Data processing method and related products

Abstract

The present disclosure discloses a data processing method and related products, in which the data processing method includes: generating, by a general-purpose processor, a binary instruction according to device information of an AI processor, and generating an AI learning task according to the binaiy instruction; transmitting, by the general-purpose processor, the AI learning task to the cloud Al processor for running; receiving, by the general-purpose processor, a running result corresponding to the AI learning task; and determining, by the general-purpose processor, an offline running file according to the running result, where the offline running file is generated according to the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. By implementing the present disclosure, the debugging between the Al algorithm model and the AI processor can be achieved in advance.

Description

DATA PROCESSING METHOD AND RELATED PRODUCTS Mated AMliattoBi Thepresent applicationclaims priority to: Chinese Patent ApplicationNo. 201910315962.9 with 5 the titleof“DataProcessing Method andRelatedProducts” and filed onApril. 18,2019; The present applicationclaimspriority to: Chinese Patent ApplicationNo. 201910436801.5 with the title of“DataProcessing MethodandRelated Products” and filed onMay. 23, 2019. Technical Field 10 The disclosure relates generally to artificial intelligence processor technologies, and more specifically to a data processingmethod andrelatedproducts. Background In theprior art, only after an artificial intelligence(AI) processor has successfiilly taped out can an 15 algorithmapplication developer start developing andtestingAIalgorithmson theAIprocessorhardware. Asaresult,the functionalandperformanceresultoftheAIalgorithm developedfor theAlprocessor can be obtained only after the AI processor has taped out. It is an urgent problem how to cany out debuggingbetween anAI algorithm model andthe AIprocessor when the AIprocessor hasnot taped out yet. 20 Summary The embodiments of the present disclosure provide a data processing method and related products. It doesn’t matter whether an AIprocessor has taped out or not. The technical solutions disclosed herein facilitate debugging between anAI algorithm model and theAI processor before andafter the tape-out. 25 To that puipose. in some embodiments, the present disclosure proposes a data processing method appliedto a data processing system that includes a general-purpose processor andacloudAlprocessing platformor system. The general-purpose processor generates a binary instructionor binary instructions according to the device information of the AI processor, generates anAI learning task according to the binary instruction, and transmits the AI learning task to the cloud AI processing platform or system 30 (herein referred to as “cloud AI processing system”) for running. The cloud AI processor receives and then executes the AI learning task to generate a running result. The general-purpose processor receives therunningresult corresponding to theAIlearning task,anddetermines an offlinerunningfile according to the running result. The offlinerunning file is generated according to the device information of the Al processor and the binary instruction when therunningresult satisfies a preset requirement 1 CA 3065651 2019-12-195 10 15 20 25 30 CA 3065651 2019-12-19 The device information includes hardware architectlire information and parameters of running environment ofthe AIprocessor. Theparameters ofrunningenvironment include atleastone ofabase clock speedoftheAIprocessor, an access bandwidth of an off-chip memory and the Al processor, a size of an on-chip memory, the numberofdie cores of die Alprocessor, and a type ofan operatingunit of the AIprocessor. Insome embodiments, thepresentdisclosure providesadataprocessingdeviceincludingamemory, a general-purpose processor, and a cloud AI processor. The memory is configured to store a computer programrunning on the general-purpose processor and / or the cloudAIprocessor. Thegeneral-purpose processor isconfigured to generateabinaryinstruction accordingto the device information of the AI processor, generate an AI learning task according to the binary instruction, and transmittheAIlearning task to the cloudAIprocessor for running. ThecloudAIprocessor is configured to receive and execute the AI learning task, andthen generate arunningresult.The general-purposeprocessor is configuredto receive therunningresultcorresponding to the AIlearning task, and determine anoffline running file according to the runningresult. The offline running file is generated according to the device information of the Al processor and the binary instruction whenthe runningresult satisfies a preset requirement. In some embodiments, the present disclosure provides a data processing method applied to a general-purpose processor. The data processing method includes: generating,by the general-purpose processor, a binary instruction according to device information of the AIprocessor, andgenerating anAIlearning task according to the binary instruction; transmitting, by the general-purpose processor, the AI learning task to the cloud Al processor for running; receiving, by the general-purpose processor, arunningresult correspondingto the AIlearning task; and determining,by the general-purpose processor, anofflinerunningfile accordingto therunningresult; where tire offline running file is generated according to the device information of the AIprocessor and the binary instruction when therunningresult satisfies a preset requirement. Insome embodiments, thepresent disclosure provides adataprocessingdeviceincludingamemory and a general-purpose processor, where a computer programrunning on the general-purpose processor is storedinthe memory, and the dataprocessing methodrealized by the general-purpose processor when executingthe computer program includes: generating abinary instruction according to device information of the AIprocessor, and generating anAI learning task according to the binary instruction; transmitting the AI learningtask to the cloud AIprocessor for running; 25 10 15 20 25 30 CA 3065651 2019-12-19 receiving a runningresult corresponding to the AI learning task; and determining an offline running file according to the running result; wherein the offline running file is generated according to the device information of theAI processor and the binary instruction when the runningresult satisfies a preset requirement. In some embodiments, the present disclosure provides a data processing method applied to a cloud AIprocessor, and the data processingmethod includes: receiving an AI learning task, where the AI learning task is determined according to the binary instructiongenerated based ondie device information of the AI processor; and executing the Al learning task to generate a running result. In some embodiments, the present disclosure provides anAI processor including: a receiving module configured to receive the AI learning task, where the AI learning task is determined according to the binary instruction generated based on the device information of the AI processor; and an executingmodule configured to execute the AI learning task to generate the running result. In some embodiments, the present disclosure provides a data processing method applied to an Al processor;and the data processing method includes: obtaining anoffline running file, where tire offline running fileis generated accordingto the device information of the AI processor and the binary instruction when the running result satisfies the preset requirement. In some embodiments, the present disclosure provides anAI processor including: a file-obtaining module configured to obtain the offline running file, where the offline running file is generated according to the device information of theAI processor and the binary instruction when the running result satisfies the preset requirement. The technical solutions ofthe present disclosure have the following technical effects: (1) thepresent disclosure providesa software development platform on which the users can test and debug the function, performance and accuracy of algorithms developed for an AI processor against the AI processor; and (2) the offline running file generated after the debuggingis completedcan be deployedon multiple SoC (system on chip) chips ofcompatible architectures. In this way,users can debugthe function, performance and accuracy ofthe algorithm against the AI processorinadvance without getting thehardware entity, whichgreatly shortens theproductdevelopment cycle. Moreover, there is no need to develop or select a well-matched development environment separately for each SoC chip. 35 10 15 20 25 30 CA 3065651 2019-12-19 Brief Description of the Drawings In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure,the drawings usedinthe embodiments will be briefly introduced below. The drawings in the followingdescription only refer to some embodiments of the present disclosure, and are not intended to limit thedisclosure. Fig.1is a system architecture diagram ofthe technical solution. Fig.2is a structural diagramofasoftware stack of anAIprocessor. Fig.3 is a diagramof anembodiment ofthe technical solution. Fig. 4 is another diagram of an embodiment of the technical solution. Fig.5 is a flow chart of adata processingmethodproposed by the present disclosure. Fig.6is adiagramof various types ofbasic operators supportedby anAI learning library. Fig.7is a second flow chart ofadata processingmethodproposed by the present disclosure. Fig. 8 is a third flow chart ofa dataprocessingmethodproposedby the present disclosure. Fig.9is a fourth flow chart ofa data processingmethodproposed by the present disclosure. Fig. 10 is afunctionalblock diagram ofadataprocessingdeviceproposedbythe present disclosure. Fig. 11is a functional block diagram ofanAIprocessor proposed by the present disclosure. Fig.12 is a second functional block diagram ofanAIprocessorproposed by the present disclosure. Fig. 13 is a structural diagram ofa data processing system proposed by the present disclosure. Detailed Description of the Examples The technical solutions in the embodiments of the present disclosure will be described clearly and in details hereinafter with reference to the accompanied drawings. The embodiments of the present disclosure and the various features and advantageous details of the embodiments are described with reference to the non-limiting embodiments shown in the drawings and in the following description. It shouldbenotedthat the features shownin the drawingsarenotnecessarily drawn to scale.The disclosure does not describe the known materials, components, and process techniques so as not to obscure the embodiments of the present disclosure. The examples given are intended only to facilitate the understanding of the implementation of the embodiments of the present disclosure and further enable those skilled in the art to implement the embodiments. These examples should not be construed as limiting the scope of the present disclosure. Unless otherwise specifically defined, the technical or scientific terms used in this disclosure shall have a general meaning understood by a person of general skill in the field to which this disclosure belongs. The terms such as “first”, “second” and the like of the present disclosure are used for distinguishing between different components rather than describing a particular order, quantity or 45 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 importance.In addition, in each embodiment of the present disclosure, the same or similar reference label represents the same or similar component. To facilitate better understanding of the technical solutions disclosed herein, the technical terms involved in the embodiments of the present disclosure are first explained below. Tape-out: in a field of integrated circuit design, tape-out refers to trial production, that is, after the design of integrated circuits that meet preset functions is finished, several or dozens of integrated circuits are first produced for testing. This is called trial production. If during trial production the integrated circuits meets the test requirements, mass production is then carried out to mass produce the new ICs. A software stack of an AI processor: referring to Fig. 2, a software stack includes an AI application 200, an AI framework 202, an AI learning library 204, an AI runtime library 206, and a driver 208. Next, a structure of the software stack is described in detail. The AI application 200 corresponds to different application examples or use cases, and provides a corresponding AI algorithm model, where the algorithm model can be directly parsed by a programming interface of the AI framework 202. In a possible implementation, the AI algorithm model is converted into a binary instraction or a set of binary instructions by die AI learning library 204, and the AI runtime library 206 is called to convert the binary instruction into an AI learning task, where the AI learning task is placed in a task queue, and is called by the driver 208, and then executed by an underlying AI processor. In another possible implementation, the AI runtime library 206 may be also directly called to ran the offline running file that have been previously generated by freezing which could reduce the overhead of intermediate software modules and improve operating efficiency. Abinary instruction: information that can be recognized, i.e., read and understood, by the underlying AI processor. An AI processor: also referred to as a dedicated processor, which is a processor designed specifically for a particular application or field, for example, a Graphics Processing Unit (GPU). AGPU is also called a display core, a visual processor, a display chip. GPU is a dedicated processor designed for image processing on personal computers, workstations, game consoles and some mobile devices(such astablets and smartphones). For another example, a Neural Processing Unit (NPU) is a dedicated processor designed for matrix multiplication in the field of AI. NPU adopts an architecture of “data-driven parallel computing” and is particularly efficient at processing massive amounts of multimedia data of videos and images. A reconfigurable architecture: if an AI processor can utilize reusable hardware resources and flexibly change its architecture according to different application requirements to provide an architecture for each specific application requirement, then this artificial intelligent processor is called a reconfigurable computing system, and the architecture of this artificial intelligent processor is called a 55 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 reconfigurable architecture. A special programming language: a high-level programming language developed for a specific hardware platform, for example: Compute Unified Device Architecture™ (Cuda™) C. Specific embodiments of a data processing method and related products provided by the present disclosure will be described in detail below with reference to the accompanying drawings. In the prior art, an algorithm application developer can only carry out adaptation and debugging between the AI algorithm model and that AI processor for which the AI algorithm model is developed after the corresponding hardware of the AI processor becomes available. In the prior art, debugging between the chip that has not been taped out and the algorithm is not possible. Extra time after the tapeout is required to complete the adaptation between the algorithm model and the chip. As a result, product release is often delayed, and market opportunities are likely to be missed. In order to solve the above problems, the present disclosure provides a technical solution in which, regardless of whether the AI processor has taped out or not, a we11-matched AI processor is selected from a cloud processing system to simulate the AI processor according to the device information of the AI processor. The software development platform executes a series of processes on the algorithm model provided by the user to obtain a corresponding AI learning task. The AI learning task is sent to run on the cloud AI processor to obtain a running result. The AI learning task can be adjusted or optimized on the software development platform according to the running result. Optimization of the AI learning task may include adjusting the AI algorithm model, optimizing the AI learning library, and / or adjusting the device information of the AI processor. Through testing and optimization, integration between the AI processor and the AI algorithm model can be achieved. Fig. 1 is a system architecture diagram of the technical solution. As shown in Fig. 1, the system architecture includes a device 101 that is located remotely, for example, in a cloud processing platform or system. The system architecture further includes a software development platform 102 and a device 103. In one exemplary implementation, the software development platform 102 provides a series of toolkits for application development, performance tuning, function debugging, and the like. Examples of an application development tool include an AI learning library, an AI runtime library, a compiler, and a software development tool for specific application (such as video analysis). A function debugging tool can meet debugging requirements at different levels such as in a programming framework or in an AI learning library. Examples of a performance tuning tool include a performance profiling tooland a system monitoring tool. Examples of a compiler include a traditional C++ compiler as well as a machine learning compiler based on the C-like languages, or a machine learning compiler based on other high-level languages or domain specific languages that are designed specifically for a domain or field. Optionally, 65 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 the software development platform may run on a processor of the device 101 in a cloud processing platform, or run on a processor of a host computer device including a general-purpose processor (such as CPU) and displays, etc., which is not specifically limited herein. Furthermore, the software development platform may run, as an end-user, on the host computer device or the cloud device, which is not specifically limited herein. As shown in Fig. 3, this figure schematically shows one of the embodiments of the technical solutions disclosed herein. Users log into the software development platform 301 on a desktop computer, generate the AI learning task corresponding to the algorithm model on the software development platform, and adjust the AI learning task according to the running result of the AI learning task that is executed on the cloud AI processor 101. As shown in Fig. 4, this figure schematically shows another embodiment of the technical solutions disclosed herein. An AI software development client 400 is set up on the device 101 at the cloud processing platform. In one embodiment, the device 101 is a computer system that includes a general-purpose processor and at least one AI processor. For example, the at least one AI processor may include eight clusters, each of which includes fourAI processor cores. In practice, the software development platform 102 maintains user records stored by means of a data block or the like, and the user records include users’ personal information (account information, etc.) and service information requested by the users. The service information includes, but is not limited to, debugging requirements, device information of the AI processor. The debugging requirements include, but are not limited to, functional debugging and performance debugging. The device information includes, but is not limited to, hardware architecture information and parameters of the running environment of the AI processor. The parameters of the running environment include, but are not limited to, the base clock speed of the AI processor, the access bandwidth of the off-chip memory, the size of the on-chip memory, the number of cores of the AI processor, and the type of the operating units of the AI processor. In one implementation, the device 101 is provided with an AI processor. Examples of AI processors may include, but is not limited to, AI processor chip, field programmable gate array, and simulators. AI processor chip may be a reconfigurable chip or a non-reconfigurable chip. In some embodiments, the device 101 may be a server board or a cluster of server boards. In one implementation, the device 103 may be a terminal device, such as a tablet or a mobile phone. The device 103 may also be an edge device, such as a camera. The device 103 may include an actual AI processor. In some embodiments, the device 103 may be an AI processor that is still in the design stage and has not been taped out. A working principle of the technical solutions is described as follows. On the software development platform 102, a driver selects, from the device 101, an AI processor that matches the device information of the AI processor in the device 103. The device 101 is a cloud processing system or 75 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 platform. One of the selection criteria is that the hardware architecture of the selected cloud AI processor should be compatible with the hardware architecture information of the AI processors in the device 103. Another one of the selection criteria is that the instruction set of the cloud AI processor should be compatible with the instruction set of the AI processor. For example, the hardware architecture information of the cloud AI processor is compatible with the hardware architecture information of the AI processor in the device 103 if the computing power of the cloud AI processor is greater than or equal to the computing power of the AI processor. On the software development platform 102, software parameters corresponding to the compiler interface of die AI learning library are configured according to the device information of theAI processor. Combined with the algorithm model obtained by the programming framework, the compiler interface of the AI learning library that has been configured is invoked to compile and obtain the binary instruction of the corresponding AI processor. The binary instruction is processed by the runtime library to generate the AI learning task, where the AI learning task is placed in a task queue, and will be called by the driver and executed by the cloud AI processor. After the AI learning task is executed by the cloud AI processor, the running result is fed back to the software development platform 102. Optionally, the software development platform 102 can display the running result. The user may issue an operation instruction based on the running result. The software development platform 102 receives the operation instruction from the user and may adjust the binary instruction according to the user’s operation instruction in three different ways. The three ways are as follows: adjusting the hardware architecture information of the AI processor, adjusting the parameters of running environment of the AI processor, and optimizing the AI learning task. The adjusted binary instruction is converted into a corresponding AI learning task, and is put into the task queue. Next, the AI learning task in the task queue is again called by the driver, and executed by the cloud AI processor. This process can be repeated until the running result fed back from the cloud AI processor satisfies an expected condition or requirement. In the technical solutions disclosed herein, the device information of the AI processor corresponds to the software parameters of the compiler interface of the AI learning library. The software parameters may include more information, such as RAM size, Cache size, whether to pass the Cache, and the like. Since such information is related to an operation field that is allocated when the binary instraction is generated, if the AI algorithm model remains the same, the binary instruction can be adjusted by updating the device information of the AI processor, so that the AI learning task can be adjusted. Regardless of whether the AI processor has taped out or not, the well-matched cloud AI processor selected from the device 101 according to the device information of the AI processor can simulate the AI processor, and the corresponding AI learning task can be executed on the device 101. Using the running result, the users 85 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 can complete debugging and testing of the function, performance and accuracy of the algorithm model against the AI processor on the software development platform. An offline running file generated after the debugging is completed can be deployed on multiple SoC chips of compatible architectures. In this way, the users can perform testing between the algorithm model and the AI processor in advance without relying on the processor hardware, which will greatly shorten the product development cycle. Moreover, there is no need to develop or select a separate development environment for each SoC chip. Furthermore, in the technical solutions disclosed herein, the current running environment parameters corresponding to the device information of the cloud AI processor may be the same as the actual running environment parameters, or different from the actual running parameters.According to an execution result of the cloud AI processor executing the specific AI learning task, it may be determined whether the device information of the AI processor satisfies an expected condition. If the device information of the AI processor does not meet the expected condition, the device information of the AI processor may be further adjusted until the device information of the AI processor satisfies the expected condition. Therefore, even when the architecture of an AI processor is still in the design stage, the design specifications of the SoC chip can be assessed using the technical solutions disclosed herein. Based on the above description, Fig. 5 is a flow chart of a data processing method proposed by the present disclosure. The method is applied to the general-purpose processor and corresponds to the software development platform shown in Fig. 1. As described above, the general-purpose processor may be a general-purpose processor in the device 101 of a cloud processing platform or a general-purpose processor of a host computer device. The data processing method includes: step 501, generating, by the general-purpose processor, the binary instruction according to the device information of the AI processor; and generating, by the general-purpose processor, the AI learning task according to the binary instruction. In the technical solutions disclosed herein, a plurality of programming frameworks can be integrated in the software development platform 102, for example, the tensor flow AI learning system TensorFlow™, the deep learning framework Caffe™, Caffe2™, MXNet™, and the like. Taking Caffe™ as an example, Caffe™ has three core modules, namely Blobs, Layers and Nets. Among them, Blobs is configured for data storage, data interaction and data processing. And through Blobs, an interface of data memory can be formulated uniformly. Layers is the core of the neural network and defines many hierarchical structures. Layers treats Blobs as input / output. Nets is a collection of Layers where these layer structures form a network through connection. For step 501, software parameters corresponding to the compiler interface of the AI learning library are configured according to the device information of the AI processor. Combined with the algorithm model obtained by the programming framework, the compiler interface of the AI learning library that has 95 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 been configured is called to compile and obtain the binary instruction of the corresponding AI processor. The binary instruction is processed by the runtime library to generate the AI learning task, where the AI learning task is placed in a task queue, invoked by the driver, and executed by the cloud AI processor. In practical applications, the AI learning library is configured to accelerate various AI learning algorithms on the AI processors, where the AI learning algorithms include, but are not limited to, a deep learning algorithm, a convolutional neural network algorithm, a recurrent neural network algorithm, and the like. Specifically, the AI learning library generally has the following features: 1. Supporting various types of basic operators in machine learning algorithms. In one implementation, a variety of machine learning algorithms can be implemented through the combination of basic operators, thereby satisfying the requirements of versatility, flexibility, and scalability. For instance, the various types of basic operators involved herein may include: a common neural network operator, a matrix operator, a vector operator, a scalar operator, a recurrent neural network operator, etc. Fig. 6 is a diagram of various types of basic operators supported by an AI learning library. As shown in Fig. 6, the various types of basic operators supported by the AI learning library include the common neural network operator 1. The common neural network operator 1 often includes a convolution / deconvolution operator 11, a pooling operator 12, an activation operator 13, a Local Response Normalization (LRN) / batch normalization operator 14, a Softmax operator 15, and a full connection operator 16. In some embodiments, the activation operator 13 includes, but is not limited to, Rectified Linear Unit (ReLU), Sigmoid, Tanh, and other operators that can be implemented by interpolation. The operator 2 of matrix, vector, and scalar may include a matrix multiplication operator 21, a tensor addition / subtraction operator 22, a tensor logical operation operator 23, a tensor transformation operator 24, a Region of Interest (ROI) Pooling operator 25, and a Proposal operator 26. The tensor transformation operator 24 includes, but is not limited to, Crop, tensor Reshape, tensor Slice, tensor Concat, etc.; the recurrent neural network operator 3 includes a Long Short-Term Memory (LSTM) operator 31, a basic Recurrent Neural Network (RNN) operator, a RNN Operator 32, and a Singular Value Decomposition Filter (SVDF) Operator 33. In practical applications, users can freely add new operators or change different versions of AI learning libraries according to their own needs, which is not detailed here. How to optimize the AI learning task on the software development platform based on the AI learning library will be described in detail when the process of debugging the AI learning task is described. 2. Supporting fusion of basic operators. In one implementation, when fused operators are being compiled, some compiler optimization 105 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 methods such as memory multiplexing, memory access optimization, instruction pipeline, and data type optimization (for example, selecting different applicable data types) may be adopted, significantly improving the overall performance of the fused operators. 3. Supporting generation of offline running file. The offline running file generated after the running result is returned from the cloud AI computer system may include necessary network structure information such as a network weight of each computing node in the AI algorithm model and an instruction configured to indicate which computing function is realized by the computing node. Specifically, the instruction may include information such as computation properties of each computing node in the AI learning model and connection relationship between each computing node. In one implementation, the offline running file can be separated from the AI learning library, and can be run independently based on die AI runtime library. In practical applications, since the offline running file is separated from the upper software stack, the execution of the offline running file has better performance and versatility. The data processing method further includes: step 502, transmitting, by the general-purpose processor, the AI learning task to the cloud AI processor, which selected from the device 101 according to the device information of the AI processor. The selected cloud AI processor can simulate the AI processor. In this way, the AI learning task generated on the software development platform 102 can be run on the cloud AI processor. The data processing method further includes: step 503, receiving, by the general-purpose processor, the running result corresponding to the AI learning task. The cloud AI processor executes the AI learning task to generate a running result that is fed back to the software development platform 102 for display. The running result may include, but is not limited to, one or more pieces of information of whether running time of the AI learning task on the cloud AI processor satisfies a first expectation, whether the load information of the cloud AI processing system when the artificial intelligent learning task is executed satisfies a second expectation, and whether the result of theAI learning task satisfies a third expectation. In some embodiments, the cloudAI processing system includes a general-purpose processor and a cloud AI processor. When the AI learning task is being executed, it may be necessary to know not only the load information of the cloud AI processor but also information of used memory and the occupancy rate of the general-purpose processor during execution. The reason why the load information is included in the running result is that if an AI learning task requires too much resources on the general-purpose processor, it is likely to work poorly or fail to run on the device 103. 115 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 The data processing method further includes: step 504, determining, by the general-purpose processor, the offline running file according to the running result, where the offline running file is generated according to the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. The offline running file may include one or more of the following: version information of the offline running file, version information of the AI processor, a binary instruction, a constant table, the size of input data / output data, data layout description information, and parameter information. Specifically, the version information of the offline running represents different versions of the offline running file; and the version information of the AI processor refers to the hardware architecture information of the AI processor. For example, the hardware architecture information may be represented by a version number of the chip architecture, or may be represented by the functional description. The data layout description information refers to preprocessing of the layout and type of the input data / output data based on hardware characteristics. The constant table, the size of input data / output data, and the parameter information may be determined by the developed AI algorithm model. The parameter information may be weight data in an AI algorithm model. Data that is needed in the execution of the binary instruction may be stored in the constant table. The device information of the AI processor includes the hardware architecture information and parameters of running environment of the AI processor. Herein the parameters of running environment include: the base clock speed of the AI processor, the access bandwidth of an off-chip memory and the AI processor, the size of an on-chip memory, the number of the cores of the AI processor, and / or the type of the operating unit or units of the AI processor. If the running result satisfies the preset requirement, the corresponding offline running file is generated according to the binary instruction that satisfies the preset requirement. If the running result does not satisfy the preset requirement, at least one optimization in one of the below-listed optimization processes may be executed by the function debugging tool and / or the performance tuning tool until the running result satisfies the preset requirement. Afterwards, the corresponding offline running file is generated according to the binary instruction that satisfies the preset requirement. The optimization processes include: adjusting the hardware architecture information of the AI processor, adjusting the parameters of running environment of the AI processor, and / or optimizing the AI learning task. When the running result fed back by the cloud AI processor satisfies the preset requirement, the binary instruction corresponding to the current AI learning task is written or “frozen” into an offline running file through an offline mode. If the running result fed back by the cloud AI processor does not satisfy the preset requirement, the debugging process of the AI learning task falls into two scenarios. The first scenario is when the AI 125 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 processor is in the chip design stage. The technical solutions disclosed herein are used to assess the chip design specifications based on the running result. In this case, the hardware architecture information and parameters of running environment of the chip can be changed. In such case, one or more of three optimizations: adjusting the hardware architecture information of the AI processor, adjusting the parameters of running environment of the AI processor, and optimizing the AI learning task, may be executed on the software development platform. All three optimization manners can adjust the binary instruction corresponding to the AI learning task accordingly. After each adjustment, flie adjusted AI learning task in the task queue is called by the driver and is executed by the corresponding cloud AI processor to obtain a new running result. If the new running result is still not as expected, the users can repeat the above steps until the running result is satisfactory. The debugged binary instruction is written or “frozen” into an offline running file through an offline mode. The second scenario is: regardless of whether the AI processor has taped out or not, the software design and development of the AI algorithm model for the AI processor may be enabled by using the software development platform to debug between the AI processor and the AI algorithm model. In this scenario, the hardware architecture information of the chip will not be modified freely unless,for example, the right to use the chip with other architecture versions has been purchased. Assuming that the hardware architecture information of the chip does not change, at least two optimization processes are available: adjusting the parameters of running environment within a range of parameters of the running environment supported by the current hardware architecture information, and optimizing the AI learning task that is executed on the software development platform. Both optimization operations can adjust the binary instruction corresponding to the AI learning task accordingly. After each adjustment, the AI learning task in the task queue iscalled by the driver and is executed by the corresponding cloud AI processor to obtain a new running result. If the new running result is still not as expected, the users can repeat the above steps until the running result is satisfactory. The debugged binary instruction is then written or “frozen” into an offline running file for future use in an offline mode. The offline running file should generate consistent result whether it runs on the cloud AI processor or on the AI processor. The results should be the same or within a certain allowable error range. To achieve that, the cloud AI processor is selected from a plurality of AI processors to simulate the correspondingAI processor according to the device information of the AI processor. The selection criteria include compatible hardware architecture and compatible instruction set between the cloud AI processor and the AI processor, in order to enable seamless migration of the offline running file from one processor to another. In some embodiments, the device information of different types of AI processor may be pre-stored in the software development platform 102. According to actual needs, the target information is selected 135 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 from the pre-stored device information, and the cloud AI processor to be used to simulate theAI processor is selected from the device 101 according to the target information. In some embodiments, each time the device information is adjusted, the user sets different device information on the software development platform 102 according to the adjusted device information. The software development platform 102 receives the newly set device information of the AI processor, and selects the cloud AI processor from the device 101 to simulate the AI processor according to the newly received device information. It should be noted that the manners of obtaining the device information of the AI processor are only examples. Those skilled in the art may come up with other variations or modifications based on the embodiments disclosed herein when they understand the essence of die applied technical solutions. For example, the device 103 may transmit the request information to the software development platform 102, and the software development platform 102 may parse the request information to obtain the device information of the AI processor. As long as the functions realized and the technicaleffects achieved are similar to what is disclosed the present application, they shall fall within the scope of protection of the disclosure. In practical applications, when selecting a cloud AI processor from the device 101 to simulate the AI processor in the device 103, the device information of the AI processor in the device 103 is written to the driver to enable the driver to select a suitable cloud AI processor from the device 101. The process of selecting a cloud AI processor includes: selecting a well-matched cloud AI processor according to the hardware architecture information of the AI processor, where the hardware architecture of the selected cloud AI processor is compatible with hardware architecture of the corresponding AI processor, and the instruction set of the cloud AI processors is compatible with the instruction set of the corresponding AI processor; and adjusting the clock speed and the memory bandwidth of the cloud AI processor according to the parameters of the running environment of the AI processor. In addition, there are four ways or approaches to optimize the AI learning task. First, users can compile a dynamic link library based on the programming language on the software development platform, and invoke the dynamic link library within the programming framework. Second, users can develop new operators based on the programming language on the software development platform to create a new offline running file, in combination with the host AI learning library that is already available. Using a Proposal operator as an example, by replacing the Proposal operator in Faster-R-Cnn with a Plugin operator(Op) and calling a proposal kemel.mlu operator written in a special programming language, users can replace the Proposal operator in a Cambricon-Caffe™ framework with Proposal Kernel written in a special programming language through PluginOp, which links the special programming language with the existing AI learning library, supports various features 145 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 of the AI learning library and running modes such as online, offline, layer-by-layer, and fusion. As indicated by the first and second approaches,a large number of layers and operators are already supported in the framework, and general models can be run on the server boards in a cloud processing system. However, operators are updated frequently and they evolve quickly. Over time, individuals or organizations may have accumulated some custom operators and algorithms. In such case, because it is better not to expose the custom algorithms and because the underlying library cannot efficiently support the actual applications to meet the requirements, a special programming language may be provided to help developers to develop algorithms independently, which will address the problem that previous development models are not sufficiently flexible. Third, the users can select one of the versions of the current host AI learning library that is already available on the software development platform and match the corresponding AI runtime library. If the current host AI learning library that is already available cannot satisfy the requirements, requests can be sent through the software development platform to upgrade the version of the host AI learning library. Upon request, the operator can provide a new version of the AI learning library and the corresponding AI runtime library to the software development platform. In this way, the users can rely on the latest version of the AI learning library and the corresponding AI runtime library on the software development platform, and obtain the debugged binary instruction based on the latest version of the AI learning library. Fourth, the users can adjust the AI algorithm model to achieve the goal of optimizing the AI learning task. In practical applications, the purpose of optimizing the AI learning task may be achieved by at least one of the above four approaches. Regardless of whether the AI algorithm model is adjusted or not, the purpose of adjusting the AI learning task can be achieved by optimizing the AI learning library and / or adjusting the device information of the AI processor, and then the adaptation between the AI processor and the AI algorithm models can also be achieved. The technical solutions as shown in Fig. 5 provide a software development platform, on which the users can conduct debugging of the function, performance and accuracy between the algoritom and the AI processor, and the offline running file generated after the debugging is completed can be deployed on multiple SoC chips of compatible architectures. In this way, toe users can debug toe function, performance and precision between the algorithm and the AI processor in advance without relying on the hardware, which greatly shortens toe product development cycle. Moreover, there is no need to develop or select a matched development environment separately for each SoC chip. Fig. 7 is another flow chart of a data processing method proposed by the present disclosure. The data processing method is applied to the cloud AI processor. The data processing method includes: step 701: receiving toe AI learning task, where toe AI learning task is determined according to the 155 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 binary instruction generated based on the device information of the AI processor; step 702: executing the AI learning task to generate the running result. In this step, the running result includes, but is not limited to, at least one or more pieces of information of whether running time of the AI learning task on the cloud AI processor satisfies the first expectation, whether load information of the cloud AI processing system when the artificial intelligent learning task is executed satisfies the second expectation, and whether the result of the AI learning task satisfies the third expectation. It should be emphasized that all the related technical solutions shown in Fig. 5 are applicable to the technical solutions shown in Fig. 7, and details are not described herein again. It should be clarified that, for the device 101, one application scenario is: in the set of the selected AI processors, the hardware architecture information of all versions of the AI processor should be included as much as possible. For example, the models corresponding to the versions of the hardware architecture information and the models of the AI processor are A, B, and C, etc. The set of the cloud AI processors selected includes an AI processor of type A, an AI processor of type B, and an AI processor of type C. Another application scenario is that in the set of selected cloud AI processors, the hardware structure of the cloud AI processors can achieve high-configuration, mid-configuration, and lowconfiguration functions. For example, according to different application scenarios and actual needs, some functions of the selected cloud AI processor may be shielded, so that the AI processors with highconfiguration functions could be converted to the AI processors with low-configuration functions or the AI processors with med-configuration functions, to meet the needs of different users. In this case, the software development platform 102 of the technical solution generates the binary instruction based on the device information of the AI processor, and the driver changes different parameters of running environment supported by the selected cloud AI processor, so that some functions of the AI processor with high-configuration functions are shielded, and only the functions implemented are adapted to the functions of the corresponding AI processor. For instance, the value ranges of the parameters of running environment of the AI processor with high-configuration functions include all parameters of the running environment supported by the AI processor. For example, the size of the on-chip memory of the cloud AI processor is 100M, and the size of the on-chip memory of the AI processor is less than 100M. Additionally, a time division multiplexing method combined with a virtual machine technology may be adopted to allocate the cloud AI processor of the device 101 efficiently, for example, based on the time period when the users use the resources of the device 101. In this way, the resources can be allocated to different AI learning tasks executed in different time periods, which could reduce the number of development environments in the cloud processing system that are needed to be deployed. Furthermore, in the set of the AI processors of the device 101, not all the AI processors must be 165 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 hardware processors. They may also be Field-Programmable Gate Array (FPGA) or simulators. According to the mainstream of modem IC design verification technology, a circuit design completed in hardware description languages (Verilog™ or Vhsic Hardware Description Language (VHDL)) can be quickly burned onto an FPGA through simple synthesis and layout. For the purposes of the present disclosure, if there is no well-matched hardware processor in the cloud AI processor, the FPGA can be used to provide simulation services for the users.According to the device information of theAI processor, an FPGA that meets the requirements is selected. The selected FPGA has a mirroring file corresponding to the hardware architecture information of the AI processor. If there is no FPGA that meets the requirements, the software development platform 102 can bum the mirroring file corresponding to the hardware architecture information of the AI processor onto an idle FPGA. The newly formed FPGA can then execute the AI learning task transmitted by the software development platform. For the cloud AI processor, a more fine-tuned resource allocation can be provided. For example, the user A generates an AI learning task on the software development platform 102 based on the AI processor. This task requires or is composed of M cores, and the well-matched cloud AI processor selected at the device 101 has Ncores. In the selected cloud AI processor, P cores out of the N cores have been used by the AI learning task initiated by the user B. If M+P<=N and the device information of A’s AI processor and the device information of B’s AI processor are the same, then the AI runtime library in the software development platform 102 can allocate the AI learning tasks initiated by different users to different cores of the same cloud AI processor. Different AI learning task can be executed on different cores to achieve finer allocation of resources of the cloud AI processor. In addition, the cloud AI processor may be a common AI processor with a non-reconfigurable architecture or an AI processor with a reconfigurable architecture. The AI processor with a reconfigurable architecture uses the device information in the driver to adjust the parameters of running environment inside a reconfigurable chip, and calls functional modules inside the reconfigurable chipaccording to the AI learning task transmitted by the software development platform 102. That is, according to an actual application, the functional modules inside the reconfigurable chip may be adjusted so that the adjusted chip can simulate the AI processor. Based on the above description of the cloud AI processor, in this disclosure, the software development platform 102 counts the number of users of the cloud AI processors with different hardware architecture information in each time period over a period of time, and estimates a minimum value V that can meet the users’ needs, where the minimum value V is the minimum number of deployed cloud AI processors. On this basis, a small number of W redundant AI processors are added for fault tolerance or to prevent the sudden increase in the number of users. Then (V+W) is the number of AI processors that need to be deployed for the device 101. At the same time, the software development platform 102 175 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 periodically tracks the change in the number of users, and may request a change of the number of AI processors deployed in the device 101, to better meet customer needs and reduce overhead in the device 101. It can be seen from the above description that the AI processors deployed in the device 101 could be dynamically adjusted according to the AI learning task transmitted by the software development platform in a real-time deployment manner. In this way, by configuring different development environments, the time division multiplexing method adapts to allocate the cloud AI processor resource as required by the AI learning tasks executed at each time period, which can reduce the number of development environments of the cloud AI processors that need to be deployed. The above cloud AI resource allocation process may be transparent to the users. Fig. 8 is another flow chart of a data processing method proposed by the present disclosure. The data processing method is applicable to the AI processor, which includes: step 801, reading the offline running file, where the offline running file is determined according to the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. It should be emphasized that all the related technical solutions shown in Fig. 5 and Fig. 7 are applicable to the technical solutions shown in Fig. 8, and the details are not described herein again. On the software development platform 102, according to the running result fed back from the cloud AI processor, the AI learning task generated for the AI processor is optimized and debugged. After the running result meets the expected requirement, the binary instruction of the debugged AI learning task is converted into an offline running file, which may incorporate previous debugging and performance information. In future practical applications, the offline running file can be compiled directly, independent from the programming framework. In this way, the offline running file can be crossly compiled on the device 103 for field deployment and still maintain adequate accuracy. Fig. 9 is another flow chart of a data processing method proposed by the present disclosure. The system includes a general-purpose processor and a cloud AI processor. The data processing method includes: stepa at 901, generating, by the general-purpose processor, the binary instruction according to the device information of the AI processor; generating the AI learning task according to the binary instruction; and transmitting the AI learning task to the cloud AI processor to run; step b at 902, receiving, by the cloud AI processor, and executing the AI learning task to generate the running result; step c at 903, receiving, by the general-purpose processor, the running result of the AI learning task, and determining the offline running file according to the running result, where the offline running 185 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 file is determined according to the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. It should be emphasized that all the related technical solutions shown in Fig. 5 and Fig. 7 are applicable to the technical solutions shown in Fig. 9, and the details are not described herein again. Fig. 10 is a functional block diagram of a data processing device proposed by the present disclosure. The data processing device includes a memory 1001 and a general-purpose processor 1002, where a computer program 1003 that can be run on the general-purpose processor is stored in Ihe memory. The data processing flow that is being implemented when the general-purpose processor executes the computer program includes: generating the binary instruction according to the device information of the AI processor, and generating the AI learning task according to the binary instruction; transmitting the AI learning task to the cloud AI processor to run; receiving the running result of the AI learning task; and determining the offline running file according to the running result, where the offline running file is generated according to the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. Optionally, the device information includes the hardware architecture information and the parameters of running environment of the AI processor. Optionally, the parameters of running environment include at least one of the base clock speed of the AI processor, an access bandwidth of an off-chip memory and the AI processor, the size of an on-chip memory, the number of cores in the AI processor, and the type of an operating unit or units of the AI processor. Optionally, the data processing flow further includes: receiving the device information of the AI processor. Optionally, the data processing flow further includes: writing the device information into the driver to select a well-matched cloud AI processor according to the device information in the driver. Optionally, the data processing flow when the general-purpose processor executes the computer program to determine the offline running file according to the running result includes: if the running result satisfies the preset requirement, generating the corresponding offline running file according to the binary instruction that satisfies the preset requirement. Optionally, the data processing flow when the general-purpose processor executes the computer program to determine the offline running file according to the running result further includes: if the running result does not satisfy the preset requirement, executing at least one of the 195 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 optimization manners in the following process until the running result satisfies the preset requirement, and generating the corresponding offline running file according to the binary instruction that satisfies the preset requirement, where the optimization manners include: adjusting the hardware architecture information of the AI processor, adjusting the parameters of running environment of the AI processor, and / or optimizing the AI learning task. Optionally, the data processing flow when the general-purpose processor optimizes the AI learning task includes: optimizing the AI learning task by a special programming language. Optionally, the data processing flow when the general-purpose processor optimizes1he AI learning task further includes: optimizing the AI learning task by updating the version of the AI learning task. Optionally, the data processing flow when the general-purpose processor optimizes1he AI learning task further includes: adjusting the AI algorithm model. Optionally, the running result may include one or more pieces of information such as whether the processing time of the AI learning task on the cloud AI processor satisfies the expectation, whether the load information of the cloud AI processing system when the artificial intelligent learning task isexecuted satisfies the expectation, and whether the result of the AI learning task satisfies the expectation. The functions implemented by the general-purpose processor can be explained in comparison with the foregoing embodiments and the same technical effects of those embodiments can be achieved as well. The details are not described herein again. In the embodiments of the present disclosure, the memory may include physical devices for storing information, where the information is typically digitalized and then stored in a device that may be electric, magnetic or optical. The memory in the embodiments may further include: a device for storing information in an electric device such as a RAM, a ROM, etc.; in a magnetic device, such as a hard disk, a floppy disk, a magnetic tape, a magnetic core memory, a magnetic bubble memory, and a USB flash drive; or in an optical device, such as a CDor a DVD. Of course, there are other types of memory devices, such as quantum memory device, graphene memory device, and the like. Fig. 11 is a functional block diagram of an AI processor proposed by the present disclosure. The AI processor includes: a receiving module 1101 configured to receive the AI learning task, where the AI learning task is determined according to the binary instruction generated based on the device information of the AI processor; and an executing module 1102 configured to execute the AI learning task to generate the running result. 205 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 Optionally, the running result generated by the executing module may include one or more pieces of information suchas whether the running time of the AI learning task on the cloud AI processor satisfies the expectation, whether the load information of an AI cloud processing system when the artificial intelligent learning task is executed satisfies the expectation, and whether the result of the AI learning task satisfies the expectation. Optionally, the cloud AI processor includes at least one of a hardware piece of the AI processor, a field programmable gate array, and a simulator. Optionally, the AI processor is an AI processor with a reconfigurable architecture. Fig. 12 is another functional block diagram of an AI processor proposed by the present disclosure. The AI processor includes: an obtaining module 1201 configured to obtain the offline running file, where the offline running file is generated according to the device information of the AI processor and the binary instruction when the running result satisfies the preset requirement. The AI processor further includes: a transmitting module configured to transmit request information, where the request information includes the device information of the AI processor. Optionally, the device information includes the hardware architecture information and the parameters of running environment of the AI processor. Optionally, the parameters of running environment include at least one of the base clock speed of the AI processor, the access bandwidth of an off-chip memory and the AI processor, the size of an onchip memory, the number of the cores of the AI processor, and the type of an operating unit or units of the AI processor. Fig. 13 is a structural diagram of a data processing system proposed by the present disclosure. The data processing system includes a general-purpose processor 1310 and a cloud AI processor 1320. It should be understood that, in this embodiment, for the specific implementation of the general-purpose processor 1310 and the cloud AI processor 1320, please refer to the foregoing description, which will not repeated herein again. The embodiment of the present disclosure further provides a readable storage medium, in which the computer program is stored. The process of executing the computer program embodies the data processing method shown in Fig. 5, Fig. 7, Fig. 8, and Fig. 9. It can be seen from the above descriptions that the embodiments of the present disclosure provide a data processing method and related products. Regardless of whether the AI processor is tape out or not, this technical solution can realize debugging between the AI algorithm model and the AI processor in advance. 215 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 Those skilled in the art also know that in addition to implementing the technical solution on the client and the server in the form of computer readable program code, the technical solution can also be implemented on the client and the server in the form of logic gate, switch, ASIC, programmable logic controller and embedded microcontroller. Therefore, the client and server may be considered as a hardware component, and the devices for implementing various functions included within the hardware component can also be considered as structures inside the hardware component. Or even the devices for implementing various functions can be considered as software modules which can implement methods or structures within the hardware component. From the above descriptions of the embodiments of the present disclosure, those skilled in the art may clearly know that this application can be achieved by means of software together with a necessary general-purpose hardware platform. Based on such understanding, the technical solutions of the present disclosure essentially, or the part of the technical solutions that contributes to the related art may be embodied in the form of a software product which is stored in a storage medium (such as ROM / RAM, Disk, and Compact Disc (CD)), where the storage medium includes instructions for causing a computer device (which may be a personal computer, a server, or a network device and so on) to perform the methods described in each embodiment of the present disclosure or in certain parts of the embodiments. Each embodiment in the present disclosure is described step by step, and the same and similar parts between all the embodiments may be referred to each other, and each embodiment focuses on differences from other embodiments. In particular, for the implementation of the client and server, reference can be made to the embodiment of the aforementioned method. The application can be embodied in the context of instructions executed by the computer, such as a program module. In general, the program module includes routines, programs, objects, components, data structures, and the like that perform particular tasks or implement particular abstract data types. The present application can also be implemented in a distributed computing environment where tasks are performed by remote processing devices that are connected through a communication network. In the distributed computing environment, the program module can be located in both host and remote computer storage medium including storage devices. Furthermore, the above descriptions can be better understood in accordance with the following terms: Al. A data processing method applied to a data processing system, which includes a generalpurpose processor and a cloud AI processor, wherein the data processing method includes: generating, by the general-purpose processor, a binary instruction according to device information of theAI processor at an end; generating, by the general-purpose processor, an AI learning task according 225 10 15 20 25 30 CA 3065651 2019-12-19 to the binary instruction; andtransmitting, by the general-purpose processor, the Allearning task to the cloud Alprocessor for running; receiving,by the cloudAlprocessor, theAIlearningtask;andexecuting,by the cloudAlprocessor, theAI learning task to generate arunning result; and receiving, by the general-purpose processor, Ihe running result corresponding to the AI learning task; determining, by the general-purpose processor, an offline running file according to the running result, wherein the offline running file is generated according to the device information of the AI processor and the binaiy instruction when the running result satisfies a preset requirement A2. The data processing method of Al, wherein the device information includes hardware architecture information and parameters ofrunning environment ofthe AIprocessor. A3. The data processing methodofA2, wherein the parameters of running environment include at least oneofa base clock speedof theAI processor, an access bandwidth of an off-chip memory and the AIprocessor, a size ofanon-chip memory, the number of the coresofthe AIprocessor, anda type of an operatingunit ofthe AIprocessor. A4.The data processingmethodofAl further includes: receiving,by the general-purpose processor, die device informationofthe AIprocessor. A5.The data processingmethodofA4 further includes: writing, by the general-purpose processor, the device information into a driver to select a wellmatchedcloudAIprocessor based on the device informationin the driver. A6.The data processingmethodofAl,whereinthe determining the offlinerunningfile according to tiie running result includes: ifthe running result satisfies a preset requirement, generating the corresponding offline running file according to the binary instruction that satisfies the preset requirement. A7.The data processingmethodofA1, wherein the determining the offlinerunningfile according to the running result further includes: if tiie running result does not satisfy the preset requirement, executing at least one of the optimization manners in the following process until tiie running result satisfies the preset requirement, andgenerating the correspondingofflinerunning file accordingto tiie binaiy instructionthat satisfies the preset requirement, where the optimization manners include: adjusting the hardware architecture informationof theAI processor, adjusting the parameters ofrunningenvironmentofthe AIprocessor, or optimizingthe AI learning task. A8.The data processingmethodofA7, wherein steps ofoptimizing tiie Allearningtask include: optimizing the AI learning task by a special programming language. A9.The data processingmethodofA7 or A8, wherein the steps ofoptimizingthe Allearning task 235 10 15 20 25 30 CA 3065651 2019-12-19 further include: optimizing the AI learning task by updating the versionof theAl learning task. A10.The data processingmethodof anyofA7-A9, whereinthe stepsofoptimizingtheAIlearning task further include: adjustingthe AI algorithm model. All. The data processing method of any of A1-A10, wherein the running result includes one or more pieces of information of whether running time of the Al learning task on the cloudAI processor satisfies the first expectation, whether load information of a cloud AI processing system when the artificial intelligent learning task is executed satisfies the second expectation, and whether die result of theAI learning task satisfies the thirdexpectation. A12. The data processing methodof Al, wherein the cloud AIprocessor includes at least one of a hardware entity of the AIprocessor, a field programmable gate array, and a simulator. A13. The data processing methodof A2, wherein the hardware entity of the Al processor is an AI processor with a reconfigurable architecture. B14. A data processing device including a memory, a general-purpose processor, and a cloud Al processor, where a computer program running on the general-purpose processor and / or the cloud AI processor is storedin the memory. The general-purpose processor is configured to generate a binary instruction according to device informationoftheAIprocessor,andthen generate anAIlearning task according to thebinaryinstruction, and transmit theAIlearning task to the cloudAI processor for running; the cloud AIprocessor is configured to receive and execute the AI learning task, and then generate arunningresult; and the general-purpose processor is configured to receive the running result corresponding to the AI learning task, and determine an offline running file according to the running result; where the offline running file is generated according to the device information of the AI processor and the binary instruction when the runningresult satisfies apreset requirement. Bl5. The data processing device of Bl4, wherein die device information includes hardware architecture information and parameters ofrunning environment ofthe AIprocessor. B16.The dataprocessingdeviceofB15, wherein the parametersofrunning environment include at least one ofa base clock speedof theAIprocessor, an access bandwidth of an off-chip memory and the AIprocessor, a size of an on-chip memory, the numberof the cores of the AIprocessor, anda type of an operatingunit ofthe AIprocessor. Bl7. The data processing device of Bl4, wherein the general-purpose processor is further configuredto: 24receive the device information ofthe AIprocessor. B18. The data processing device of B17, wherein the general-purpose processor is further configured to: write the device informationinto a driver to select a well-matched cloudAIprocessor based on the 5 device information in the driver. B19. The data processing device of B14, wherein the general-purpose processor is configured to determine the offline running file according to the runningresult, which includes: iftherunning result satisfies apreset requirement, the general-purpose processor isconfigured to generatethe correspondingofflinerunning file according tothebinary instruction that satisfies the preset 10 requirement. B20. The data processing device of B14, wherein the general-purpose processor is further configured to determine the offline running file according to the running result, which includes: if the running result does not satisfy the preset requirement, the general-purpose processor is configuredto execute at least oneoftheoptimizationmannersin the following process untilthe running 15 result satisfies the preset requirement, and generate the corresponding offline running file according to the binary instruction that satisfies the preset requirement, where the optimization manners include: adjusting the hardware architecture information of the AI processor, adjusting the parameters ofrunning environment ofthe Al processor, or optimizing the AI learning task. B21.The data processing deviceofB20, wherein stepsofoptimizingthe AIlearningtask include: 20 optimizing the AI learning task by a special programming language. B22. The data processing device of B20 or B21, wherein the steps of optimizing the AI learning task further include: optimizing the AIlearning task by updating the versionof theAI learning task. B23. The data processing device of any of B20-B22, wherein the steps of optimizing the Al 25 learning task further include: adjusting the AIalgorithmmodel. B24. The data processing device of any of B14-B23, wherein the running result includes one or more pieces of information of whether running time of the AI learning task on the cloudAI processor satisfies the first expectation, whether load information of a cloud AI processing system when the 30 artificialintelligent learning task is executed satisfies the second expectation, and whether the result of theAI learning task satisfies the thirdexpectation. B25.The data processingdevice ofBl4, wherein the cloudAIprocessor includes at least one of a hardware entity of the AIprocessor, a field programmable gate array, and a simulator. B26.The data processing device ofB25, wherein thehardware entity of theAlprocessor is an AI 25 CA 3065651 2019-12-195 10 15 20 25 30 CA 3065651 2019-12-19 processor with a reconfigurable architecture. Cl.A dataprocessing methodappliedto a general-purpose processor,including: generating, by the general-purpose processor, a binary instruction according to device information ofthe Alprocessor, andgenerating anAl learning task according to the binary instruction; transmitting, by the general-purpose processor, the AI learning task to the cloud AI processor for running; receiving,by the general-purpose processor, arunningresult corresponding to the AIlearning task; and determining,by the general-purposeprocessor, anofflinerunning file accordingto therunningresult; wherein the offline running file is generated according to the device information ofthe Alprocessor and the binary instruction when therunningresult satisfies a preset requirement. C2. The data processing method of Cl, wherein the device information includes hardware architecture information and parameters ofrunning environment of the AIprocessor. C3. The data processingmethodof C2, wherein the parameters of running environment include at least one ofa base clock speedof the AI processor, an access bandwidth of an off-chip memory and the Alprocessor, a size ofanon-chip memory, thenumberofthe coresofthe AIprocessor, anda type of an operatingunit ofthe AIprocessor. C4.The data processing methodofCl further includes: receiving,by the general-purpose processor, the device information ofthe Alprocessor. C5.The data processingmethodofCl further includes: writing, by the general-purpose processor, the device information into a driver to select a wellmatchedcloudAIprocessor based on the device informationin the driver. C6.The data processingmethodofC1,wherein the determining the offlinerunningfile according to the running result includes: ifdie running result satisfies a preset requirement, generating the corresponding offline running file according to the binary instructionthat satisfies the preset requirement. C7.The data processingmethodofC1, whereinthe determining the offlinerunningfile according to the runningresult further includes: if the running result does not satisfy the preset requirement, executing at least one of the optimization manners in the following process until the running result satisfies the preset requirement, andgeneratingthe corresponding offline running file according to thebinary instruction that satisfies the preset requirement, where the optimization manners include: adjusting the hardware architecture informationof the AI processor,adjusting the parameters ofrunningenvironment of the AIprocessor, or optimizingthe AI learning task. 265 10 15 20 25 30 CA 3065651 2019-12-19 C8.The data processingmethodofC7, wherein steps ofoptimizing the Allearningtask include: optimizing the AIlearning task by a special programming language. C9.The dataprocessingmethodofC7 or C8, wherein the steps ofoptimizingthe Allearning task further include: optimizing the AIlearning task by updating the version ofdieAl learning task. CIO.The data processingmethodof any of C7-C9, whereinthe stepsofoptimizingtheAIlearning task further include: adjusting the AI algorithm model. Cl1. The data processing method of any of C1-C10, wherein the runningresult includes one or more pieces of information of whether running time of the AI learning task on the cloudAI processor satisfies the first expectation, whether load information of a cloud AI processing system when the artificial intelligent learning task is executed satisfies the second expectation, and whether the result of the Al learningtask satisfies the thirdexpectation. D12. A data processing device including a memory and a general-purpose processor, wherein a computer program that can be run on die general-purpose processor is stored in die memory. Hie data processing flow implemented when the general-purpose processor executes the computer program includes: generating the binary instruction according to the device information of the AI processor, and generatingthe AI learning task according to the binary instruction; transmitting the AI learning task to the cloudAIprocessor torun; receivingthe runningresult ofthe AI learning task; and determining the offline running file according to the runningresult, where theofflinerunning file is generated according to the device information oftheAIprocessor and the binary instruction when the runningresult satisfies a preset requirement. D13. The data processing device of D12, wherein the device information includes die hardware architecture information and the parameters ofrunning environment of the AIprocessor. D14. The data processing device ofD13, wherein the parameters ofrunning environment include at least one of a base clock speed of the Al processor, an access bandwidth of an off-chip mernoiy and theAIprocessor, a size ofan on-chip memory, the number ofthe cores oftheAIprocessor,and a type of an operatingunit of the AIprocessor. D15. The data processing device ofD12, wherein the data processing flow implemented when the general-purpose processor executes the computer program further includes: receiving the device information of the AIprocessor. D16. The data processing device ofDI2 or D15, wherein the data processing flow implemented 275 10 15 20 25 30 CA 3065651 2019-12-19 when the general-purpose processor executes the computer program further includes: writing the device information into the driver to select the well-matched cloud AI processor accordingto the device information in the driver. D17. The data processingdeviceofD12, the dataprocessing flow implemented whenthe generalpurpose processor executes the computer program to determine the offline running file according to the runningresult includes: iftherunning result satisfies the preset requirement, generating the correspondingoffline running file according to the binary instruction that satisfies the preset requirement. D18. The data processingdeviceofD12, the dataprocessing flow implemented when the generalpurpose processor executes the computer program to determine the offline running file according to the runningresult further includes: if the running result does not satisfy the preset requirement, executing at least one of the optimization manners in the following process until the running result satisfies the preset requirement, and generatingthe corresponding offline running file according to the binary instruction that satisfies the preset requirement, where the optimization manners include: adjusting the hardware architecture informationof the AI processor, adjusting the parameters ofrunning environment ofthe AIprocessor, or optimizing the AI learning task. D19. The data processingdevice ofD18, the data processing flow implemented when the generalpurpose processor optimizes the AI learning task includes: optimizing the AI learning task by a special programming language. D20. The data processing device ofDI8 or D19, the data processing flow implemented when the general-purpose processor optimizes the AI learning task further includes: optimizing the AI learning task by updating the version of the Al learning task. D21.The data processing device of any ofD18-D20, the dataprocessing flow implementedwhen the general-purpose processor optimizes the AI learning task further includes: adjusting the Al algorithmmodel. D22. The data processing device of any of D12-D21, the running result includes one or more pieces of information of whether the running time of the AI learning task on the cloud AI processor satisfies the expectation, whether load information of a cloud AI processing system when the artificial intelligent learning task is executed satisfies the expectation, and whether the result of the AI learning task satisfies the expectation. E23.A data processing method applied to a cloudAI processor, including: receiving the AI learning task, where the AI learning task is determined according to the binary instruction generated based on the device information of the Al processor; and 285 10 15 20 25 30 CA 3065651 2019-12-19 executing the AI learning task to generate the running result. E24. The dataprocessingmethodofE23, wherein the runningresult includes one or more pieces of information of whether running time of the Al learning task on the cloud Al processor satisfies the first expectation, whether load information of a cloud Al processing system when the artificial intelligent learning task is executed satisfies the second expectation, and whether the result of the AI teaming task satisfies the third expectation. E25. The data processing methodof E23, wherein the cloudAIprocessor includes at least one of a hardware entity of the AI processor, a fieldprogrammable gate array, and a simulator. E26. The data processing methodof E25, wherein the cloud AI processor is an AI processor with a reconfigurable architecture. F27. AnAIprocessor, including: a receiving module configured to receive the AI learning task, where the AI learning task is determined according to the binary instruction generated based on the device information of the AI processor; and an executing module configured to execute the Al learning task to generate the running result. F28. The Al processor of F27, wherein the running result generated by the executing module includes at least one ofpieces ofinformationof whether the running time of the AI learning task on the cloud Al processor satisfies the first expectation, whether load information of a cloud Al processing system when the artificial intelligent learning task is executed satisfies the second expectation, and whether the result of the AI learning task satisfies the third expectation. F29. The AI processor ofF27, wherein die cloudAI processor includes at least oneofa hardware entity ofthe Al processor, a fieldprogrammable gate array, and a simulator. F30. The AI processor of F29, wherein the cloud AI processor is an AI processor with a reconfigurable architecture. G31. A dataprocessing method applied to an AI processor, including: obtaining anoffline running file, where the offline running file is generated accordingto the device information of the AI processor and the binary instruction when the running result satisfies a preset requirement. G32. The dataprocessingmethodofG31farther includes: transmitting request information, where the request information includes the device information ofthe AIprocessor. G33. The data processing method of G3I, wherein the device information includes the hardware architecture information and the parameters ofrunning environment of the AI processor. G34. The data processing method of G33, wherein the parameters ofrunning environment include 29at least one of a base clock speed of the AI processor, an access bandwidth of an off-chip memory and the AI processor, a size of an on-chip memory, the number ofthe cores of the Al processor,and a type of anoperating unit of the AI processor. H35.An AI processor, including: 5 anobtaining module configured to obtain the offline running file, where the offline running file is generated according to the device information of the Al processor and the binary instruction when the runningresult satisfies the preset requirement. H36.The Alprocessor ofH35, further including: a transmitting module configured to transmit request information, where the request information 10 includes the device information of theAl processor. H37. The AI processor ofH36, wherein the device information includes the hardware architecture information and the parameters ofrunning environment of the AI processor. H38.TheAIprocessor ofH37. wherein the parameters ofrunning environment include at least one of a base clock speed of the Al processor, an access bandwidth of an off-chip memory and the AI 15 processor, a size of an on-chip memory, the number of the cores of the Al processor, and a type of an operatingunit of the AI processor. 30 CA 3065651 2019-12-19

Claims

5 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 CLAIMS What is claimed is:

1. A data processing method implemented in a data processing system comprising a generalpurpose processor and a cloud artificial intelligence (AI) processor, wherein the data processing method includes: writing, by the general-purpose processor, device information of a terminal AI processor into a driver to select the cloud AI processor based on the device information in the driver, so as to simulate the terminal AI processor; converting, by the general-purpose processor, an AI algorithm model to obtain a binary instruction according to device information of a terminal AI processor; generating, by the general-purpose processor, an AI learning task according to the binary instruction;and transmitting, by the general-purpose processor, the AI learning task to the cloud AI processor for running; receiving, by the cloud AI processor, the AI learning task; and executing, by the cloud AI processor, the AI learning task to generate a running result; and receiving, by the general-purpose processor, the running result; if the running result does not satisfy a preset requirement, adjusting, by the general-purpose processor, the device information of the terminal AI processor, thereby adjusting the binary instruction and adjusting the AI learning task generated from the binary instruction, until the running result satisfies the preset requirement; if the running result satisfies the preset requirement, generating, by the general-purpose processor, an offline running file according to the binary instruction, wherein the offline running file is generated based on the device information of the terminal AI processor and the binary instruction when the running result satisfies the preset requirement.

2. The data processing method of claim 1, wherein the device information includes hardware architecture information and parameters of running environment in the terminal AI processor, where the parameters of the running environment include at least one of: a base clock speed of the terminal AI processor, an access bandwidth of an off-chip memory and the terminal AI processor, a size of an onchip memory, a number of cores in the terminal AI processor, and a type of an operating unit in the terminal AI processor.

3. The data processing method of claim 1, further comprising: receiving, by the general-purpose processor, the device information of the terminal AI processor.

4. The data processing method of claim 1, wherein the determining of the offline running file according to the running result further includes: 315 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 if the running result does not satisfy the preset requirement, executing at least one of optimization operations until the running result satisfies the preset requirement, and generating the offline running file according to the binary instruction that satisfies the preset requirement, where the optimization operations include one or more of the following: adjusting hardware architecture information of the terminal AI processor, adjusting parameters of running environment of the terminal AI processor, and optimizing the AI learning task.

5. The data processing method of claim 4, wherein steps of optimizing the AI learning task include one or more of the following: optimizing the AI learning task by a special programming language, optimizing the AI learning task by updating a version of the AI learning task, and adjusting an AI algorithm model.

6. The data processing method of any one of claims 1-5, wherein the running result includes one or more pieces of information of the following: whether running time of theAI learning task on a cloud AI processor satisfies a first expectation, whether load information of a cloud AI processing system when the artificial intelligent learning task is executed satisfies a second expectation, and whether a result of the AI learning task satisfies a third expectation.

7. The data processing method of claim 1, wherein the cloud AI processor includes at least one of the following: a hardware entity the cloud AI processor, a field programmable gate array,and a simulator.

8. A data processing device, comprising a memory, a general-purpose processor, and a cloud AI processor, wherein a computer program running on the general-purpose processor and / or the cloud AI processor is stored in the memory; wherein the general-purpose processor is configured to write device information of a terminal AI processor into a driver to select the cloud AI processor based on the device information in the driver, so as to simulate the terminal AI processor, convert an AI algorithm model to a binary instraction according to device information of the terminal AI processor, generate an AI learning task according to the binary instruction, and transmit the AI learning task to the cloud AI processor for running; the cloud AI processor is configured to receive and execute the AI learning task to generate a running result; and the general-purpose processor is configured to receive the running result corresponding to the AI learning task, and adjust the device information of the terminal AI processor if the running result does not satisfy a preset requirement, thereby adjusting the binary instruction and adjust the AI learning task generated from the binary instruction until the running result satisfies the preset requirement; generate an offline running file according to the binary instruction if the running result satisfies the preset requirement; where the offline running file is generated according to the device information of the 325 10 15 20 25 30 Date Re<;ue / Date Received 2023-12-22 terminal AI processor and the binary instruction when the running result satisfies the preset requirement.

9. The data processing device of claim 8, wherein the device information includes hardware architecture information and parameters of running environment of the terminal AI processor, where the parameters of running environment include at least one of the following: a base clock speed of the terminal AI processor, an access bandwidth of an off-chip memory and the terminal AI processor, a size of an on-chip memory, a number of cores of the terminal AI processor, and a type of an operating unit of the terminal AI processor.

10. The data processing device of claim 8, wherein the general-purpose processor is further configured to: receive the device information of the terminal AI processor.

11. The data processing device of claim 8, wherein the general-purpose processor is further configured to determine the offline running file according to the running result, and wherein: if the running result does not satisfy the preset requirement, the general-purpose processor is configured to execute at least one optimization operation until the running result satisfies the preset requirement, and generate a corresponding offline running file according to the binary instruction that satisfies the preset requirement, where the optimization operation includes one or more of the following: adjusting the hardware architecture information of the terminal AI processor, adjusting the parameters of running environment of the terminal AI processor, and optimizing the AI learning task.

12. The data processing device of claim 11, wherein steps of optimizing the AI learning task include one or more of the following: optimizing the AI learning task by a special programming language, optimizing the AI learning task by updating a version of the AI learning task, and adjusting an AI algorithm model.

13. The data processing device of any one of claims 8-12, wherein the running result includes one or more pieces of information of the following: whether running time of the AI learning task on a cloud AI processor satisfies a first expectation, whether load information of a cloud AI processing system when the artificial intelligent learning task is executed satisfies a second expectation, and whether a result of the AI learning task satisfies a third expectation.

14. The data processing device of claim 8, wherein the cloud AI processor includes at least one of: a hardware entirety of the cloud AI processor, a field programmable gate array, and a simulator. 15.A data processing method applied to a general-purpose processor, comprising: writing, by the general-purpose processor, the device information into a driver to select a wellmatched cloud artificial intelligence (AI) processor based on the device information in the driver; converting, by the general-purpose processor, an AI algorithm model to obtain a binary instruction 335 10 15 20 25 30 Date Reyue / Date Received 2023-12-22 according to device information of the terminal AI processor, and generating an AI learning task according to the binary instruction; transmitting, by the general-purpose processor, the AI learning task to a cloud AI processor for running; receiving, by the general-purpose processor, a running result corresponding to the AI learning task; and if the running result does not satisfy a preset requirement, adjusting, by the general-purpose processor, the device information of the terminal AI processor, thereby adjusting the binary instruction and adjusting the AI learning task generated from the binary instruction, until the running result satisfies the preset requirement; if the running result satisfies the preset requirement, generating, by the general-purpose processor, an offline running file according to the binary instruction; wherein the offline running file is generated according to the device information of the terminal AI processor and the binary instruction when the running result satisfies the preset requirement.

16. The data processing method of claim 15, wherein the device information includes hardware architecture information and parameters of running environment of the terminal AI processor. 17.The data processing method of claim 16, wherein the parameters of running environment include at least one of a base clock speed of the terminal AI processor, an access bandwidth of an offchip memory and the terminal AI processor, a size of an on-chip memory, a number of cores of the terminal AI processor, and a type of an operating unit of the terminal AI processor. 18.The data processing method of claim 15, further comprising: receiving, by the general-purpose processor, the device information of the terminal AI processor. 19.The data processing method of claim 15, wherein the determining the offline running file according to the running result further includes: if the running result does not satisfy the preset requirement, executing at least one of optimization manners in a following process until the running result satisfies the preset requirement, and generating the offline running file according to the binary instruction that satisfies the preset requirement, wherein the optimization manners include: adjusting the hardware architecture information of the terminal AI processor, adjusting the parameters of running environment of the terminal AI processor, or optimizing the AI learning task. 2O.The data processing method of claim 19, wherein steps of optimizing the AI learning task include: optimizing the AI learning task by a special programming language. 21.The data processing method of claim 20, wherein the steps of optimizing the AI learning task 34further include: optimizing the AI learning task by updating a version of the AI learning task. 22.The data processing method of any one of claims 20-21, wherein the steps of optimizing the AI learning task further include: 5 adjusting the AI algorithm model. Date Re<;ue / Date Received 2023-12-22