operator pair. The cost of synchronous component disturbance; , as well as These are the preset weights corresponding to the perturbation costs of the cache, bandwidth, and synchronization components, respectively. This is the cost of the total disturbance.

[0109] In one embodiment, the formula for calculating the complementary gain value by performing absolute difference calculations on the cache perturbation index, bandwidth perturbation index, and synchronization perturbation index values ​​of the two candidate operators in the basic operator pair is as follows:

[0110] ; ; ;

[0111] in, To cache the complementary gain values ​​of the sub-items, This represents the complementary gain value for the bandwidth components. These are the complementary gain values ​​for synchronous components.

[0112] The preset complementary gain threshold can be a pre-set critical value for judging whether the complementary gain value of a sub-item has practical value. When the complementary gain value of a sub-item exceeds the preset complementary gain threshold, it means that the difference in demand intensity between the two candidate operators in the sub-item resource dimension has reached the effective complementarity standard. That is, the high demand of one operator for the sub-item resource can be balanced by the low demand of the other operator. Therefore, the total complementary gain value is obtained by summing the complementary gain values ​​of the sub-items that exceed the preset complementary gain threshold. This can screen out complementary sub-items with practical optimization value and avoid meaningless differences from affecting the scoring results.

[0113] The preset complementary gain coefficient can be a coefficient used to adjust the contribution weight of the total complementary gain value to the matching score.

[0114] The preset base score can be a pre-set benchmark score for matching scores.

[0115] In one embodiment, a first calculation result is obtained by subtracting the total perturbation cost from the preset base score, and a second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The formula for calculating the matching score of the base operator pair by summing the first and second calculation results is as follows:

[0116] ; ;

[0117] ;

[0118] in, The first calculation result, The preset base score; This is the second calculation result, where k is the preset complementary gain coefficient; The matching score of the basic operator pair.

[0119] The advantage of this approach is that by accurately identifying the complementary potential of operator pairs in terms of resource dimensions such as caching, bandwidth, and synchronization, it ensures that the selected operator pairs can both reduce hardware operation interference and maximize resource utilization.

[0120] S104, select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs for assisted driving when the vehicle is driving.

[0121] The preset quantity can be the number of operator pairs selected in advance based on the requirements of the assisted driving scenario and the parallel processing capability of the TPU.

[0122] Among them, the candidate operator pairs can be a preset number of basic operator pairs selected from all basic operator pairs based on their matching scores.

[0123] In one embodiment, the method of selecting a preset number of basic operator pairs as candidate operator pairs according to the matching score from high to low can be achieved by sorting the matching scores of all basic operator pairs in descending order to obtain a sorted list of operator pairs, and then selecting a preset number of basic operator pairs as candidate operator pairs starting from the head of the list.

[0124] The reconfigurable operator combination can be a dynamically adjustable set of operators constructed based on candidate operator pairs. Correspondingly, the reconfigurable operator combination to be applied to the vehicle can be the optimal operator combination determined from candidate operator pairs for the current vehicle (the vehicle's preset driving scenario and TPU hardware state).

[0125] In one embodiment, the method for determining the reconfigurable operator combination to be applied to the vehicle based on candidate operator pairs can be as follows: arbitrarily combine candidate operator pairs and remove duplicate operators to obtain candidate reconfigurable operator combinations; sum the matching scores of the candidate operator pairs constituting the candidate reconfigurable operator combinations to obtain a total matching score; count the number of operators in the candidate reconfigurable operator combinations; divide the total matching score by the number of operators to obtain an evaluation score for the candidate reconfigurable operator combinations; and determine the candidate reconfigurable operator combination with the highest evaluation score as the reconfigurable operator combination to be applied to the vehicle.

[0126] In this embodiment, key feature information of preset road condition data of the vehicle and hardware performance information of the vehicle's tensor processing unit are acquired. The preset road condition data includes image data and at least one type of radar data. Candidate operators are selected from a preset operator pool based on the key feature information and the hardware performance information, and the perturbation index of the candidate operators is determined. The perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index. The candidate operators are paired to obtain basic operator pairs, and the matching score of the basic operator pairs is determined based on the perturbation indices of the two candidate operators in each basic operator pair. A preset number of basic operator pairs are selected as candidate operator pairs according to the matching scores from high to low, and the reconfigurable operator combination to be applied to the vehicle is determined based on the candidate operator pairs for assisted driving during vehicle operation. The above-described method for determining reconfigurable operator combinations for assisted driving achieves reasonable screening of reconfigurable operator combinations by accurately matching road condition data features with TPU hardware characteristics and quantifying the impact of operator-coordinated perturbations, thereby improving the real-time performance of assisted driving control.

[0127] Figure 2 This is a flowchart illustrating another method for determining the combination of reconfigurable operators for assisted driving, provided in an embodiment of this application. Figure 2 As shown, the specific steps include the following:

[0128] S201, acquire key feature information of the vehicle's preset road condition data, as well as the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data.

[0129] S202, based on the key feature information and the hardware performance information, candidate operators are selected from the preset operator pool, and the perturbation index of the candidate operators is determined; wherein, the perturbation index includes cache perturbation index, bandwidth perturbation index and synchronization perturbation index.

[0130] S203, pair the candidate operators to obtain basic operator pairs, and determine the matching score of the basic operator pairs based on the perturbation indexes of the two candidate operators in the basic operator pairs.

[0131] S204, select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low.

[0132] S205, count the number of times each operator in the candidate operator pair appears in all candidate operator pairs, and determine the initial score of each operator based on the number of appearances and the matching score.

[0133] The occurrence count can be the total number of times a single operator is included in a pair of candidate operators.

[0134] The preliminary score can be the matching score and the number of occurrences of the candidate operator pair to which the comprehensive operator belongs, which is a quantitative evaluation value of the operator's own adaptability.

[0135] In one embodiment, the method for determining the initial score of each operator based on the number of occurrences and the matching score can be as follows: for each operator, sum the matching scores of the candidate operator pair to which the operator belongs to obtain the evaluation score of the operator, and divide the evaluation score of the operator by the number of occurrences of the operator to obtain the initial score of the operator.

[0136] S206, arbitrarily combine the operators to obtain candidate reconfigurable operator combinations, and determine the estimated collaborative efficiency of the candidate reconfigurable operator combinations on the tensor processing unit based on the hardware performance information.

[0137] Among them, the candidate reconfigurable operator combination can be a set consisting of multiple operators included in all candidate operator pairs.

[0138] In one embodiment, the method of arbitrarily combining each operator to obtain a candidate reconfigurable operator combination can be achieved by generating a candidate reconfigurable operator combination by performing a non-repeating combination of each operator, provided that the number of operators included in the candidate reconfigurable operator combination is not less than two and not more than the total number of all operators.

[0139] Among them, the estimated collaborative efficiency can be a quantitative indicator of the comprehensive performance of the predicted combination of candidate reconfigurable operators when running collaboratively on the current TPU hardware.

[0140] In one embodiment, the method for determining the estimated collaborative efficiency of candidate reconfigurable operator combinations on the tensor processing unit based on hardware performance information can be as follows: The total resource utilization rate of the candidate reconfigurable operator combination is calculated based on the attribute information of each operator in the candidate reconfigurable operator combination and the TPU hardware performance information. Then, the current estimated processing latency is obtained through a pre-trained latency prediction model (inputting total computation and total resource utilization rate, outputting estimated processing latency). The estimated processing latency is divided by a preset maximum tolerable processing latency to obtain a latency normalization coefficient. 1 is subtracted from the total resource utilization rate to obtain a first difference, and 1 is subtracted from the latency normalization coefficient to obtain a second difference. The average of the first difference and the second difference is calculated and converted to a percentage to obtain the estimated collaborative efficiency of the candidate reconfigurable operator combination on the tensor processing unit.

[0141] Optionally, the hardware performance information includes the peak computing power of the computing unit, memory bandwidth, and cache capacity;

[0142] Accordingly, determining the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit based on the hardware performance information includes:

[0143] The computing unit utilization weight is determined based on the peak computing power of the computing unit, and the memory access overhead weight is determined based on the memory bandwidth and the cache capacity.

[0144] A collaborative efficiency calculation function is constructed based on the utilization weight of the computing unit and the memory access overhead weight.

[0145] Substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the cooperative efficiency calculation function, the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit is obtained.

[0146] Among them, the peak computing power of the computing unit can be the maximum number of operations that the computing core in the TPU can complete per unit time; the memory bandwidth can be the data transfer rate between the TPU and the memory; and the cache capacity can be the total storage capacity of the TPU's built-in cache.

[0147] Among them, the computing unit utilization weight can be a coefficient used to adjust the contribution of computing unit utilization to the estimated collaborative efficiency; the memory access overhead weight can be a coefficient used to adjust the contribution of memory access overhead to the estimated collaborative efficiency.

[0148] In one embodiment, the method of determining the computing unit utilization weight based on the peak computing power of the computing unit and the memory access overhead weight based on the memory bandwidth and cache capacity can be achieved by normalizing the peak computing power, memory bandwidth, and cache capacity of the computing unit, calculating the average of the normalized memory bandwidth and normalized cache capacity as the comprehensive value of memory-related performance, multiplying the normalized peak computing power of the computing unit by a preset performance balance coefficient (e.g., 0.5) to obtain the first weight, calculating the difference between 1 and the preset performance balance coefficient and multiplying the difference by the comprehensive value of memory-related performance to obtain the second weight, and normalizing the first weight and the second weight to ensure that the sum of the normalized first weight and the normalized second weight is 1. The obtained normalized first weight is the computing unit utilization weight, and the obtained normalized second weight is the memory access overhead weight.

[0149] The collaborative efficiency calculation function can be a mathematical model that combines the utilization rate of computing units and memory access overhead to quantify the collaborative operation efficiency of candidate reconfigurable operators. The specific mathematical expression is as follows:

[0150] ;

[0151] in, This is to estimate the cooperative efficiency of candidate reconfigurable operator combinations on tensor processing units. To calculate the unit utilization weight, Weights for memory access overhead. The sum of computational costs for all operators in the candidate combinatorial reconfigurable operator group. T represents the peak computing power of the TPU computing unit, where T is the preset maximum tolerable processing time. This represents the total memory accesses of all operators in the candidate reconfigurable operator combination. For TPU memory bandwidth, This represents the TPU cache capacity.

[0152] Correspondingly, by substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the co-efficiency calculation function, the estimated co-efficiency of the candidate reconfigurable operator combination on the tensor processing unit can be obtained.

[0153] The advantage of this approach is that it accurately anchors the core hardware performance parameters of the TPU, balances the impact of computing power and memory performance through dynamic weighting, avoids the one-sidedness of evaluating a single performance dimension, and thus ensures that the selected combination of refactoring operators achieves optimal collaborative operation under hardware resource constraints.

[0154] S207, determine the total score of the candidate reconfigurable operator combination based on the initial score and the estimated collaborative efficiency.

[0155] The total score can be the final quantitative evaluation value of the overall fitness of the combination of individual operators in the combination of candidate reconfigurable operators and the overall collaborative efficiency of the combination.

[0156] In one embodiment, the total score of the candidate reconfigurable operator combination can be determined by weighting and summing the initial score and the estimated collaborative efficiency according to preset weights.

[0157] S208, based on the total score, determine the combination of reconfigurable operators to be applied to the vehicle from the candidate combinations of reconfigurable operators, for use in assisting driving when the vehicle is driving.

[0158] In one embodiment, the method of determining the reconfigurable operator combination to be applied to the vehicle from the candidate reconfigurable operator combinations based on the total score can be as follows: for each candidate reconfigurable operator combination, calculate the ratio of the total score of the candidate reconfigurable operator combination to the number of operators it includes, and determine the candidate reconfigurable operator combination with the highest ratio as the reconfigurable operator combination to be applied to the vehicle.

[0159] The advantage of this scheme is that it ensures the basic adaptability of individual operators, highlights the overall collaborative performance of the combination of reconfigurable operators, and avoids the decline in hardware scheduling efficiency caused by overly complex combinations of reconfigurable operators by filtering the total score to the number of operators, thereby maximizing hardware resource utilization and reducing running latency.

[0160] Figure 3 This is a flowchart illustrating another method for determining the combination of reconfigurable operators for assisted driving, provided in an embodiment of this application. Figure 3 As shown, the specific steps include the following:

[0161] S301, acquire key feature information of the vehicle's preset road condition data, as well as the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data.

[0162] S302, based on the key feature information, determine the compatibility information of each operator in the preset operator pool with the preset traffic data; wherein, the key feature information includes data type information, real-time requirement information, and resource requirement information.

[0163] Among them, data type information can be quantitative information on the form and structural characteristics of preset traffic data, including data format information, data dimension information, and data distribution density information of preset traffic data; real-time requirement information can be time constraint information for processing preset traffic data, such as processing latency limit, data update frame rate, etc.; resource requirement information can be quantitative indicators of hardware resources required to process preset traffic data, such as data storage occupancy, computing power requirement, and bandwidth requirement.

[0164] Among them, compatibility information can be a quantitative evaluation result of whether the operator can meet the processing requirements of preset road condition data and adapt to data characteristics.

[0165] In one embodiment, the method for determining the compatibility information of each operator in the preset operator pool with the preset traffic data based on key feature information can be as follows: based on the key feature information and the attribute information of each operator in the preset operator pool, each operator is scored according to the preset scoring rules in three dimensions: data type adaptability, real-time requirement adaptability, and resource requirement adaptability. The compatibility information of the operator with the preset traffic data is obtained by weighted summation of the three scoring results.

[0166] S303, based on the hardware performance information, determine the adaptation reference information of each operator in the preset operator pool to the tensor processing unit; wherein, the adaptation reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption.

[0167] The adaptation reference information can be a comprehensive set of information on the performance and resource utilization status of the quantization operator when running on the current TPU, which may include parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption.

[0168] Specifically, the parallel computing core utilization rate can be the actual proportion of TPU parallel computing cores used during operator runtime; the parallel computing core utilization rate can be the ratio of the number of cores used in the attribute information to the total number of cores in the hardware performance information.

[0169] Specifically, data processing efficiency can be the amount of data processed by the operator per unit time. The data processing efficiency is obtained by multiplying the peak computing power of the computing unit in the hardware performance information with the preset efficiency discount factor and dividing it by the amount of computing per frame in the attribute information.

[0170] Specifically, memory bandwidth utilization can be the actual proportion of TPU memory bandwidth used during operator runtime; memory bandwidth utilization can be the ratio of used memory bandwidth in attribute information to total memory bandwidth in hardware performance information.

[0171] Specifically, computing power loss can be the ratio of the difference between the theoretical computing power and the actual available computing power when the operator runs on the TPU to the theoretical computing power. The computing power loss is calculated by dividing the difference between the peak computing power of the computing unit in the hardware performance information and the running computing power in the attribute information by the peak computing power of the computing unit.

[0172] Optionally, determining the adaptation reference information of each operator in the preset operator pool to the tensor processing unit based on the hardware performance information includes:

[0173] Construct a flow resistance topology diagram based on the hardware performance information;

[0174] Computational fluid data is generated based on the data type information and the attribute information of each operator in the preset operator pool;

[0175] The predicted flow velocity of the computational fluid data is determined based on the flow resistance topology diagram as the data processing efficiency.

[0176] Among them, the flow resistance topology diagram can be a topological model that transforms hardware performance information into resistance nodes and transmission paths in fluid mechanics; the flow resistance topology diagram includes nodes, node resistance values, and connection paths between nodes.

[0177] In one embodiment, the method of constructing a flow resistance topology map based on hardware performance information can be achieved by constructing each hardware structure (memory unit, bus transmission unit, cache unit, parallel computing core unit) in the hardware performance information as a node, and connecting each node sequentially according to the data processing logic flow (memory read, bus transmission, cache temporary storage, computing core processing, cache write-back, bus transmission) in the hardware performance information to obtain the connection path between nodes.

[0178] Figure 4 This is an example diagram of a flow resistance topology provided in an embodiment of this application. For example... Figure 4 As shown, the flow resistance topology diagram includes 6 nodes and 6 edges: Node 1 is the input node, Node 2 is the first bus transmission node, Node 3 is the cache node, Node 4 is the parallel computing core node, Node 5 is the second bus transmission node, and Node 6 is the output node; Node 1→Node 2, Node 2→Node 3, Node 3→Node 4, Node 4→Node 3, Node 3→Node 5, and Node 5→Node 6 correspond to the memory read, bus transmission, cache temporary storage, computing core processing, cache write-back, and bus transmission stages, respectively.

[0179] Among them, computational fluid data can be a fluid in fluid dynamics that represents the process of operator processing data, that is, it is used to simulate the processing state of preset road condition data in TPU.

[0180] In one embodiment, the method of generating computational fluid data based on data type information and attribute information of each operator in a preset operator pool can be as follows: the data complexity of preset road condition data is calculated based on data type information as fluid viscosity; the single-frame data storage occupancy in resource requirement information is multiplied by the data update frame rate in real-time requirement information, and the result is divided by the bandwidth requirement threshold in resource requirement information to obtain the initial flow rate; the computational complexity in the attribute information of an operator is determined as the fluid density corresponding to that operator; and the resulting fluid viscosity, initial flow rate, and fluid density corresponding to an operator are the fluid attributes of the computational fluid data corresponding to that operator.

[0181] The predicted flow velocity of computational fluid data can be the flow velocity in the flow resistance topology diagram corresponding to the computational fluid data simulated by a fluid dynamics model, and can be expressed as the data processing efficiency of the corresponding operator on the TPU.

[0182] In one embodiment, the method of determining the predicted flow velocity of computational fluid data as the data processing efficiency based on the flow resistance topology diagram can be achieved by importing the flow resistance topology diagram and the fluid properties of the computational fluid data into fluid dynamics software (e.g., Tecplot 360 EX, ANSYS Fluent, etc.) to obtain the predicted flow velocity of the computational fluid data as the data processing efficiency.

[0183] Optionally, determining the predicted flow velocity of the computational fluid data as data processing efficiency based on the flow resistance topology map includes:

[0184] The fluid properties of the computational fluid dynamics data are mapped to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and a baseline initial pressure value is assigned to the remaining nodes of the flow resistance topology graph.

[0185] Based on the pressure distribution of the flow resistance topology, the instantaneous flow rate of the computational fluid data flowing through each side is calculated, and the pressure values ​​of all nodes in the flow resistance topology are updated based on the instantaneous flow rate.

[0186] Repeat the above steps until the variation norm of the pressure distribution in the flow resistance topology is less than the preset convergence threshold.

[0187] The instantaneous flow rate of the output edge of the flow resistance topology is determined as the predicted flow velocity to obtain the data processing efficiency.

[0188] The input node can be a memory cell node in the TPU memory reading stage of the flow resistance topology diagram, and the injection pressure value of the input node can be a pressure driving value obtained by quantizing the fluid density and initial flow rate based on the computational fluid data.

[0189] In one embodiment, mapping the fluid properties of computational fluid dynamics data to the injection pressure values ​​of the input nodes in the flow resistance topology can be achieved by weighted summation of the fluid density and initial flow rate of the computational fluid dynamics data to obtain the injection pressure value. The weighting coefficients for the weighted summation are equal to preset mapping coefficients corresponding to the fluid density and initial flow rate, respectively.

[0190] The reference initial pressure value can be a preset uniform initial pressure constant.

[0191] In one embodiment, the method for assigning baseline initial pressure values ​​to the remaining nodes in the flow resistance topology can be to directly assign the baseline initial pressure values ​​to the parameters used to store pressure values ​​in all nodes except the input node.

[0192] The pressure distribution in the flow resistance topology diagram can be a set of pressure values ​​for all nodes in the flow resistance topology diagram.

[0193] Instantaneous flow rate can be the amount of fluid flow per unit time on each edge (connection path between nodes) of the flow resistance topology graph.

[0194] In one embodiment, the instantaneous flow rate of computational fluid data flowing through each edge can be calculated based on the pressure distribution of the flow resistance topology. This can be achieved by calculating the pressure difference between the two nodes connected to an edge, dividing this pressure difference by the corresponding link resistance value, and obtaining the instantaneous flow rate for that edge. The link resistance value is calculated by combining the resistance values ​​of the nodes at both ends of the edge with the inherent resistance of the transmission link between the nodes; specifically, it is the sum of the average resistance values ​​of the nodes at both ends and the inherent resistance of the transmission link.

[0195] In one embodiment, the method of updating the pressure values ​​of all nodes in the flow resistance topology graph based on instantaneous flow can be as follows: For each non-input node, first calculate the sum of the instantaneous flow of all edges flowing into the node, then calculate the sum of the instantaneous flow of all edges flowing out of the node, calculate the difference between the sum of inflow and outflow, if the sum of inflow is greater than the sum of outflow, increase the pressure value of the node by multiplying the difference by a preset pressure adjustment coefficient, if the sum of inflow is less than the sum of outflow, decrease the pressure value of the node by multiplying the difference by the preset pressure adjustment coefficient, and if the two are equal, keep the pressure value of the node unchanged.

[0196] The pressure distribution variation norm can be the L2 norm of the pressure value changes of all nodes in two iterations; the preset convergence threshold can be a pre-set variation norm threshold. If the pressure distribution variation norm of the flow resistance topology map is less than the preset convergence threshold, it means that the pressure values ​​of each node in the flow resistance topology map have stabilized and no longer change significantly. At this time, the instantaneous flow rate can accurately reflect the stable flow state of the preset road condition data in the hardware processing link.

[0197] The instantaneous flow rate of the output edge represents the stable flow rate from the output node (memory storage link) after the preset traffic data flows through the entire TPU hardware processing link. Therefore, the instantaneous flow rate of the output edge can be determined as the predicted flow rate, i.e., the data processing efficiency.

[0198] The advantage of this solution is that it can simulate the real flow of preset road condition data in the TPU hardware link, which not only matches the actual hardware processing logic, but also completes the accurate prediction of data processing efficiency without actually running the operator.

[0199] The advantage of this approach is that it allows for accurate prediction of data processing efficiency without actually running the operators, thus solving the problems of high cost and time consumption in offline testing.

[0200] S304, Based on the compatibility information and the adaptability reference information, select candidate operators from the preset operator pool.

[0201] In one embodiment, the method of selecting candidate operators from a preset operator pool based on compatibility information and adaptability reference information can be achieved by pre-setting threshold conditions for compatibility information, parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption, and determining the operators in the preset operator pool that simultaneously meet each threshold condition as candidate operators.

[0202] S305, determine the perturbation index of the candidate operator; wherein the perturbation index includes the cache perturbation index, the bandwidth perturbation index, and the synchronization perturbation index.

[0203] S306, pair the candidate operators to obtain basic operator pairs, and determine the matching score of the basic operator pairs based on the perturbation exponents of the two candidate operators in the basic operator pairs.

[0204] S307, select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs for assisted driving when the vehicle is driving.

[0205] The advantage of this approach is that it ensures that the candidate operators not only meet the processing requirements of the preset road condition data but also adapt to the performance characteristics of the TPU hardware, providing a reliable selection basis for determining the final reconfigurable operator combination.

[0206] Figure 5 This is a schematic diagram of a reconfigurable operator combination determination device for assisted driving, provided in an embodiment of this application. Figure 5 As shown, the device includes:

[0207] The information acquisition module 510 is used to acquire key feature information of the vehicle's preset road condition data and the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data.

[0208] The candidate determination module 520 is used to select candidate operators from a preset operator pool based on the key feature information and the hardware performance information, and to determine the perturbation index of the candidate operators; wherein, the perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index;

[0209] The score determination module 530 is used to pair the candidate operators one by one to obtain basic operator pairs, and to determine the matching score of the basic operator pairs based on the perturbation indexes of the two candidate operators in the basic operator pairs.

[0210] The combination determination module 540 is used to select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs, for assisted driving when the vehicle is driving.

[0211] Optionally, the combination determination module 540 is specifically used for:

[0212] The occurrence count of each operator in all candidate operator pairs is counted, and the initial score of each operator is determined based on the occurrence count and the matching score.

[0213] Arbitrarily combine the operators to obtain candidate reconfigurable operator combinations, and determine the estimated cooperative efficiency of the candidate reconfigurable operator combinations on the tensor processing unit based on the hardware performance information;

[0214] Based on the initial score and the estimated collaborative efficiency, the total score of the candidate reconfigurable operator combination is determined;

[0215] The combination of reconfigurable operators to be applied to the vehicle is determined from the candidate combinations of reconfigurable operators based on the total score.

[0216] Optionally, the hardware performance information includes the peak computing power of the computing unit, memory bandwidth, and cache capacity;

[0217] Accordingly, the combination determination module 540 is specifically used for:

[0218] The computing unit utilization weight is determined based on the peak computing power of the computing unit, and the memory access overhead weight is determined based on the memory bandwidth and the cache capacity.

[0219] A collaborative efficiency calculation function is constructed based on the utilization weight of the computing unit and the memory access overhead weight.

[0220] Substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the cooperative efficiency calculation function, the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit is obtained.

[0221] Optionally, the candidate determination module 520 is specifically used for:

[0222] Based on the key feature information, the compatibility information of each operator in the preset operator pool with the preset traffic data is determined; wherein, the key feature information includes data type information, real-time requirement information, and resource requirement information;

[0223] Based on the hardware performance information, the fitness reference information of each operator in the preset operator pool for the tensor processing unit is determined; wherein, the fitness reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption;

[0224] Based on the compatibility information and the adaptability reference information, candidate operators are selected from the preset operator pool.

[0225] Optionally, the candidate determination module 520 is specifically used for:

[0226] Construct a flow resistance topology diagram based on the hardware performance information;

[0227] Computational fluid data is generated based on the data type information and the attribute information of each operator in the preset operator pool;

[0228] The predicted flow velocity of the computational fluid data is determined based on the flow resistance topology diagram as the data processing efficiency.

[0229] Optionally, the candidate determination module 520 is specifically used for:

[0230] The fluid properties of the computational fluid dynamics data are mapped to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and a baseline initial pressure value is assigned to the remaining nodes of the flow resistance topology graph.

[0231] Based on the pressure distribution of the flow resistance topology, the instantaneous flow rate of the computational fluid data flowing through each side is calculated, and the pressure values ​​of all nodes in the flow resistance topology are updated based on the instantaneous flow rate.

[0232] Repeat the above steps until the variation norm of the pressure distribution in the flow resistance topology is less than the preset convergence threshold.

[0233] The instantaneous flow rate of the output edge of the flow resistance topology is determined as the predicted flow velocity to obtain the data processing efficiency.

[0234] Optionally, the score determination module 530 is specifically used for:

[0235] The buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the component perturbation cost, and the component perturbation cost is weighted and summed to obtain the total perturbation cost.

[0236] The absolute difference values ​​of the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are calculated to obtain the individual complementary gain values. The individual complementary gain values ​​that exceed the preset complementary gain threshold are summed to obtain the total complementary gain value.

[0237] The first calculation result is obtained by subtracting the total disturbance cost from the preset base score, and the second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The matching score of the basic operator pair is obtained by summing the first calculation result and the second calculation result.

[0238] In this embodiment, an information acquisition module is used to acquire key feature information of preset road condition data of the vehicle, and hardware performance information of the tensor processing unit of the vehicle; wherein, the preset road condition data includes image data and at least one radar data; a candidate determination module is used to select candidate operators from a preset operator pool based on the key feature information and the hardware performance information, and determine the perturbation index of the candidate operators; wherein, the perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index; a score determination module is used to pair the candidate operators in pairs to obtain basic operator pairs, and determine the matching score of the basic operator pairs according to the perturbation indices of the two candidate operators in the basic operator pairs; a combination determination module is used to select a preset number of basic operator pairs as candidate operator pairs in descending order of the matching scores, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs, for assisted driving when the vehicle is driving. The aforementioned reconfigurable operator combination determination device for assisted driving achieves reasonable selection of reconfigurable operator combinations by accurately matching road condition data characteristics with TPU hardware characteristics and quantifying the impact of cooperative perturbations of operators, thereby improving the real-time performance of assisted driving control.

[0239] The reconfigurable operator combination determination device for assisted driving in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. This application embodiment does not impose specific limitations.

[0240] The reconfigurable operator combination determination device for assisted driving in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0241] The reconfigurable operator combination determination device for assisted driving provided in this application embodiment can realize each process implemented in the above embodiments. To avoid repetition, it will not be described again here.

[0242] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 6 As shown, this application embodiment also provides an electronic device 600, including a processor 601, a memory 602, and a program or instructions stored in the memory 602 and executable on the processor 601. When the program or instructions are executed by the processor 601, they implement the various processes of the above-described embodiment of the method for determining the combination of reconfigurable operators for assisted driving, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0243] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0244] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described method for determining the combination of reconfigurable operators for assisted driving, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0245] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0246] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0247] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0248] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0249] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. A method for determining the combination of reconfigurable operators for assisted driving, characterized in that, The method includes: The system acquires key feature information of preset road condition data of the vehicle, as well as hardware performance information of the vehicle's tensor processing unit; wherein the preset road condition data includes image data and at least one type of radar data. Candidate operators are selected from a preset operator pool based on the key feature information and the hardware performance information, and the perturbation index of the candidate operators is determined; wherein, the perturbation index includes cache perturbation index, bandwidth perturbation index and synchronization perturbation index; The step of selecting candidate operators from a preset operator pool based on the key feature information and the hardware performance information includes: determining the compatibility information of each operator in the preset operator pool with the preset traffic data based on the key feature information; wherein, the key feature information includes data type information, real-time requirement information, and resource requirement information; determining the adaptability reference information of each operator in the preset operator pool with the tensor processing unit based on the hardware performance information; wherein, the adaptability reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption; and selecting candidate operators from the preset operator pool according to the compatibility information and the adaptability reference information. The step of determining the adaptation reference information of each operator in the preset operator pool to the tensor processing unit based on the hardware performance information includes: constructing a flow resistance topology map based on the hardware performance information; generating computational fluid data based on the data type information and the attribute information of each operator in the preset operator pool; and determining the predicted flow rate of the computational fluid data as the data processing efficiency based on the flow resistance topology map. The step of determining the predicted flow rate of the computational fluid data as the data processing efficiency based on the flow resistance topology graph includes: mapping the fluid properties of the computational fluid data to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and assigning baseline initial pressure values ​​to the remaining nodes in the flow resistance topology graph; calculating the instantaneous flow rate of the computational fluid data through each edge based on the pressure distribution of the flow resistance topology graph, and updating the pressure values ​​of all nodes in the flow resistance topology graph based on the instantaneous flow rate; repeating the above steps until the variation norm of the pressure distribution of the flow resistance topology graph is less than a preset convergence threshold; and determining the instantaneous flow rate of the output edge of the flow resistance topology graph as the predicted flow rate to obtain the data processing efficiency. The candidate operators are paired up to obtain basic operator pairs, and the matching score of the basic operator pairs is determined based on the perturbation exponents of the two candidate operators in the basic operator pairs. A preset number of basic operator pairs are selected as candidate operator pairs according to the matching scores from high to low, and the reconfigurable operator combination to be applied to the vehicle is determined based on the candidate operator pairs for assisted driving when the vehicle is driving.

2. The method for determining the combination of reconfigurable operators for assisted driving according to claim 1, characterized in that, The step of determining the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pair includes: The occurrence count of each operator in all candidate operator pairs is counted, and the initial score of each operator is determined based on the occurrence count and the matching score. Arbitrarily combine the operators to obtain candidate reconfigurable operator combinations, and determine the estimated cooperative efficiency of the candidate reconfigurable operator combinations on the tensor processing unit based on the hardware performance information; Based on the initial score and the estimated collaborative efficiency, the total score of the candidate reconfigurable operator combination is determined; The combination of reconfigurable operators to be applied to the vehicle is determined from the candidate combinations of reconfigurable operators based on the total score.

3. The method for determining the combination of reconfigurable operators for assisted driving according to claim 2, characterized in that, The hardware performance information includes the peak computing power of the computing unit, memory bandwidth, and cache capacity. Accordingly, determining the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit based on the hardware performance information includes: The computing unit utilization weight is determined based on the peak computing power of the computing unit, and the memory access overhead weight is determined based on the memory bandwidth and the cache capacity. A collaborative efficiency calculation function is constructed based on the utilization weight of the computing unit and the memory access overhead weight. Substituting the computational and memory access costs of each operator in the candidate reconfigurable operator combination into the cooperative efficiency calculation function, the estimated cooperative efficiency of the candidate reconfigurable operator combination on the tensor processing unit is obtained.

4. The method for determining the combination of reconfigurable operators for assisted driving according to claim 1, characterized in that, Determining the matching score of the basic operator pair based on the perturbation exponents of the two candidate operators in the basic operator pair includes: The buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are summed to obtain the component perturbation cost, and the component perturbation cost is weighted and summed to obtain the total perturbation cost. The absolute difference values ​​of the buffer perturbation index, bandwidth perturbation index, and synchronization perturbation index of the two candidate operators in the basic operator pair are calculated to obtain the individual complementary gain values. The individual complementary gain values ​​that exceed the preset complementary gain threshold are summed to obtain the total complementary gain value. The first calculation result is obtained by subtracting the total disturbance cost from the preset base score, and the second calculation result is obtained by multiplying the total complementary gain value by the preset complementary gain coefficient. The matching score of the basic operator pair is obtained by summing the first calculation result and the second calculation result.

5. A reconfigurable operator combination determination device for assisted driving, characterized in that, The device includes: The information acquisition module is used to acquire key feature information of the vehicle's preset road condition data, as well as the hardware performance information of the vehicle's tensor processing unit; wherein, the preset road condition data includes image data and at least one type of radar data. The candidate determination module is used to select candidate operators from a preset operator pool based on the key feature information and the hardware performance information, and to determine the perturbation index of the candidate operators; wherein, the perturbation index includes a cache perturbation index, a bandwidth perturbation index, and a synchronization perturbation index; The candidate determination module is specifically used for: determining the compatibility information of each operator in the preset operator pool with the preset road condition data based on the key feature information; wherein the key feature information includes data type information, real-time requirement information, and resource requirement information; determining the adaptability reference information of each operator in the preset operator pool with the tensor processing unit based on the hardware performance information; wherein the adaptability reference information includes parallel computing core utilization, data processing efficiency, memory bandwidth utilization, and computing power consumption; and selecting candidate operators in the preset operator pool according to the compatibility information and the adaptability reference information. The candidate determination module is specifically used for: constructing a flow resistance topology map based on the hardware performance information; generating computational fluid data based on the data type information and the attribute information of each operator in the preset operator pool; and determining the predicted flow rate of the computational fluid data as the data processing efficiency based on the flow resistance topology map. The candidate determination module is specifically used for: mapping the fluid properties of the computational fluid dynamics data to the injection pressure values ​​of the input nodes of the flow resistance topology graph, and assigning baseline initial pressure values ​​to the remaining nodes in the flow resistance topology graph; calculating the instantaneous flow rate of the computational fluid dynamics data through each edge according to the pressure distribution of the flow resistance topology graph, and updating the pressure values ​​of all nodes in the flow resistance topology graph according to the instantaneous flow rate; repeating the above steps until the change norm of the pressure distribution of the flow resistance topology graph is less than a preset convergence threshold; determining the instantaneous flow rate of the output edge of the flow resistance topology graph as the predicted flow velocity, and obtaining the data processing efficiency; The score determination module is used to pair the candidate operators one by one to obtain basic operator pairs, and to determine the matching score of the basic operator pairs based on the perturbation indexes of the two candidate operators in the basic operator pairs; The combination determination module is used to select a preset number of basic operator pairs as candidate operator pairs according to the matching scores from high to low, and determine the reconfigurable operator combination to be applied to the vehicle based on the candidate operator pairs, for assisted driving when the vehicle is driving.

6. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method for determining the combination of reconfigurable operators for assisted driving as described in any one of claims 1-4.

7. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method for determining the combination of reconfigurable operators for assisted driving as described in any one of claims 1-4.

Images