Convolution operation parallel computing and result aggregation method, system, equipment and medium

Abstract

The invention discloses a convolution operation parallel computing and result aggregation method, system and device and a medium, relates to the technical field of deep learning hardware acceleration, is applied to a terminal, and is characterized in that a convolution neural network model is arranged in the terminal, and the convolution operation parallel computing and result aggregation method comprises the following steps: obtaining at least one input image; dividing the input image into a plurality of block images according to convolutional layer parameters of a convolutional neural network model CNN; the plurality of block images are processed in parallel according to a plurality of convolution operator units in the CNN, and the block images processed by the convolution operator units are different from one another; and performing multi-layer aggregation processing on a processing result of processing the block image by each convolution operator unit to obtain a convolution result, and outputting the convolution result as a prediction result. The technical problem of low utilization rate of multi-task parallel computing resources during image processing is solved.

Description

Technical Field

[0001] The present application relates to the technical field of deep learning hardware acceleration, and in particular to methods, systems, devices, and media for parallel computing and result aggregation of convolution operations. Background Art

[0002] In the field of deep learning, convolutional neural networks have attracted widespread attention due to their outstanding performance in image recognition and processing tasks. The core of convolutional neural networks is the convolution operation, which extracts features by performing dot products between the convolution kernel and the local area of ​​the input data. It is of great significance for high-dimensional data such as images. However, with the growth of data scale and the increase of model complexity, the demand for computing resources for the convolution operation of convolutional neural networks has increased dramatically. When performing these computationally intensive operations, traditional hardware architectures are limited by the fixed nature of physical connections and resource allocation, making it difficult to achieve efficient computing, resulting in slow processing speed and low resource utilization. Summary of the invention

[0003] The main purpose of this application is to provide a method, system, device, and medium for parallel computing and result aggregation of convolution operations, aiming to solve the technical problem of low utilization of multi-task parallel computing resources during image processing.

[0004] To achieve the above purpose, the present application proposes a method for parallel computing and result aggregation of convolution operations, which is applied to a terminal. A convolutional neural network model is set in the terminal, including:

[0005] obtaining at least one input image;

[0006] According to the convolution layer parameters of the convolutional neural network model CNN, the input image is divided into multiple block images;

[0007] Processing a plurality of block images in parallel according to a plurality of convolution operator units in the CNN, wherein the block images processed by each convolution operator unit are different from each other;

[0008] The processing results of the block images processed by each convolution operator unit are subjected to multi-layer aggregation processing to obtain a convolution result, and the convolution result is output as a prediction result.

[0009] In one embodiment, the step of dividing the input image into a plurality of block images according to the convolution layer parameters of the convolutional neural network model CNN includes:

[0010] Determine the convolution parameters of any convolution operator unit in the CNN, wherein the convolution parameters include the convolution kernel size and stride;

[0011] The input image is divided according to the image size and convolution parameters of the input image to obtain a task decomposition matrix, wherein each region in the task decomposition matrix includes a block image after the input image is divided.

[0012] In one embodiment, the step of performing multi-layer aggregation processing on the processing results of the block images processed by each convolution operator unit includes:

[0013] Determine the aggregation path in CNN, optimize the aggregation path according to the configuration matrix, and obtain the target aggregation path;

[0014] The adder in CNN is used to aggregate the processing results to obtain the first aggregation result, and multi-level aggregation processing is performed according to the target aggregation path and the first aggregation result to obtain the final aggregation result, and the final aggregation result is used as the convolution result. Among them, for the adder of each layer, the aggregation result of the current layer is obtained according to the target aggregation path and the aggregation result of the previous layer, and the aggregation result of the current layer is transmitted to the adder of the next layer through the target aggregation path.

[0015] In one embodiment, before the step of processing a plurality of block images in parallel according to a plurality of convolution operator units in a CNN, the method includes:

[0016] Determine a configuration matrix according to the task decomposition matrix and a plurality of convolution operator units in the CNN, wherein the configuration matrix includes a corresponding relationship between each block image in the task decomposition matrix and the convolution operator unit;

[0017] According to the configuration matrix, each block image in the task decomposition matrix is ​​allocated to its corresponding convolution operator unit.

[0018] In one embodiment, the step of allocating each block image in the task decomposition matrix to the corresponding convolution operator unit according to the configuration matrix includes:

[0019] If the input feature map is a multi-channel feature map, the task load of the multi-channel feature map is determined according to the number of channels and convolution parameters of the multi-channel feature map;

[0020] A distribution matrix is ​​determined according to the task load, wherein the distribution matrix is ​​used to control the calculation allocation of each channel feature map.

[0021] In one embodiment, the method for parallel computing and result aggregation of convolution operations further includes:

[0022] Calculate the sum of the loads of all convolution operator units;

[0023] When the sum of the loads is greater than a preset high load threshold, the configuration matrix is ​​updated by increasing the number of convolution operator units;

[0024] When the sum of the loads is less than a preset low load threshold, the configuration matrix is ​​updated to reduce the number of convolution operator units.

[0025] In one embodiment, after the step of outputting the convolution result as the prediction result, the following steps are included:

[0026] Determine the utilization and corresponding weight of each convolution operator unit;

[0027] The configuration effect is obtained by performing weighted summation based on the utilization rate and weight of each convolution operator unit;

[0028] If the configuration effect is lower than the preset effect threshold, the convolution operator units in the configuration matrix are reallocated.

[0029] In addition, to achieve the above purpose, the present application also proposes a convolution operation parallel computing and result aggregation system, which includes:

[0030] A data acquisition module acquires at least one input image;

[0031] The image decomposition module divides the input image into multiple block images according to the convolution layer parameters of the convolutional neural network model CNN;

[0032] A convolution operator module processes a plurality of block images in parallel according to a plurality of convolution operator units in the CNN, wherein the block images processed by each convolution operator unit are different from each other;

[0033] The result aggregation module performs multi-layer aggregation processing on the processing results of the block image processed by each convolution operator unit to obtain the convolution result, and outputs the convolution result as the prediction result.

[0034] In addition, to achieve the above-mentioned objectives, the present application also proposes a device for parallel computing and result aggregation of convolution operations, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the method for parallel computing and result aggregation of convolution operations as described above.

[0035] In addition, to achieve the above-mentioned purpose, the present application also proposes a medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the convolution operation parallel calculation and result aggregation method as described above are implemented.

[0036] In addition, to achieve the above-mentioned purpose, the present application also provides a product, which is a computer program product. The computer program product includes a computer program. When the computer program is executed by a processor, it implements the steps of the convolution operation parallel calculation and result aggregation method as described above.

[0037] One or more technical solutions proposed in this application have at least the following technical effects:

[0038] The present application is applied to a terminal, in which a convolutional neural network model is set up to obtain at least one input image. By optimizing the task scheduling strategy, the image preprocessing time is effectively reduced and the data input efficiency is improved. According to the convolution layer parameters of the convolutional neural network model CNN, the input image is divided into multiple block images, and the parallel processing preparation of the image data is realized, which lays the foundation for subsequent parallel computing. Multiple block images are processed in parallel according to multiple convolution operator units in CNN, wherein the block images processed by each convolution operator unit are different from each other. This parallel processing mechanism significantly improves the computing efficiency, so that large-scale convolutional neural networks can make full use of hardware resources. Multi-layer aggregation processing is performed on the processing results of the block images processed by each convolution operator unit to obtain the convolution result, and the convolution result is output as the prediction result. In summary, the present application solves the technical problem of low utilization of multi-task parallel computing resources during image processing, and improves the performance and response speed of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a flow chart of the first embodiment of the method for parallel calculation and result aggregation of convolution operations of the present application;

[0042] Figure 2 This is another flowchart of the method for parallel computing and result aggregation of convolution operations in this application;

[0043] Figure 3 This is another flowchart of the method for parallel calculation and result aggregation of convolution operations in this application;

[0044] Figure 4 This is a schematic diagram of the module structure of the convolution operation parallel computing and result aggregation system in the embodiment of the present application;

[0045] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the convolution operation parallel computing and result aggregation method in the embodiment of the present application.

[0046] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0047] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0048] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0049] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a terminal system, etc. The following takes the system as an example to illustrate this embodiment and the following embodiments.

[0050] Based on this, this embodiment provides a method for parallel computing and result aggregation of convolution operations, referring to Figure 1 , Figure 1 This is a flow chart of the method for parallel calculation and result aggregation of convolution operations of the present application. The method for parallel calculation and result aggregation of convolution operations includes steps S10 to S40:

[0051] Step S10, obtaining at least one input image;

[0052] It should be noted that the input image refers to the original image data to be processed, which is the starting input for the convolutional neural network to perform tasks such as image recognition and classification.

[0053] Obtain the image data to be processed from the data source (such as a camera, image library, etc.). As the starting point for the convolutional neural network model processing, the input image is the basis for all subsequent steps.

[0054] Step S20, dividing the input image into a plurality of block images according to the convolution layer parameters of the convolutional neural network model CNN;

[0055] It should be noted that CNN (Convolutional Neural Networks) is a deep learning model that is particularly suitable for image processing tasks. It extracts features and classifies images through structures such as convolutional layers, pooling layers, and fully connected layers. Convolutional layer parameters refer to the relevant configuration information of the convolutional layer in CNN, such as convolution kernel size, stride, padding method, etc. These parameters determine the specific method of convolution operation. Block image is a smaller image block obtained by dividing the input image according to certain rules. Each block image is the basic unit processed by the convolution operator unit.

[0056] In this step, the system divides the input image into multiple small blocks or sub-regions (block images) according to the convolution layer parameters of the CNN model (such as convolution kernel size, stride, etc.). The purpose of division is to facilitate subsequent parallel processing and improve computational efficiency. The division method may vary depending on the convolution layer parameters. For example, the size and position of each block image are determined according to the convolution kernel size and stride.

[0057] For example, when a 3x3 convolution kernel performs convolution on a 256x256 input feature map, the system divides the input into small blocks of 16x16 to ensure that the block size processed by each convolution operator unit is within the computing power of the convolution operator unit.

[0058] Step S30, processing a plurality of block images in parallel according to a plurality of convolution operator units in the CNN, wherein the block images processed by each convolution operator unit are different from each other;

[0059] It should be noted that the convolution operator unit is the basic unit responsible for performing convolution operations in CNN, which can process multiple block images in parallel to improve computing efficiency.

[0060] In the systolic array architecture, multiple convolution operator units can work in parallel. Each convolution operator unit is responsible for processing a different block image. In this way, multiple block images can be processed at the same time, which significantly improves the computational efficiency. The convolution operator unit performs convolution operations on the block image according to the convolution layer parameters of the CNN model to extract the feature information in the image.

[0061] Step S40, performing multi-layer aggregation processing on the processing results of the block images processed by each convolution operator unit to obtain a convolution result, and outputting the convolution result as a prediction result.

[0062] It should be noted that multi-layer aggregation processing is to aggregate and integrate the processing results of multiple convolution operator units to obtain the final convolution result. This step usually involves hardware structures such as adders and multipliers. The prediction result is the final result processed and output by CNN, which is usually used for the final judgment of tasks such as image classification and recognition.

[0063] After all convolution operator units complete their respective convolution operations, the system needs to aggregate these output results. The aggregation process may include multiple levels of data integration and addition operations to ensure that the final feature representation of the entire input image is obtained. The design and optimization of the aggregation module is crucial to improving the overall performance and accuracy of the system. For example, by optimizing the aggregation path and reducing data transmission delays, the computational efficiency of the system can be further improved.

[0064] For example, the system designs the convolution operator unit in the systolic array as a multifunctional module that supports variable convolution kernel size and stride. Assuming the convolution kernel size is K x K, the stride is s, and the input feature map size is H x W, the computational effort of a single convolution task is C conv As shown in formula (1):

[0065] C conv =(H-K+1)×(W-K+1)×K 2 Formula (I)

[0066] On this basis, the convolution operation is decomposed into subtasks Ti, each of which consists of a convolution operator unit calculating the area of ​​a specific convolution kernel to ensure maximum utilization of system resources.

[0067] This embodiment is applied to a terminal, in which a convolutional neural network model is set to obtain at least one input image. By optimizing the task scheduling strategy, the image preprocessing time is effectively reduced and the data input efficiency is improved. According to the convolution layer parameters of the convolutional neural network model CNN, the input image is divided into multiple block images, and the parallel processing preparation of the image data is realized, which lays a foundation for subsequent parallel computing. Multiple block images are processed in parallel according to multiple convolution operator units in CNN, wherein the block images processed by each convolution operator unit are different from each other. This parallel processing mechanism significantly improves the computing efficiency, so that large-scale convolutional neural networks can make full use of hardware resources. Multi-layer aggregation processing is performed on the processing results of the block images processed by each convolution operator unit to obtain the convolution result, and the convolution result is output as the prediction result. In summary, this embodiment solves the technical problem of low utilization of multi-task parallel computing resources during image processing, and improves the performance and response speed of the overall system.

[0068] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above introduction, and will not be repeated in the following. Figure 2 The step S20 further includes steps A10 to A20:

[0069] Step A10, determining the convolution parameters of any convolution operator unit in the CNN, wherein the convolution parameters include the convolution kernel size and stride;

[0070] It should be noted that convolution parameters refer to the specific properties that define the convolution operation, including the convolution kernel size (also called filter size) and stride. The convolution kernel size determines the size of the input image area covered by each convolution operation, while the stride determines the distance the convolution kernel moves on the input image.

[0071] First, the convolution parameters need to be determined for each convolution operator unit in the CNN. These parameters are usually determined based on the design requirements of the CNN and the characteristics of the input image. For example, a 3x3 convolution kernel size and a stride of 1 can be selected for feature extraction. The choice of convolution parameters has an important impact on the performance and accuracy of CNN. The convolution kernel size determines the granularity of feature extraction. Smaller convolution kernels can capture finer features, while larger convolution kernels can capture more global features. The stride determines the density of feature extraction. A smaller stride can produce more feature maps, but it will also increase the amount of computation.

[0072] Step A20, dividing the input image according to the image size and convolution parameters of the input image to obtain a task decomposition matrix, wherein each region in the task decomposition matrix includes a block image after the input image is divided.

[0073] It should be noted that the task decomposition matrix is ​​a matrix used to manage and organize the block images obtained after the input image is divided. In the task decomposition matrix, each area corresponds to a block image, and these block images will be processed as independent processing units by different convolution operator units.

[0074] After determining the convolution parameters of the convolution operator unit, the next step is to divide the input image according to the image size of the input image and these parameters. This process involves calculating the size and position of each block image to ensure that they can fully cover the input image and meet the convolution parameter requirements of the convolution operator unit. These block images are then organized into a task decomposition matrix, where each area corresponds to a block image. The division of the input image is to decompose the complex convolution task into smaller and more manageable subtasks. Each subtask is processed independently by a convolution operator unit, which can achieve parallel computing and improve computational efficiency. The task decomposition matrix provides a convenient way to manage and organize these subtasks so that they can be assigned to different convolution operator units for processing in an orderly manner.

[0075] Exemplarily, a task decomposition matrix M_{T} is defined to allocate input feature map regions to different convolution operator units. If the feature map is divided into n\times m regions, each regional task is represented by M_{T}(i,j), where i,j represent the region index. The task decomposition matrix works with the configuration matrix C_d(t) to enable the system to automatically adjust the calculation allocation according to the convolution kernel size and stride.

[0076] This embodiment realizes the division and task decomposition of the input image, and provides a basis for subsequent parallel processing. Each convolution operator unit will process one or more block images according to its convolution parameters, so as to extract features in the input image.

[0077] In one possible implementation, reference Figure 3 , before step S20, steps B10 to B20 are also included:

[0078] Step B10, determining a configuration matrix according to the task decomposition matrix and a plurality of convolution operator units in the CNN, wherein the configuration matrix includes a correspondence between each block image in the task decomposition matrix and the convolution operator unit;

[0079] It should be noted that the configuration matrix is ​​a matrix whose elements represent the correspondence between the block images in the task decomposition matrix and the convolution operator units (CPEs) in the CNN. The configuration matrix is ​​used to guide how to assign block images to different CPEs for processing.

[0080] First, an empty configuration matrix is ​​created based on the number of tile images defined in the task decomposition matrix and the number of convolution operator units (CPEs) available in the CNN. Then, each tile image is assigned to a CPE based on certain strategies (such as load balancing, task priority, etc.). The result of this assignment process is the element value of the configuration matrix, where each element represents the correspondence between a tile image and a CPE. The process of determining the configuration matrix is ​​actually a task scheduling problem. In CNN, since the computing power and resources of each CPE are limited, it is necessary to reasonably assign tasks (i.e., tile images) to each CPE to ensure maximum computing efficiency and resource utilization. The role of the configuration matrix is ​​to record this assignment relationship and provide guidance for subsequent task execution.

[0081] Step B20, allocating each block image in the task decomposition matrix to its corresponding convolution operator unit according to the configuration matrix.

[0082] Once the configuration matrix is ​​determined, the actual block image allocation can be performed based on it. Specifically, it is to traverse each element of the configuration matrix and send the corresponding block image to the corresponding CPE for processing according to the element value (i.e. the index of the CPE). This step is the actual application stage of the configuration matrix. With the guidance of the configuration matrix, the system can efficiently allocate tasks to each CPE, thereby achieving parallel computing and accelerated processing. This allocation method not only improves computing efficiency, but also ensures the rational use of resources.

[0083] For example, assume that there are 4 tile images in the task decomposition matrix and 3 CPEs in the CNN. Then the configuration matrix may be a 4x3 matrix (or more concisely, an array containing 4 elements, each element pointing to the index of a CPE). Through some strategy (such as polling, priority, etc.), we can assign tile image 1 to CPE1, tile image 2 to CPE2, tile image 3 to CPE3, and tile image 4 to CPE1 again (or to another CPE according to other strategies). In this way, the configuration matrix records this allocation relationship. If the configuration matrix indicates that tile image 1 is assigned to CPE1, the system will send the data of tile image 1 to CPE1 for processing. Similarly, tile image 2 will be sent to CPE2, tile image 3 will be sent to CPE3, and tile image 4 will be reallocated to a CPE (it may be CPE1 or another CPE) according to the instructions of the configuration matrix.

[0084] This embodiment realizes the effective correspondence and allocation between the task decomposition matrix and the convolution operator unit in CNN. This correspondence and allocation method not only improves the computing efficiency, but also ensures the rational use of resources and the parallel processing of tasks.

[0085] In a feasible implementation manner, step B20 further includes steps B201 to B202:

[0086] Step B201, if the input feature map is a multi-channel feature map, determine the task load of the multi-channel feature map according to the number of channels and convolution parameters of the multi-channel feature map;

[0087] It should be noted that in deep learning, especially in convolutional neural networks, feature maps usually have multiple channels, which are called multi-channel feature maps. Each channel represents a different feature of the input image or the output of the previous layer. These channels together constitute the input feature map for subsequent calculations and processing.

[0088] First, it is necessary to analyze the number of channels of the multi-channel feature map, that is, the depth of the feature map. Then, the convolution parameters (such as convolution kernel size, stride, padding, and number of output channels) are combined to calculate the amount of computation required for each channel feature map. This amount of computation can be estimated based on the complexity of the convolution operation and the size of the feature map. The computational complexity of the convolution operation is usually proportional to the size of the feature map, the number of channels, and the size and number of convolution kernels. Therefore, by comprehensively considering these factors, the task load of each channel feature map can be estimated more accurately.

[0089] Step B202, determining a distribution matrix according to the task load, wherein the distribution matrix is ​​used to control the calculation allocation of each channel feature map.

[0090] It should be noted that in parallel computing or distributed systems, task load refers to the amount of computing tasks assigned to each processing unit. In this embodiment, the task load specifically refers to the amount of processing required for each channel in the multi-channel feature map, which depends on the number of channels and the configuration of the convolution parameters. The distribution matrix is ​​used to determine how to assign the computing tasks of the multi-channel feature map to different computing units (such as convolution operator units). Each element of the distribution matrix may represent a correspondence or allocation strategy between a channel feature map and a computing unit.

[0091] After determining the task load of each channel feature map, the next step is to determine the distribution matrix based on the availability of computing resources and the principle of load balancing. Each element of the distribution matrix may represent the correspondence between a channel feature map and a computing unit (such as a convolution operator unit). This correspondence can be determined by a variety of strategies, such as polling, priority scheduling, load balancing algorithms, etc. The process of determining the distribution matrix is ​​actually a problem of task scheduling and resource allocation. In parallel computing or distributed systems, reasonable task scheduling and resource allocation can significantly improve computing efficiency and resource utilization. Therefore, it is necessary to formulate a suitable distribution strategy based on the actual situation of the task load and the availability of computing resources.

[0092] For example, for multi-channel convolution, let the input feature map have C in channels, and the convolution operation of each channel is performed by C out The convolution operation distribution formula is shown in formula (2):

[0093]

[0094] Where T_{ij} represents the computational task of the i convolution kernel on the j channel. conv To calculate the task load, the application layer follows T conv Distribute convolution operations to different convolution operator units to balance the load.

[0095] For multi-channel convolution, a distribution matrix M_{dist} is defined to control the calculation distribution of each channel feature map. The element M_{dist}(p,q) of the distribution matrix M_{dist} represents the calculation convolution operator unit index of the qth convolution block of the pth channel feature map.

[0096] This embodiment achieves accurate estimation of the multi-channel feature map task load and reasonable allocation of computing resources. This distribution strategy not only improves computing efficiency, but also ensures effective utilization of resources and realization of load balancing.

[0097] Based on the first or second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the first or second embodiment can be referred to the above introduction, and will not be described in detail later. Step S40 also includes steps C10 to C20:

[0098] Step C10, determining the aggregation path in the CNN, optimizing the aggregation path according to the configuration matrix, and obtaining the target aggregation path;

[0099] It should be noted that the aggregation path is in the convolutional neural network (CNN). The aggregation path refers to the path or method of merging multiple processing results (such as the output of different convolution kernels) into a single result. It determines how to combine multiple feature maps or feature vectors to form a more representative feature representation. The target aggregation path is the aggregation path after the configuration matrix is ​​optimized. It represents the optimal or desired aggregation method for merging multiple processing results into the final convolution result.

[0100] Before the system starts or each convolution task begins, the configuration matrix is ​​initialized to store the current network configuration and path information. During each layer of aggregation, the aggregation path is automatically optimized through the configuration matrix so that the data is transmitted over the shortest distance during the aggregation process, thereby reducing the delay in data transmission. The optimization process can be achieved by developing an adaptive routing algorithm that can dynamically select the optimal path based on the current network status and data traffic, or by applying graph theory algorithms such as the Dijkstra algorithm or the Bellman-Ford algorithm to calculate the shortest path. Furthermore, machine learning technology can be used to learn network behavior patterns based on historical data, predict and optimize path selection.

[0101] After obtaining the optimal path, the optimal path is mapped to the transmission matrix, which is used to store and update the information of the optimal transmission path and calculate the optimal path in each layer of aggregation, so that the convolution output can be quickly integrated. According to the results and efficiency of data transmission, the configuration matrix is ​​continuously adjusted through the feedback mechanism to optimize future path selection. When the convolution calculation task is completed, the configuration matrix is ​​updated to reflect the new network status in preparation for the next task.

[0102] Step C20, use the adder in CNN to aggregate the processing results to obtain a first aggregation result, and perform multi-level aggregation processing based on the target aggregation path and the first aggregation result to obtain a final aggregation result, and use the final aggregation result as the convolution result, wherein for the adder of each layer, the aggregation result of the current layer is obtained according to the target aggregation path and the aggregation result of the previous layer, and the aggregation result of the current layer is transmitted to the adder of the next layer through the target aggregation path.

[0103] Note that in a CNN, an adder is usually used to perform element - wise addition operations, such as element - by - element addition of feature maps. In this embodiment, the adder is used to aggregate multiple processing results. Multilevel aggregation processing refers to performing aggregation operations at multiple levels or stages. Each level may use different aggregation strategies or paths to gradually construct a more complex feature representation.

[0104] After determining the target aggregation path, the adder in the CNN is used to aggregate each processing result. This usually means adding multiple feature maps or feature vectors element - by - element to form a new feature map or feature vector. After obtaining the first aggregation result, multilevel aggregation processing is performed according to the target aggregation path and the aggregation result of the previous level. This means that in each level, the adder is used to further process the aggregation result of the previous level to form the aggregation result of the current level. Then, this result is transmitted through the target aggregation path to the adder of the next level for the next round of aggregation processing. Multilevel aggregation processing allows the model to gradually construct a more complex feature representation. In each level, the model can use the aggregation result of the previous level and the feature representation of the current level to generate new features. This approach helps the model capture deep - level features in the input data, thereby improving the performance and accuracy of the model.

[0105] Exemplarily, the aggregation process of the convolution result involves hierarchical addition aggregation. The output of the convolution operator unit is gradually passed to higher - level aggregation nodes after being aggregated by the primary adder. If the convolution output is divided into n layers, where n is a positive integer, the aggregation process is shown in formula (III):

[0106]

[0107] where R conv is the final aggregation result, G k is the aggregation output of the k - th layer, where k is a positive integer and 1 < k < n. This hierarchical aggregation ensures the gradual integration of data and reduces the overall transmission delay. The transmission path is optimized through the configuration matrix, making the transmission distance the shortest during the aggregation process. The transmission matrix calculates the optimal path during each layer of aggregation, enabling the rapid integration of the convolution output.

[0108] This embodiment realizes the optimization of the aggregation path and multilevel aggregation processing, thereby improving the performance and accuracy of the CNN.

[0109] In a feasible implementation, after step S40, steps D10 - D30 are further included:

[0110] Step D10, determining the utilization rate and corresponding weight of each convolution operator unit;

[0111] It should be noted that utilization refers to the actual degree of use of the convolution operator unit when processing tasks. It reflects whether the computing resources of the convolution operator unit are effectively utilized. High utilization means that the computing resources of the convolution operator unit are fully utilized, while low utilization may indicate a waste of computing resources. In this embodiment, weights are used to measure the importance or influence of each convolution operator unit. The weight can be determined based on the performance, contribution or other relevant factors of the convolution operator unit. The larger the weight, the greater the influence of the convolution operator unit in the configuration effect evaluation.

[0112] First, it is necessary to collect the actual usage of each convolution operator unit when processing tasks, including the amount of calculation, memory usage, etc., to calculate its utilization. Then, a weight is assigned to each convolution operator unit based on the performance, contribution or other relevant factors of the convolution operator unit. The determination of utilization and weight is based on an in-depth analysis and understanding of the actual usage of the convolution operator unit. By calculating the utilization, we can understand whether the computing resources of each convolution operator unit are effectively utilized; and by assigning weights, we can reflect the importance or influence of each convolution operator unit in the overall configuration.

[0113] Step D20, performing weighted summation according to the utilization rate and weight of each convolution operator unit to obtain the configuration effect;

[0114] It should be noted that the configuration effect is a comprehensive indicator used to evaluate the performance and efficiency of the current convolution operator unit configuration. It is obtained by weighted summation based on the utilization and weight of each convolution operator unit, reflecting the computational efficiency and resource utilization of the entire convolution layer or network.

[0115] The utilization rate of each convolution operator unit is multiplied by its corresponding weight to obtain the weighted utilization rate. Then, all weighted utilization rates are added together to obtain the configuration effect of the entire convolution layer.

[0116] Step D30: If the configuration effect is lower than the preset effect threshold, the convolution operator units in the configuration matrix are reallocated.

[0117] When the configuration effect is lower than the preset threshold, it indicates that the configuration of the current convolution operator unit is insufficient or unreasonable. At this time, the convolution operator units in the configuration matrix need to be reallocated to optimize the configuration effect. Reallocation can include increasing or decreasing the number of convolution operator units, adjusting the parameters or types of convolution operator units, etc. The purpose of reallocating convolution operator units is to optimize the configuration effect and improve the performance and efficiency of the entire convolution layer or network.

[0118] This embodiment implements dynamic monitoring and optimization of the convolution operator unit configuration, which helps to ensure that the convolutional neural network maintains a balance between computational efficiency and performance, and improves the overall performance of the model.

[0119] Exemplarily, the system performance E(t) is defined as an evaluation index of the configuration effect, as shown in formula (IV):

[0120]

[0121] Among them, for a system with n convolution operator units, u k (t) is the utilization rate of the kth convolution operator unit, k is a positive integer, 1 <k<n,ω k The feedback module notifies the configuration matrix to adjust the allocation of convolution operator units when E(t) is lower than the preset effect threshold to improve the utilization of system resources. Furthermore, the system also has an adaptive adjustment function. When it is detected that the load change of the convolution calculation exceeds the set range, the feedback module notifies the task decomposition and configuration matrix module to reallocate the convolution operator units to form an adaptive optimization.

[0122] In a feasible implementation, the method for parallel computing of convolution operations and aggregation of results may include steps E10 to E30:

[0123] Step E10, calculating the sum of the loads of all convolution operator units;

[0124] It should be noted that the sum of the loads refers to the total computational amount of the convolution tasks currently being processed by all convolution operator units, which is usually obtained by calculating the task amount of each convolution operator unit and summing them up.

[0125] First, we need to determine the amount of computational tasks currently undertaken by each convolution operator unit. This can be obtained by analyzing the input feature map of the convolution layer, the convolution kernel parameters, and the size of the output feature map. Then, we add up the loads of all convolution operator units to get the sum of the loads.

[0126] Step E20, when the sum of the loads is greater than a preset high load threshold, updating the configuration matrix by increasing the number of convolution operator units;

[0127] It should be noted that the preset high load threshold is a pre-set threshold used to determine whether the convolution calculation load of the current system is too high. If the sum of the loads exceeds this threshold, the number of convolution operator units needs to be increased to cope with it.

[0128] When the sum of the loads exceeds the high load threshold, it indicates that the computational requirements of the current convolutional layer are too high, which may lead to performance bottlenecks. To solve this problem, the configuration matrix needs to be updated to increase the number of convolutional operator units. This can be achieved by adding one or more rows to the configuration matrix, each row representing a new convolutional operator unit. Increasing the number of convolutional operator units can share the computational tasks and reduce the average load of each convolutional operator unit, thereby improving the computational efficiency of the entire convolutional layer. However, this may also lead to an increase in model parameters and an increase in computational cost. Therefore, it is necessary to find a balance between high load and low computational cost.

[0129] Step E30: When the sum of the loads is less than a preset low load threshold, the configuration matrix is ​​updated to reduce the number of convolution operator units.

[0130] It should be noted that the preset low load threshold is a pre-set threshold used to determine whether the convolution calculation load of the current system is too low. If the sum of the loads is lower than this threshold, the number of convolution operator units can be reduced to save resources.

[0131] When the sum of the loads is lower than the low load threshold, it indicates that the computational requirements of the current convolutional layer are too low and there is a waste of computational resources. In order to reduce the computational cost, the configuration matrix needs to be updated to reduce the number of convolution operator units. This can be achieved by deleting one or several rows in the configuration matrix, each row representing a removed convolution operator unit. Reducing the number of convolution operator units can reduce the complexity and computational cost of the model and improve computational efficiency. However, this may also lead to a decrease in model performance. Therefore, it is very important to find a balance between reducing computational costs and maintaining model performance.

[0132] For example, assuming that the high load threshold T high , when the convolution operation load L conv Meet L conv >T high When , the configuration matrix automatically increases the number of convolution operator units, as shown in formula (V):

[0133]

[0134] Among them, C d (t): represents the number of convolution operator units at time t. d (t-1): represents the number of convolution operator units at time t-1, that is, the number of PEs at the previous moment. δwhereδ represents the number of convolution operator units (PEs) increased or decreased at the current moment t relative to the previous moment t-1.

[0135] On the contrary, if L conv If the load is lower than the low load threshold, some convolution operator units are gradually shut down.

[0136] Furthermore, the system can also reduce the frequency and voltage of low-load convolution operator units according to the distribution of convolution tasks to minimize power consumption. i The calculation formula of (t) is shown in formula (VI):

[0137]

[0138] Where V i is the voltage, f i The power consumption control strategy adjusts the voltage and frequency of the inactive convolution operator units.

[0139] This embodiment implements dynamic monitoring of the convolutional layer load and power consumption and adaptive updating of the configuration matrix, which helps to ensure that the model maintains a balance between computational efficiency and performance and improves the overall performance of the model.

[0140] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the parallel computing and result aggregation methods of the convolution operations of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0141] This application also provides a convolution operation parallel computing and result aggregation system, please refer to Figure 4 , the convolution operation parallel computing and result aggregation system includes:

[0142] A data acquisition module 10 acquires at least one input image;

[0143] The image decomposition module 20 divides the input image into a plurality of block images according to the convolution layer parameters of the convolutional neural network model CNN;

[0144] A convolution operator module 30 processes a plurality of block images in parallel according to a plurality of convolution operator units in the CNN, wherein the block images processed by each convolution operator unit are different from each other;

[0145] The result aggregation module 40 performs multi-layer aggregation processing on the processing results of the block images processed by each convolution operator unit to obtain a convolution result, and outputs the convolution result as a prediction result.

[0146] The convolution operation parallel computing and result aggregation system provided by the present application adopts the convolution operation parallel computing and result aggregation method in the above embodiment, which can solve the technical problem of low utilization rate of multi-task parallel computing resources during image processing. Compared with the prior art, the beneficial effects of the convolution operation parallel computing and result aggregation system provided by the present application are the same as the beneficial effects of the convolution operation parallel computing and result aggregation method provided by the above embodiment, and other technical features of the convolution operation parallel computing and result aggregation system are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0147] The present application provides a device for parallel computing and result aggregation of convolution operations, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method for parallel computing and result aggregation of convolution operations in the above-mentioned embodiment one.

[0148] Reference below Figure 5 , which shows a schematic diagram of the structure of a device for parallel computing and result aggregation of convolution operations suitable for implementing the embodiments of the present application. The device for parallel computing and result aggregation of convolution operations in the embodiments of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The convolution operation parallel computing and result aggregation device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0149] like Figure 5As shown, the convolution operation parallel calculation and result aggregation device may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. Various programs and data required for the operation of the convolution operation parallel calculation and result aggregation device are also stored in the RAM 1004. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the convolution operation parallel computing and result aggregation device to communicate wirelessly or wired with other devices to exchange data. Although the convolution operation parallel computing and result aggregation device with various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have instead.

[0150] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0151] The convolution operation parallel calculation and result aggregation device provided by the present application adopts the convolution operation parallel calculation and result aggregation method in the above embodiment, which can solve the technical problem of low utilization rate of multi-task parallel calculation resources during image processing. Compared with the prior art, the beneficial effects of the convolution operation parallel calculation and result aggregation device provided by the present application are the same as the beneficial effects of the convolution operation parallel calculation and result aggregation method provided by the above embodiment, and other technical features in the convolution operation parallel calculation and result aggregation device are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0152] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0153] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0154] The present application provides a medium, which is a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, and the computer-readable program instructions are used to execute the convolution operation parallel calculation and result aggregation method in the above-mentioned embodiment.

[0155] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.

[0156] The above-mentioned computer-readable storage medium may be included in the device for parallel computing of convolution operations and aggregation of results; or it may exist independently without being assembled into the device for parallel computing of convolution operations and aggregation of results.

[0157] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the convolution operation parallel computing and result aggregation device, the convolution operation parallel computing and result aggregation device:

[0158] obtaining at least one input image;

[0159] According to the convolution layer parameters of the convolutional neural network model CNN, the input image is divided into multiple block images;

[0160] Processing multiple block images in parallel according to multiple convolution operator units in CNN, wherein the block images processed by each convolution operator unit are different from each other;

[0161] The processing results of the block images processed by each convolution operator unit are subjected to multi-layer aggregation processing to obtain a convolution result, and the convolution result is output as a prediction result.

[0162] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0163] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0164] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0165] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned method for parallel computing of convolution operations and aggregation of results, and can solve the technical problem of low utilization of multi-task parallel computing resources during image processing. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the method for parallel computing of convolution operations and aggregation of results provided by the above-mentioned embodiments, and will not be elaborated here.

[0166] The present application also provides a product, which is a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned convolution operation parallel calculation and result aggregation method.

[0167] The computer program product provided by the present application can solve the technical problem of low resource utilization of multi-task parallel computing during image processing. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the convolution operation parallel computing and result aggregation method provided in the above embodiment, which will not be described in detail here.

[0168] The above are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for parallel computing and result aggregation of convolution operations, characterized in that: Applied to a terminal, a convolutional neural network model is set in the terminal, and the convolution operation parallel calculation and result aggregation method includes: Obtaining at least one input image; According to the convolution layer parameters of the convolutional neural network model CNN, the input image is divided into a plurality of block images; Processing the multiple block images in parallel according to the multiple convolution operator units in the CNN, wherein the block images processed by each of the convolution operator units are different from each other; Multi-layer aggregation processing is performed on the processing results of the block image processed by each convolution operator unit to obtain a convolution result, and the convolution result is output as a prediction result.

2. The method for parallel computing and result aggregation of convolution operations according to claim 1, characterized in that: The step of dividing the input image into a plurality of block images according to the convolution layer parameters of the convolutional neural network model CNN comprises: Determine the convolution parameters of any convolution operator unit in the CNN, wherein the convolution parameters include the convolution kernel size and stride; The input image is divided according to the image size of the input image and the convolution parameter to obtain a task decomposition matrix, wherein each region in the task decomposition matrix includes a block image after the input image is divided.

3. The method for parallel computing and result aggregation of convolution operations according to claim 1, characterized in that: The step of performing multi-layer aggregation processing on the processing results of each of the convolution operator units processing the block image includes: Determine an aggregation path in the CNN, optimize the aggregation path according to the configuration matrix, and obtain a target aggregation path; The adder in the CNN is used to aggregate the processing results to obtain a first aggregation result, and multi-level aggregation processing is performed according to the target aggregation path and the first aggregation result to obtain a final aggregation result, and the final aggregation result is used as the convolution result, wherein, for the adder of each layer, the aggregation result of the current layer is obtained according to the target aggregation path and the aggregation result of the previous layer, and the aggregation result of the current layer is transmitted to the adder of the next layer through the target aggregation path.

4. The method for parallel computing and result aggregation of convolution operations according to claim 1, characterized in that: Before the step of processing the plurality of block images in parallel according to the plurality of convolution operator units in the CNN, the method further comprises: Determine a configuration matrix according to the task decomposition matrix and a plurality of convolution operator units in the CNN, wherein the configuration matrix includes a corresponding relationship between each block image in the task decomposition matrix and the convolution operator unit; The block images in the task decomposition matrix are allocated to the corresponding convolution operator units according to the configuration matrix.

5. The method for parallel computing and result aggregation of convolution operations according to claim 4, characterized in that: The step of allocating each of the block images in the task decomposition matrix to the corresponding convolution operator units according to the configuration matrix includes: If the input feature map is a multi-channel feature map, determining the task load of the multi-channel feature map according to the number of channels of the multi-channel feature map and the convolution parameters; A distribution matrix is ​​determined according to the task load, wherein the distribution matrix is ​​used to control the calculation allocation of each channel feature map.

6. The method for parallel computing and result aggregation of convolution operations according to claim 4, characterized in that: The convolution operation parallel computing and result aggregation method further includes: Calculating the sum of the loads of all the convolution operator units; When the sum of the loads is greater than a preset high load threshold, updating the configuration matrix by increasing the number of convolution operator units; When the sum of the loads is less than a preset low load threshold, the configuration matrix is ​​updated to reduce the number of convolution operator units.

7. The method for parallel computing and result aggregation of convolution operations according to claim 4, characterized in that: After the step of outputting the convolution result as a prediction result, the method further comprises: Determining the utilization rate and corresponding weight of each of the convolution operator units; Performing weighted summation according to the utilization rate and weight of each of the convolution operator units to obtain a configuration effect; If the configuration effect is lower than a preset effect threshold, the convolution operator units in the configuration matrix are reallocated.

8. A convolution operation parallel computing and result aggregation system, characterized in that: The convolution operation parallel computing and result aggregation system includes: A data acquisition module acquires at least one input image; An image decomposition module, which divides the input image into a plurality of block images according to the convolution layer parameters of the convolutional neural network model CNN; A convolution operator module, processing the plurality of block images in parallel according to the plurality of convolution operator units in the CNN, wherein the block images processed by each of the convolution operator units are different from each other; The result aggregation module performs multi-layer aggregation processing on the processing results of the block image processed by each convolution operator unit to obtain a convolution result, and outputs the convolution result as a prediction result.

9. A device for parallel computing and result aggregation of convolution operations, characterized in that: The device for parallel computing and result aggregation of convolution operations comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the method for parallel computing and result aggregation of convolution operations as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for parallel calculation and result aggregation of convolution operations as described in any one of claims 1 to 6.

Images