Neural network acceleration chip capability test data generation method based on virtual simulation

Abstract

The invention discloses a neural network acceleration chip capability test data generation method based on virtual simulation. The method is completed through scene parameter planning, basic scene generation, environment data superposition and scene data fusion. According to the method, the number and positions of the sensors can be set to generate test data consistent with a task working environment. Meanwhile, a data transmission interface capable of simulating a real sensor is included in the method, and the interface brings performance testing capacity under the condition that data collection and calculation are asynchronous. A virtual simulation technology is adopted to simulate a task working environment of a neural network acceleration chip, and the problem that a test result and task application experience are greatly deviated due to the fact that a public data set is used in an existing test method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural network acceleration chips, and more particularly, to a method for generating capability test data of a neural network acceleration chip based on virtual simulation. Background Art

[0002] Current neural network acceleration chip performance testing typically uses public data sets as test data. The overall process involves collecting public data sets on the Internet and designing a data interface suitable for the neural network acceleration chip. The neural network acceleration chip reads the data in the data set through the data interface and relies on the computing units in the chip to perform neural network inference. The speed and frame rate of the neural network inference process are used as the basis for evaluating the chip performance.

[0003] Commonly used public datasets include ImageNet for image classification, the COCO dataset for object detection and instance segmentation, the SQuAD dataset for natural language processing, and the Criteo Terabyte Dataset for recommendation systems. Currently, datasets are typically stored in the board's built-in EMMC or SD card, and the data is read through the board's I / O as neural network input.

[0004] Neural network accelerator chips are typically used in relatively fixed production environments, and the data they need to process varies significantly in different environments. For example, testing a neural network accelerator chip designed to process low-resolution images using high-resolution images is not very useful as a reference. Similarly, testing a neural network accelerator chip designed to identify vehicles in traffic using factory parts identification data is also not very useful as a reference. Because currently available public datasets contain few specific production environments, using specific test data for specific production environments is not feasible with existing methods.

[0005] Neural network acceleration chips often need to process time-series data, and sensor and chip computation are typically asynchronous. For example, if a sensor can capture 30 frames per second, but the chip can only process 10 frames per second, 20 frames will be unprocessed. Existing input interfaces require the chip to actively read all data, making existing methods for testing the performance of neural network acceleration chips in this asynchronous environment infeasible.

[0006] Neural network acceleration chips have the multi-task requirement of processing data from multiple sensors simultaneously. For example, they need to process two camera images placed at specific angles at the same time, and test data is difficult to find at this time. Summary of the Invention

[0007] The purpose of the present invention is to provide a test data generation method for neural network acceleration chip performance testing in a multi-task parallel scenario based on a virtual simulation environment. It is achieved through scenario parameter planning, basic scenario generation, environmental data superposition and scenario data fusion. The method can set the number and position of sensors to generate test data consistent with the task working environment. At the same time, the method includes a data transmission interface that can simulate real sensors, which brings the ability to test performance in the case of asynchronous data acquisition and calculation. The use of virtual simulation technology to simulate the task working environment of the neural network acceleration chip solves the problem that the test results of the current test method using public data sets deviate too much from the task application experience. The method can set the number and position of sensors to generate test data consistent with the task working environment. At the same time, the method includes a data transmission interface that can simulate real sensors, which brings the ability to test performance in the case of asynchronous data acquisition and calculation.

[0008] This method, based on virtual simulation technology, primarily utilizes 3D modeling and physical simulation techniques to create a digital model on a computer platform that closely resembles the actual production environment. This simulation environment realistically recreates the equipment, physical conditions, and external environmental factors involved in the production process, accurately simulating the operation of machinery and equipment, energy transfer, and the overall coordination of the production line. This simulation not only simulates operating conditions in various production scenarios but also allows for testing production processes under different conditions by adjusting parameters.

[0009] Data generation is achieved by precisely placing virtual sensor models within the simulation environment. These models represent the various sensors used in the neural network acceleration task environment, such as visible light sensors, infrared sensors, and lidar. Users can flexibly adjust the number and placement of sensors within the virtual environment according to specific needs, and can even place them on dynamic objects (such as a moving car). These sensors capture data consistent with the input to the accelerator chip in the task environment during simulation, generating highly realistic test data and providing reliable input for subsequent neural network accelerator performance evaluation. These sensors accurately capture data consistent with the input to the accelerator chip in the task environment. Leveraging advanced sensing technology and efficient data collection mechanisms, the sensors can perceive various data changes in the environment in real time and ensure that the characteristics of the collected data closely match those of the input to the accelerator chip. Based on this collected data, the sensors can then generate highly realistic test data. This test data serves as input for subsequent neural network accelerator performance evaluation, providing a realistic, reliable, and valuable foundation for the evaluation.

[0010] The data transmission interface, capable of simulating real sensors, is designed based on a broadcast mechanism. It stores sensor-collected data in a data cache along with a timestamp. Multiple receivers then read the latest data from the cache based on the timestamp, accurately simulating the data transmission behavior of sensors in a mission environment. This interface reproduces data transmission delays, bandwidth limitations, and concurrent processing, ensuring that each receiver receives the same data stream simultaneously in the virtual simulation. This provides a more realistic simulation environment for testing the performance of neural network accelerator chips in multi-tasking parallel and asynchronous computing scenarios.

[0011] The present invention provides a test data generation method for neural network acceleration chip performance testing in a multi-task parallel scenario based on a virtual simulation environment. The method also includes a data transmission interface that simulates the behavior of real sensors.

[0012] 1. This method uses virtual simulation technology to simulate the task working environment of the neural network acceleration chip, solving the problem that the test results of the current test method using public data sets deviate too much from the task application experience.

[0013] 2. This method generates test data by setting up different types and quantities of sensors in a virtual work environment using virtual simulation technology. This overcomes the drawback of current testing methods that use datasets containing only one type of sensor data, and enables the performance evaluation of neural network accelerator chips in a multi-task parallel environment.

[0014] 3. The data transmission interface of this method can simulate the data transmission behavior of real sensors, making the test data set closer to the task sensor, and bringing the ability to evaluate the performance of neural network acceleration chips under asynchronous data acquisition and calculation conditions.

[0015] The present invention provides a method for generating neural network acceleration chip capability test data based on virtual simulation, which is completed through scene parameter planning, basic scene generation, environmental data superposition and scene data fusion.

[0016] Scenario parameter planning

[0017]

[0018] Basic scene generation

[0019] First, import the basic terrain (grassland, city, etc.), and add buildings as the basic scene in the empty 3D terrain according to the parameters entered above.

[0020] Environmental data overlay

[0021] Use Unity's weather system to generate rain, snow and fog according to scene parameters.

[0022] Scene data fusion

[0023] Overlay rain, snow and fog onto the base scene.

[0024] The architecture of the test platform of the present invention is mainly divided into the data layer, platform layer and application layer. The functions of each module in the application layer are described as follows:

[0025] Environment Generation: Once the system receives input criteria, such as a specific location, weather conditions, and a clear test subject, it immediately initiates the generation of a virtual test 3D scene. For location information, the system quickly links to a high-precision geographic information database, accurately extracting key data such as the location's latitude, longitude, and altitude. Using this data, the system uses digital elevation model (DEM) data and interpolation algorithms to transform discrete terrain elevation points into continuous landforms. For example, it constructs rolling mountain ranges and gentle plains based on elevation data. For urban buildings, the system references urban planning data and selects appropriate 3D models from a rich library of architectural models for instantiation, taking into account architectural style and detail, precisely placing them according to the city's layout. Road generation combines topography and planning to determine its course, while river generation uses water system distribution and flow simulation algorithms to determine its source and direction. In the meteorological environment overlay phase, based on pre-input meteorological conditions, the system utilizes a lighting model, integrating time, season, and location latitude and longitude information, to accurately calculate the sun's position and angle, thereby simulating sunlight effects at different times of day and seasons. When simulating rainfall, a particle system is used to generate raindrop particles, with the size, density, falling speed, and collision effects set to create a rainfall scene. Simulating snowfall uses a particle system to control the trajectory of snowflakes and take into account the effects of snow accumulation. Simulating foggy days adjusts atmospheric scattering parameters to evenly distribute fog particles throughout the scene, creating fog effects of varying concentrations and realistically recreating various weather conditions, creating a highly realistic virtual 3D test scene for the test subject. Based on the input conditions (location, weather, test subject, etc.), the corresponding virtual 3D test scene is loaded and generated, including topography, urban buildings, roads, rivers, etc., and the corresponding meteorological environment is superimposed on this to simulate various weather conditions such as sunlight effects, rain, snow, and fog.

[0026] Scene fusion: After completing the construction of the basic environment, the system loads the corresponding moving or stationary objects from a rich resource library based on the test object type. If the test object is an aircraft, various drone, helicopter and other models will be loaded and given corresponding flight properties; for moving vehicles, cars, trucks, etc. will be loaded and the power parameters will be set. For obstacles, building ruins, roadblocks, etc. are placed according to the scene requirements. Subsequently, the system uses navigation and dynamic algorithms based on pre-set properties to drive the moving object along the specified route, and adjusts the speed in real time according to the scene conditions, such as aircraft according to the flight stage and vehicles according to the road conditions, to achieve accurate and realistic dynamic simulation. On top of the basic environment, moving or stationary objects such as aircraft, moving vehicles, obstacles, etc. are loaded and integrated according to the test object type, and the specified object is driven to move along the specified route and speed according to the pre-set properties.

[0027] Sensor Data Simulation: Within the virtual environment, a rich and accurate set of sensor models is built in to comprehensively simulate various sensor performance parameters. For visual sensors, the field of view (FOV) can be flexibly adjusted from ultra-wide to narrow to suit different scenario requirements. The detection range can also be set based on the actual application scenario, enabling both close-range detection and long-range perception. Furthermore, the acquisition rate can be finely controlled, enabling both high-frequency rapid acquisition and low-frequency stable acquisition. Generated sensor data is quickly and stably transmitted to the designated receiver using efficient network transmission protocols. Furthermore, the system keenly captures target objects in the scene and automatically generates reference images for intelligent algorithm detection. These intuitive and clear reference images are perfect for visual demonstrations, helping users intuitively understand the operation of intelligent algorithms. Sensor performance parameters (FOV, detection range, acquisition rate, etc.) are simulated, sensor data is generated, and transmitted over the network to the designated receiver. The system automatically generates reference images for intelligent algorithm detection based on target objects in the scene for visual demonstrations. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a structural block diagram for testing neural network acceleration chips in multi-task parallel and asynchronous computing scenarios.

[0029] Figure 2 This is the main structure diagram of the sensor data generation module.

[0030] Figure 3 This is a schematic diagram of the LOD method based on the quadtree.

[0031] Figure 4 It is a visible light sensor imaging model.

[0032] Figure 5 It is an architectural diagram of test data generation of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and examples. The parameters listed are merely exemplary embodiments of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention.

[0034] See also Figure 1 、 Figure 2 As shown, the present invention is a method for generating data for testing the capability of a neural network acceleration chip based on virtual simulation. Taking an image sensor as an example, the method uses graphics rendering technology to realize data generation for a visible light camera sensor, and utilizes high-performance rendering algorithms (such as ray tracing algorithms, radiosity algorithms, image-based rendering algorithms, etc.) to avoid frame skipping or memory exhaustion and crashes, so as to ensure the stability of the data simulation and testing process. Since any polygon can ultimately be decomposed into triangles that are easy to calculate and represent, the project team uses polygons and NURBS curves to realize the abstraction of scene materials in a three-dimensional rendering engine, and grades the degree of terrain rendering details in the scene, i.e., more close-up details and fewer distant details, thereby further reducing the rendering burden and improving the stability of the test.

[0035] In the sensor data generation module, the project team used different methods to render the sensor viewpoint scene based on the imaging principles of different sensors, that is, the data characteristics, and the scene characteristics. To ensure the imaging precision of the close-up part of the viewpoint and render the sensor viewpoint image under different visible light conditions, the project team used LOD (multi-level of detail) rendering technology based on quadtree. The workflow is as follows: Figure 3 As shown in the figure, graphics rendering technology enables data generation from visible light camera sensors. High-performance rendering algorithms are used to avoid frame skipping or memory exhaustion crashes, ensuring the stability of the data simulation and testing process. Since any polygon can ultimately be decomposed into easily calculated and represented triangles, the project team used polygons and NURBS curves to abstract scene material within the 3D rendering engine. The team also graded the level of detail in the scene's terrain rendering, with more detail in the foreground and less detail in the background. This further reduced the rendering burden and improved test stability.

[0036] The basic working principle of quadtree: Quadtree is a tree-like data structure used to organize two-dimensional spatial data. In scene rendering, starting from the root node of the entire scene, it is divided into four equal sub-areas. Each sub-area becomes a sub-node, and the recursive division continues until the preset stopping condition is reached. Different nodes of the quadtree are given different levels of detail based on factors such as their distance from the viewpoint and their importance in the scene. Usually, nodes that are close to the viewpoint or in critical positions use high LOD, and their models contain more geometric details and high-precision textures; nodes that are far away from the viewpoint or relatively unimportant use low LOD, and their models are relatively simplified.

[0037] The viewing cone is the range of space visible to the observer. When rendering, the system only renders the quadtree nodes within the viewing cone. Let the near clipping plane distance of the viewing cone be n, the far clipping plane distance be f, the horizontal viewing angle be α, and the vertical viewing angle be β. For a quadtree node, if its center coordinates are (x, y, z), we can determine whether it is within the viewing cone by the condition Only nodes that meet all these conditions will be considered for rendering, thus reducing unnecessary rendering calculations.

[0038] When constructing a quadtree model, it is required that the size of each terrain block of the spatial terrain segmentation meets (2n+1)(2n+1), and the insufficient part is padded with null values. The field of view of the visible light sensor of the constructed quadtree node is the part represented by pink, and the squares in the terrain grid represent the nodes of the quadtree. Each constructed quadtree node stores information about a certain area in the terrain, such as the height of the center point of the area, the coordinates of the spatial point, etc. A two-dimensional array is used to store the data information of the full-resolution terrain, so that the terrain data stored in the quadtree node can be directly reflected in the array. At the same time, in order to record the node status of the quadtree data structure, a flag array of the same size as the terrain array is established. In order to save resources and determine whether the node needs to be further divided, the following expression is used in combination with the sensor viewpoint distance to achieve the effect of "more details in the near field and less details in the distant field"

[0039] In the figure, I represents the distance between the node center and the viewpoint, and d represents the size of the node itself. When the above equation holds true, the node needs to be segmented further. C is a controllable variable; increasing its value increases detail; conversely, decreasing detail decreases. Furthermore, to ensure a higher level of detail in areas with more rugged terrain while minimizing the waste of primitives in flat areas, the project team evaluated roughness using the following formula.

[0040] r is the ratio of the node's maximum elevation to its size, and C2 is the roughness adjustment factor. A larger C2 indicates a higher level of detail. Combining the above formulas, we obtain the following final evaluation formula. When the following formula is met, the node needs to be segmented further.

[0041] Nodes that are not further split will be rendered. At this stage, the project team uses a recursive algorithm and a triangle fan method to achieve terrain rendering and node drawing.

[0042] A triangle fan consists of a common vertex and a series of vertices around it. Assume there is a common vertex V0, and a vertex sequence V1, V2, ..., V n . Starting from the common vertex V0, the adjacent vertices V i and V i+1 (i=1,2,…,n+1) form a triangle, so that a complex polygon can be constructed with fewer vertex definitions, which greatly reduces the amount of data compared to defining each triangle vertex one by one. When performing graphic transformations such as translation, rotation, and scaling on the triangle fan, the vertex coordinates need to be transformed accordingly. Taking rotation transformation as an example, if the rotation angle around the z axis is θ, the transformation matrix is Vertex V i The transformed coordinates are The transformation formula is used to flexibly adjust the position, direction and size of the triangle fan according to different scene requirements.

[0043] In the process of projecting three-dimensional points onto the imaging plane, the project team adopted the pinhole imaging model, and its calculation formula is:

[0044] In the above formula, a spatial point P(X, Y, Z) in the camera coordinate system is imaged on the physical imaging plane through the sensor's optical center. Scaling and offsetting then yield a point p(u, v) on the pixel plane. fx and fy are the imaging scale parameters, cx and cy are the imaging offset parameters, and K is the camera's intrinsic parameter matrix. This model formula, combined with the spatial point information in the scene, generates a generated image from the sensor's perspective.

[0045] Based on the above analysis, the correspondence between 3D points and pixels has been established. The process of generating visible light data is to determine the visible light value of each pixel. 3D model data can be thought of as a simple 3D array of cubes, each of which is called a voxel. Each voxel represents a unit of space. Each unit stores the optical model for that point in 3D space, which mainly includes the following types. Absorption-only: The volume is assumed to be composed of cold, black particles that completely absorb all light that impinges on it. They neither emit nor scatter light. Emission-only: The volume is assumed to be composed of particles that emit light but do not absorb it, as absorption is negligible. Absorption + Emission: This optical model is the most common one in direct voxel rendering. Particles emit light and also block (absorb) incoming light. However, there is no scattering or indirect light. Scattering and Shading / Shading: This model includes scattering from external lighting sources. Scattered light can be assumed to be unaffected by distant light sources and can also be blocked by particles between the considered light source and the voxel. Multiple Scattering: This complex model supports indirect light that can be scattered by multiple particles.

[0046] The test platform includes a virtual zz task software platform and a chip test hardware and software platform, which are deployed on two different computing servers. The computing servers establish a web socket connection through the local area network for data communication. Figure 5 The main functions of the zz task platform are to generate test scenarios, test data, and display performance indicators. Test scenario generation is mainly divided into the scenario environment and the modeling of the intelligent unmanned system. The required test scenario environment can be generated by setting parameters in the configuration file. The intelligent unmanned system is equipped with a motion control software interface to connect to the parallel task load in the neural network acceleration chip.

[0047] The architecture of the test platform is mainly divided into the data layer, platform layer and application layer. The functions of each module in the application layer are described as follows:

[0048] Environment Generation: Based on input conditions (location, weather, test object, etc.), the system loads and generates a corresponding virtual test 3D scene, including topography, urban buildings, roads, and rivers. It then overlays the corresponding meteorological environment to simulate various weather conditions, such as sunlight effects, rain, snow, and fog. Once the system receives input conditions, such as a specific location, specific weather conditions, and a clear test object, it immediately initiates the generation of the virtual test 3D scene. For location information, the system quickly links it to a high-precision geographic information database, accurately extracting key data such as the location's latitude, longitude, and altitude. Using this data, the system uses digital elevation model (DEM) data and interpolation algorithms to convert discrete terrain elevation points into continuous topography. For example, it can construct rolling mountains and gentle plains based on elevation data. For urban buildings, urban planning data is referenced, and appropriate 3D models are instantiated from a rich library of architectural models, taking into account architectural style and details, and accurately placed according to the city layout. Road generation combines topography and planning to determine its course, while river generation uses water system distribution and flow simulation algorithms to determine its source and flow direction. In the meteorological environment overlay phase, based on pre-entered meteorological conditions, the lighting model integrates time, season, and location latitude and longitude information to accurately calculate the sun's position and angle, thereby simulating sunlight effects at different times and seasons. When simulating rainfall, a particle system is used to generate raindrop particles, setting the size, density, falling speed, and collision effects to create a rainfall scene. Snowfall simulation uses a particle system to control the falling trajectory of snowflakes and take into account the effect of snow accumulation. Fog simulation adjusts atmospheric scattering parameters to evenly distribute fog particles in the scene, creating fog effects of varying concentrations, realistically recreating various weather environments and creating highly realistic virtual test 3D scenes for test subjects.

[0049] Scene fusion: Based on the basic environment, moving or stationary objects such as aircraft, moving vehicles, and obstacles are loaded and integrated according to the test object type, and the specified object is driven to move along the specified route and speed according to the pre-set attributes. After completing the construction of the basic environment, the system loads the corresponding moving or stationary objects from a rich resource library based on the test object type. If the test object is an aircraft, various drones, helicopters and other models will be loaded and given corresponding flight attributes; for moving vehicles, cars, trucks, etc. will be loaded and power parameters will be set. For obstacles, building ruins, roadblocks, etc. are placed according to the scene requirements. Subsequently, the system uses navigation and dynamic algorithms based on the pre-set attributes to drive the moving object along the specified route, and adjusts the speed in real time according to the scene conditions, such as aircraft according to the flight stage and vehicles according to the road conditions, to achieve accurate and realistic dynamic simulation.

[0050] Sensor Data Simulation: Sensor performance parameters (field of view, detection range, acquisition rate, etc.) are simulated, sensor data is generated, and transmitted over the network to a designated receiver. The system automatically generates reference images for intelligent algorithm detection based on target objects in the scene for visual demonstration. A rich and accurate sensor model is built into the virtual environment to comprehensively simulate various sensor performance parameters. For visual sensors, the field of view can be flexibly adjusted from ultra-wide to narrow to accommodate diverse scenario requirements. The detection range can also be set based on the actual application scenario, enabling both close-range detection and long-range perception. Furthermore, the acquisition rate can be finely controlled, enabling both high-frequency rapid acquisition and low-frequency stable acquisition. The generated sensor data is quickly and stably transmitted to the designated receiver using efficient network transmission protocols. Furthermore, the system keenly captures target objects in the scene and automatically generates reference images for intelligent algorithm detection. These intuitive and clear reference images are perfect for visual demonstrations, helping users intuitively understand the operation of the intelligent algorithms.

[0051] The method of the present invention sets up multiple real-world combat missions for unmanned equipment and establishes a multi-task parallel test program set for each mission to more comprehensively obtain performance data for the neural network acceleration chip in 1) different scenarios and 2) different tasks. Furthermore, the test dataset generation module is divided into sensor data generation and sensor data transmission. Sensor data is simulated using the Unity rendering engine and transmitted via web socket broadcast. Furthermore, the project team has prepared a variety of indicator collection and calculation methods for the system to meet the needs of testing neural network acceleration chips in different parallel mission scenarios. roadList is an integer array of arbitrary length, with elements ranging from 0 to 6, representing the seven different road sections mentioned above. All road sections specified in roadList constitute the test environment for the entire test process. roadParamList is an array of the same length as roadList, with each element containing the value of an adjustable parameter for the corresponding road section in the roadList array. roadWeight is an array of floating-point numbers corresponding to the contribution of each road section in the entire environmental adaptability test process. The elements in roadWeight sum to 1.0. lightList is an array of floating-point numbers, corresponding to the light intensity of each road section in roadList, with a data range of 0.0-1.0, where 0.0 is the lowest light intensity and 1.0 is the highest light intensity. SnowList is an array of floating-point numbers, corresponding to the snowfall degree of each road section in roadList, with a data range of 0.0-1.0, where 0.0 is the lowest snowfall and 1.0 is the highest snowfall. RainList is an array of floating-point numbers, corresponding to the rainfall amount of each road section in roadList, with a data range of 0.0-1.0, where 0.0 is the lowest rainfall and 1.0 is the highest rainfall. FogList is an array of floating-point numbers, corresponding to the fog concentration of each road section in roadList, with a data range of 0.0-1.0, where 0.0 is the lowest fog intensity and 1.0 is the highest fog intensity.

[0052] To address the inability of existing neural network accelerator chip testing methods to evaluate the multi-mission capabilities of military neural network accelerator chips, this study investigated the operating environment of neural network accelerator chips for unmanned systems and constructed a task-driven, multi-task parallel test suite tailored to future mission requirements. Real-world combat scenarios were simulated using mission scenario generation, unmanned system simulation, and sensor simulation technologies. Typical applications were developed based on combat scenarios and unmanned system types, and a test suite was developed in conjunction with existing neural network algorithms. Finally, metrics were designed to evaluate the parallel computing capabilities of neural network accelerator chips, which were collected and displayed via a dedicated communication interface.

[0053] This method overcomes the drawback of previous neural network accelerator chip evaluation methods, which were unable to assess the chip's multitasking capabilities. It also employs sensor simulation, replacing previous methods that used public datasets, improving the authenticity of sensor data. This method effectively quantifies the task switching capabilities of neural network accelerator chips, the multi-core sharing capabilities of processing units, the software stack's algorithm support, the execution efficiency of operator libraries, and the level of compiler optimization, providing a comprehensive evaluation. This provides technical support for the practical deployment of neural network accelerator chips in intelligent unmanned equipment.

Claims

1. A method for generating data for testing the capability of a neural network acceleration chip based on virtual simulation, which is accomplished through scenario parameter planning, basic scenario generation, environmental data superposition, and scenario data fusion; characterized by: Data generation is achieved by accurately placing virtual sensor models in the simulation environment. These models can represent the various sensors used in the working environment of the neural network acceleration task. Users can flexibly adjust the number and location of sensors in the virtual environment according to specific needs and set them on dynamic objects; these sensors can capture data consistent with the input accelerator chip in the task working environment in simulation, thereby generating highly realistic test data.

2. The method for generating neural network acceleration chip capability test data based on virtual simulation according to claim 1, characterized in that: Sensor data generation utilizes graphics rendering technology to generate data from visible light camera sensors. High-performance rendering algorithms are employed to avoid frame skipping and memory exhaustion crashes, ensuring stability during data simulation and testing. Since any polygon can ultimately be decomposed into easily calculated and represented triangles, the project team employed polygons and NURBS curves to abstract scene material within the 3D rendering engine. The team also implemented a gradation of the level of detail in the scene's terrain rendering, with more detail in the foreground and less detail in the background. This further reduced the rendering burden and improved test stability.

3. The method for generating neural network acceleration chip capability test data based on virtual simulation according to claim 1, characterized in that: roadList is an integer array of arbitrary length, with elements between 0 and 6, representing the seven different road sections mentioned above. All the road sections set in roadList constitute the test environment for the entire test process. roadParamList is an array of the same length as roadList, with each element being the value of the adjustable parameters in the corresponding road section in the roadList array. roadWeight is an array of floating-point numbers, corresponding to the proportion of each road section in roadList in the entire environmental adaptability test process. The sum of the elements in roadWeight is 1.

0. lightList is an array of floating-point numbers, corresponding to the light intensity of each road section in roadList, with a data range of 0.0-1.0, where 0.0 indicates the lowest light intensity and 1.0 indicates the highest light intensity. SnowList is an array of floating-point numbers, corresponding to the degree of snowfall for each road section in roadList, with a data range of 0.0-1.0, where 0.0 indicates the lowest snowfall and 1.0 indicates the highest snowfall. RainList is an array of floating-point numbers corresponding to the rainfall for each road section in roadList, with a range of 0.0-1.0, where 0.0 is the lowest rainfall and 1.0 is the highest rainfall. FogList is an array of floating-point numbers corresponding to the fog density for each road section in roadList, with a range of 0.0-1.0, where 0.0 is the lowest fog intensity and 1.0 is the highest fog intensity.

4. A test platform constructed using the capability test data of the neural network acceleration chip according to claim 1, characterized in that: The test platform architecture is mainly divided into the data layer, platform layer and application layer; the functions of each module in the application layer are described as follows: Environment generation: Load and generate the corresponding virtual test 3D scene based on input conditions (location, weather, test object, etc.), including landforms, urban buildings, roads and rivers, etc., and superimpose the corresponding meteorological environment on this basis to simulate various weather environments such as sunlight effects, rain, snow, fog, etc. Scene fusion: Based on the basic environment, moving or stationary objects such as aircraft, vehicles, and obstacles are loaded and integrated according to the test object type, and the specified objects are driven to move along the specified route and speed according to pre-set properties; Sensor data simulation: Sensor performance parameters (field of view, detection distance, acquisition rate, etc.) are simulated, sensor data is generated, and transmitted to the designated receiving end through the network; the system automatically generates reference images for detecting intelligent algorithms based on the target objects in the scene for visual demonstration.

Images