CN114330689A - Data processing method, device, electronic device and storage medium - Google Patents

Abstract

The embodiment of the disclosure provides a data processing method, a data processing device, an electronic device and a storage medium, wherein the method comprises the following steps: acquiring neural network parameters corresponding to the target neural grid model; generating shaders to be used corresponding to each network level according to the neural network parameters; wherein the target neural network model comprises a plurality of network hierarchies; and determining a computing pipeline corresponding to the shader to be used according to target equipment parameters of the equipment to which the target neural network model belongs, and calling the computing pipeline to process the data to be processed according to the target equipment parameters when the data to be processed is received to obtain a target processing result. According to the technical scheme of the embodiment of the disclosure, the limitation of the GPU communication bandwidth to the model calculation process is reduced in a mode of greatly reducing the interaction between the CPU and the GPU, and the performance of the neural network model is improved.

Description

Data processing method, device, electronic device and storage medium

technical field

The embodiments of the present disclosure relate to the technical field of data processing, and in particular, to a data processing method, apparatus, electronic device, and storage medium.

Background technique

With the continuous development of artificial intelligence technology, the neural network model has been widely used in various fields. Relying on its characteristics of large-scale parallel processing and distributed information storage, it can process a variety of data.

However, when the computer performs calculations based on the neural network model, a large number of communication interactions between the central processing unit (CPU) and the graphics processing unit (Graphics Processing Unit, GPU) reduce the computational efficiency of the neural network, and there is a bottleneck in the performance of the model. .

SUMMARY OF THE INVENTION

The present disclosure provides a data processing method, device, electronic device and storage medium, which greatly reduce the interaction between the CPU and the GPU, reduce the limitation of the GPU communication bandwidth on the model calculation process, and improve the performance of the neural network model.

In a first aspect, an embodiment of the present disclosure provides a data processing method, which is applied to a central processing unit, including:

Obtain the neural network parameters corresponding to the target neural grid model;

According to the neural network parameters, a shader to be used corresponding to each network level is generated; wherein, the target neural network model includes a plurality of network levels;

Determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so as to call the computing pipeline according to the target device parameters when receiving the data to be processed The data to be processed is processed to obtain a target processing result.

In a second aspect, an embodiment of the present disclosure further provides a data processing method, which is applied to a graphics processor, including:

When receiving the data to be processed, load the pre-determined computing pipeline and neural network parameters; wherein, the computing pipeline is based on the neural network parameters of the target neural network model by the central processing unit and the target neural network model belongs to The target device parameters are determined;

According to the target device parameters, determine the target processing mode of each shader to be used to process the data to be processed; wherein, the shader to be used is obtained by processing the neural network parameters based on the central processing unit;

The data to be processed is processed based on the target processing manner to obtain a target processing result.

In a third aspect, an embodiment of the present invention further provides a data processing device, the device is configured in a central processing unit, and includes:

The network parameter determination module is used to obtain the neural network parameters corresponding to the target neural grid model;

a shader determination module, configured to generate shaders to be used corresponding to each network level according to the neural network parameters; wherein, the target neural network model includes a plurality of network levels;

A pipeline determination module, configured to determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so that when the data to be processed is received, according to the target device parameters, The computing pipeline is called to process the to-be-processed data to obtain a target processing result.

In a fourth aspect, an embodiment of the present invention further provides a data processing apparatus, where the apparatus is configured in a graphics processor and includes:

The network parameter loading module is used to load the pre-determined calculation pipeline and neural network parameters when receiving the data to be processed; wherein, the calculation pipeline is based on the neural network parameters and the neural network parameters of the target neural network model based on the central processing unit. Determined by the parameters of the target device to which the target neural network model belongs;

A processing mode determination module, configured to determine, according to the target device parameters, a target processing mode for each shader to be used to process the data to be processed; wherein, the shader to be used is based on the processing of the neural network parameters by the central processing unit obtained after

A processing result determination module, configured to process the data to be processed based on the target processing manner to obtain a target processing result.

In a fifth aspect, an embodiment of the present disclosure further provides an electronic device, the electronic device comprising:

one or more processors;

storage means for storing one or more programs,

When the one or more programs are executed by the one or more processors, the one or more processors implement the data processing method according to any one of the embodiments of the present disclosure.

In a sixth aspect, an embodiment of the present disclosure further provides a storage medium containing computer-executable instructions, when executed by a computer processor, the computer-executable instructions are used to perform the data processing according to any one of the embodiments of the present disclosure method.

According to the technical solution of the embodiment of the present disclosure, the neural network parameters corresponding to the target neural network model are obtained; according to the neural network parameters, the to-be-used shaders corresponding to each network level are generated, that is, for multiple shaders in the target neural network model The network level generates the corresponding shader to be used; further, according to the target device parameters of the device to which the target neural network model belongs, the computing pipeline corresponding to the shader to be used is determined, so that when the data to be processed is received, according to the target device parameters , call the computing pipeline to process the data to be processed, obtain the target processing result, generate the shader corresponding to each network level of the neural network model through the CPU, and then perform coarse-grained calculation based on each shader after receiving the data to be processed, In a way that greatly reduces the interaction between the CPU and the GPU, the limitation of the GPU communication bandwidth on the model calculation process is reduced, and the performance of the neural network model is improved.

Description of drawings

The above and other features, advantages and aspects of various embodiments of the present disclosure will become more apparent when taken in conjunction with the accompanying drawings and with reference to the following detailed description. Throughout the drawings, the same or similar reference numbers refer to the same or similar elements. It should be understood that the drawings are schematic and that the originals and elements are not necessarily drawn to scale.

1 is a schematic flowchart of a data processing method according to Embodiment 1 of the present disclosure;

2 is a schematic flowchart of a data processing method according to Embodiment 2 of the present disclosure;

3 is a schematic flowchart of a data processing method provided in Embodiment 3 of the present disclosure;

4 is a schematic flowchart of a data processing method provided in Embodiment 4 of the present disclosure;

FIG. 5 is a schematic structural diagram of a data processing apparatus according to Embodiment 5 of the present disclosure;

6 is a schematic structural diagram of a data processing apparatus according to Embodiment 6 of the present disclosure;

FIG. 7 is a schematic structural diagram of an electronic device according to Embodiment 7 of the present disclosure.

Detailed ways

Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but rather are provided for the purpose of A more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes, and are not intended to limit the protection scope of the present disclosure.

It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and/or in parallel. Furthermore, method embodiments may include additional steps and/or omit performing the illustrated steps. The scope of the present disclosure is not limited in this regard.

As used herein, the term "including" and variations thereof are open-ended inclusions, ie, "including but not limited to". The term "based on" is "based at least in part on." The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions of other terms will be given in the description below.

It should be noted that concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or interdependence. It should be noted that the modifications of "a" and "a plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless the context clearly indicates otherwise, they should be understood as "one or a plurality of". multiple".

The names of messages or information exchanged between multiple devices in the embodiments of the present disclosure are only for illustrative purposes, and are not intended to limit the scope of these messages or information.

Example 1

FIG. 1 is a schematic flowchart of a data processing method according to Embodiment 1 of the present disclosure. The embodiment of the present disclosure is suitable for enabling a central processing unit to generate a coarse-grained shader according to neural network parameters, so that when receiving data to be processed, based on When the shader corresponding to each network level is to be used to quickly process the data, the method can be executed by a data processing device, which can be implemented in the form of software and/or hardware, or, optionally, through an electronic device. For implementation, the electronic device may be a mobile terminal, a PC terminal, a server, or the like.

As shown in Figure 1, the method includes:

S110. Obtain neural network parameters corresponding to the target neural grid model.

It should be noted that the solution in this embodiment can be implemented based on a target device, wherein the target device can be any terminal device equipped with a CPU and a GPU. After the target device receives the data to be processed, it can use the CPU and GPU The processing capability processes the received data.

Among them, the neural network model (Neural Networks, NN) is a mathematical method to simulate the actual human neural network. Specifically, each neural network is widely interconnected by a large number of simple processing units (which can be called neurons). The formed complex network system can at least reflect many basic features of human brain function, and is a highly complex nonlinear dynamic learning system. Therefore, the neural network model has large-scale parallel, distributed storage and processing, self-organization, self-adaptation and self-learning capabilities, and is suitable for dealing with inaccurate and ambiguous information processing problems that need to consider many factors and conditions at the same time.

In this embodiment, the target neural network model may be one or more neural network models associated with a specific service. For example, when a graph-related service needs to be processed, a corresponding convolutional neural network (Convolutional Neural Network) can be determined. Neural Networks, CNN) model, when the business needs to be processed with multiple time periods, a corresponding Recurrent Neural Network (RNN) model can be determined.

In this embodiment, in order for the computer device to perform the calculation corresponding to the target neural network model, it is also necessary to acquire the neural network parameters corresponding to the target neural network. Among them, the neural network parameters are the multi-dimensional algorithm information involved in the target neural network model, which can include the identifier of the target neural network model, the network structure, the number of neural network layers, the tensors of the neural network, the execution order of each layer, and the number of layers of the neural network. At least one of the input and output of the layer. For example, when the target neural network model is a CNN model, the determined neural network parameters can be the data types of the input and output of the CNN, and the execution order of the input layer, convolution layer, activation function, pooling layer, and fully connected layer . It can be understood that, based on the acquired neural network parameters, the computer can at least construct the target neural network model in the memory.

It can be understood that the structure of the target neural network model can be set according to the actual situation, and its specific neural network parameters correspond to each network level of the model. Build to obtain the target neural network model.

It should be noted that, in the actual application process, the neural network parameters corresponding to each target business can be obtained in advance and stored in the corresponding server, and called when the model-related calculation needs to be performed, or the target neural network can be executed when the target business needs to be executed. During model-related heterogeneous computing, the required neural network parameters are obtained in a dynamic manner in real time. Those skilled in the art should understand that the specific manner of acquiring the neural network parameters can be selected according to the actual situation, which is not specifically limited in this embodiment of the present disclosure.

S120. Generate shaders to be used corresponding to each network level according to the neural network parameters.

In this embodiment, the heterogeneous computing of neural networks is a special form of parallel and distributed computing. It can be understood that the performance characteristics required by heterogeneous computing are actually very similar to graph-related algorithms, usually involving a large number of parameters. , buffers of activation values, gradient values, each of which is updated during each training iteration or data processing process, these buffers are too large and will exceed the traditional desktop computer cache, therefore, can be used with high memory bandwidth GPU to perform neural network related computations.

Therefore, in this embodiment, after acquiring the neural network parameters, in order to use the GPU to process the data when receiving the business-related data, it is first necessary to use the CPU to generate the corresponding target neural network model, which can be run in The shader to use in the GPU. Among them, the shader to be used can be understood as a calculation shader unrelated to conventional graphics rendering. When the GPU loads the generated calculation shader, the calculation related to the target neural network model can be performed.

At the same time, since the target neural network model determined by the CPU includes multiple network layers, when generating shaders to be used, it is usually necessary to generate corresponding shaders to be used for each network layer. Specifically, the shader generation module needs to be called to process the neural network parameters to obtain shaders to be used at each network level in the target neural network model. Among them, the shader generation module refers to the processing module deployed on the CPU side, and at least has the shader to be used for construction. Since the neural network parameters correspond to each network level of the target neural network model, the CPU uses the shader generation module. Using shaders also corresponds to each network level of the model.

In this embodiment, the CPU constructs coarse-grained shaders to be used for each network level of the target neural network model, instead of constructing fine-grained shaders for specific operators in each network level, so that in the process of neural network calculation , avoids frequent interaction between CPU and GPU, and reduces the limitation of GPU communication bandwidth on neural network performance.

Exemplarily, after it is determined that the target neural network model is a CNN model, and the neural network parameters of the CNN model are obtained, the CPU can determine the input layer, convolution layer, pooling layer and the CNN based on the obtained neural network parameters. After the fully connected layer, the shader to be used corresponding to each network level can be generated. It can be understood that after generating the corresponding shader to be used for each layer of the target neural network model, which shader will be executed by the GPU in the subsequent process , which means to perform the computation at the network level corresponding to this shader.

S130, according to the target device parameters of the device to which the target neural network model belongs, determine the computing pipeline corresponding to the shader to be used, so that when receiving the data to be processed, according to the target device parameters, call the computing pipeline to process the data to be processed, Get the target processing result.

In this embodiment, since not all GPUs use the same instruction set, only by converting the program into a binary file can the GPU load and use the shaders to be used corresponding to each network level based on its own architecture set. Therefore, after the CPU generates shaders to be used corresponding to each network level of the target neural network model, it is also necessary to invoke the driving capability of the GPU to generate computing pipelines corresponding to the shaders to be used.

Among them, for the computer, the calculation pipeline can be understood as the pipeline required by the GPU to run the program, and the process of determining the calculation pipeline can be understood as the CPU calling the GPU's driving ability, based on the preset data structure. The process of compiling, packaging, and storing into the target cache. Exemplarily, after pre-setting the json format as the data structure of the computing pipeline, the CPU can invoke the driving capability of the GPU to compile, encapsulate and cache the shaders to be used. Data is the key data in each shader to be used. When processing specific business data later, the GPU can load these key data as the calculation pipeline based on the target cache, so as to execute the execution of each network level according to the target neural network model. The corresponding heterogeneous computations are performed sequentially.

In this embodiment, each computing pipeline may be determined according to target device parameters of the device to which the target neural network model belongs. The target device may be a device that performs computation related to the target neural network model based on the GPU, and correspondingly, the target device parameter is attribute information of the GPU mounted on the target device.

In the actual application process, the parameters of the target device include the indirect buffer parameter (IndirectBuffer), which can be understood as the information indicating whether the GPU supports the indirect buffer function. When processing, the execution order of the loaded shaders to be used is determined based on itself. When the GPU does not support indirect buffering, when the GPU is processing business data, it needs to execute a shader to be used. Instructions sent by the CPU to determine the next shader to be used that needs to be executed.

It can be understood that when the GPU supports indirect caching, the overall heterogeneous computing process of the target neural network model can be completed without the need for the CPU to send instructions. In the process, multiple interactions with the CPU are also required, and the next shader to be used to be executed is determined based on each instruction sent by the CPU.

Based on this, in the case of different target device parameters, there are also differences in the way the CPU determines each computing pipeline. It can be understood that, according to the judgment on whether the GPU supports indirect caching, the CPU has two corresponding ways to determine the The compute pipeline corresponding to the shader.

The first way is, if the indirect buffer parameter of the device to which the target neural network model belongs is the first parameter, generate a scheduling shader corresponding to each shader to be used; by compiling the scheduling shader and the shader to be used, Get the computation pipeline.

Specifically, when the indirect buffer parameter is the first parameter, it means that the GPU mounted on the target device supports indirect buffering, and therefore, the scheduling shader corresponding to each shader to be used can be generated. The scheduling shader may be understood as the program code running in the GPU and used to control the execution order of each shader to be used. For example, when it is determined that the target device parameter is the first buffer parameter, corresponding scheduling shaders can be generated for the input layer, convolutional layer, pooling layer and fully connected layer in the CNN model, and when the related image data is subsequently processed , the GPU can determine the execution order of each layer in the CNN model based on the scheduling shader. It can be understood that after the GPU finishes executing the shader to be used corresponding to the input layer, the next shader to be executed can be determined according to the scheduling shader. The shader is the shader to be used corresponding to the convolutional layer.

In this embodiment, after the scheduling shader and each shader to be used are generated, each generated shader can be compiled based on the CPU to obtain a computing pipeline corresponding to each shader. Among them, the computing pipeline includes sub-computation pipelines corresponding to the shaders to be used, and control pipelines corresponding to each sub-computation pipeline. It contains key data used to control the sequence-related procedures at each network level.

The second way is, if the indirect buffer parameter of the device to which the target neural network model belongs is the second parameter, determine the sub-computing pipeline corresponding to each shader to be used; and determine the computing pipeline based on each sub-computing pipeline .

Specifically, when the indirect buffer parameter is the second parameter, it means that the GPU mounted on the target device does not support indirect buffering. Therefore, it is not necessary to generate the scheduling shader corresponding to each shader to be used, but to directly generate the corresponding shader to be used. The sub-computation pipeline corresponding to the shader, and further, the determined sub-computation key is integrated, that is, the computation pipeline corresponding to the target neural network model is obtained. It can be understood that when the indirect buffer parameter is the second parameter, when the GPU processes each service data, the execution order of each shader to be used cannot be determined based on itself. Therefore, it is necessary to execute each segment based on the instructions sent by the CPU. Use shaders.

In this embodiment, after the calculation pipeline is determined, the CPU can cache the neural network parameters and the calculation pipeline of the target neural network model, so as to call the neural network parameters and the calculation pipeline to process the to-be-processed data when receiving the to-be-processed data . The data to be processed may be data related to the business associated with the target neural network model, and at the same time, the data to be processed is also the input of the target neural network model.

Exemplarily, after the CPU determines the neural network parameters of the CNN model, the sub-computation pipelines corresponding to each network level, and the control pipelines corresponding to the scheduling shaders, these data can be stored in the target cache. After receiving the business data associated with the CNN model (that is, the data to be processed), the GPU of the target device can load the sub-computing pipeline and the control pipeline corresponding to the target neural network model based on the target cache to process the data to be processed, that is, , under the scheduling of the control pipeline, first input the business data to the input layer, after the shader to be used corresponding to the input layer is executed, the output result of this layer is input to the convolutional layer, The pooling layer and the fully connected layer, and execute the shader to be used corresponding to each network level, so as to obtain the processing result corresponding to the data to be processed for the service.

According to the technical solution of the embodiment of the present disclosure, the neural network parameters corresponding to the target neural network model are obtained; according to the neural network parameters, the to-be-used shaders corresponding to each network level are generated, that is, for multiple shaders in the target neural network model The network level generates the corresponding shader to be used; further, according to the target device parameters of the device to which the target neural network model belongs, the computing pipeline corresponding to the shader to be used is determined, so that when the data to be processed is received, according to the target device parameters , call the computing pipeline to process the data to be processed, obtain the target processing result, generate the shader corresponding to each network level of the neural network model through the CPU, and then perform coarse-grained calculation based on each shader after receiving the data to be processed, In a way that greatly reduces the interaction between the CPU and the GPU, the limitation of the GPU communication bandwidth on the model calculation process is reduced, and the performance of the neural network model is improved.

Embodiment 2

As an optional embodiment of the foregoing embodiment, FIG. 2 is a schematic flowchart of a data processing method provided in Embodiment 2 of the present disclosure. In order to clearly introduce the technical solution of this embodiment, it can be introduced by taking the application scenario that the shader program is dynamically generated based on the central processing unit, and the situation that the target device processes the data to be processed when receiving the data to be processed as an example, but It is not limited to the above scenarios, and can be applied to various scenarios where data related to the target neural network model needs to be processed.

Referring to Figure 2, in the process of building a neural network by the CPU, it is first necessary to obtain the model parameters of the target neural network model, which can be determined based on the configuration parameters input by the caller. After entering the relevant configuration parameters of the CNN model, the processing module on the CPU side can obtain these parameters and use them as neural network parameters.

Continue to refer to FIG. 2, after the CPU determines the target neural network model and the corresponding neural network parameters, it can generate the shader to be used for neural network calculation for each level of the model; further, determine the GPU of the target device. Whether indirect cache is supported, when it is determined that the GPU supports indirect cache, the CPU also needs to generate scheduling shaders at each level, and call the GPU driver to compile the shaders and the shaders to be used, and get the corresponding computing pipeline, understandable , the computing pipeline includes sub-computing pipelines corresponding to the shaders to be used, and control pipelines corresponding to the scheduling shaders. When it is determined that the GPU does not support indirect caching, the CPU directly generates sub-computation pipelines corresponding to the shaders to be used.

Continuing to refer to Figure 2, after the CPU generates the calculation pipeline, the neural network parameters and the calculation pipeline can be cached, so that when the data to be processed is received, the associated GPU can retrieve the above data in the cache and process the data to be processed. .

In the technical solution of the embodiments of the present disclosure, shaders corresponding to each network level of the neural network model are generated by the CPU, and then coarse-grained calculations are performed based on each shader after receiving the data to be processed, so as to greatly reduce the difference between the CPU and the GPU. The interactive method reduces the limitation of the GPU communication bandwidth on the model calculation process and improves the performance of the neural network model.

Embodiment 3

FIG. 3 is a schematic flowchart of a data processing method provided by Embodiment 3 of the present disclosure. The embodiment of the present disclosure is applicable to a situation in which a computing pipeline in a cache is loaded based on a graphics processor to process data to be processed. The processing apparatus can be implemented in the form of software and/or hardware. Optionally, it can be implemented by an electronic device, and the electronic device can be a mobile terminal, a PC terminal, a server, or the like.

As shown in Figure 3, the method includes:

S210. When receiving the data to be processed, load the pre-determined computing pipeline and neural network parameters.

In this embodiment, since the computing pipeline is determined by the central processing unit on the neural network parameters of the target neural network model and the parameters of the target device to which the target neural network model belongs, at the same time, the CPU has already calculated the neural network parameters of each target neural network model, And the computing pipeline is stored in the target cache. Therefore, when the target device receives the data to be processed, the CPU first needs to load the determined data based on the cache. Further, the GPU on the device is based on the application programming that interacts with the CPU. Interface (Application Programming Interface, API), load the target neural network model and each computing pipeline, that is, deploy the model, that is, the computing pipeline, on the GPU to process the data to be processed.

Exemplarily, when the CPU determines the calculation pipeline based on the neural network parameters of the CNN model and the target device parameters of the target device (that is, the information representing whether the target device supports indirect caching), the neural network parameters and the calculation pipeline are pre-stored in the target device. After the cache, when the target device receives the pending data of the business associated with the CNN model, the CPU can load the CNN model and the corresponding computing pipelines, and at the same time, send a program execution instruction to the GPU, so as to use the corresponding API interface Deploy the model and computing pipeline in the GPU, that is, make the GPU load the CNN model corresponding to the service, and the computing pipeline corresponding to the input layer, convolution layer, pooling layer, and fully connected layer, so as to execute the above-mentioned functions. Heterogeneous computing corresponding to the network level.

S220. Determine, according to the target device parameters, a target processing mode for each shader to be used to process the data to be processed.

The shader to be used is obtained after processing the neural network parameters based on the central processing unit, that is to say, the CPU can process the neural network parameters according to the method of the first embodiment of the present disclosure, so as to obtain the parameters corresponding to the target neural network model. The shader to be used corresponding to the network level will not be repeated in this embodiment of the present disclosure.

In this embodiment, when the parameters of the target devices are different, the processing methods of the GPU are also different. Optionally, if the target device parameter is the first parameter, determine the target processing mode of each shader to be used to process the data to be processed as the first target processing mode; if the target device parameter is the second parameter, determine the target processing mode The processing method is the second target processing method.

It can be understood that when it is determined that the GPU mounted on the target device supports indirect caching, the first target processing method is determined, so as to process the data to be processed based on each shader to be used, and when it is determined that the GPU mounted on the target device does not support indirect caching , the second target processing mode is determined, so as to process the data to be processed based on each shader to be used.

S230. Process the data to be processed based on the target processing method to obtain a target processing result.

In this embodiment, optionally, when the target processing mode is the first target processing mode, the execution order of each sub-computing pipeline is determined based on the control pipeline corresponding to the computing pipeline; based on the execution order and the corresponding The sub-computing pipeline sends a program execution instruction to the corresponding shader to be used, so that each shader to be used processes the data to be processed to obtain a processing result corresponding to the target.

Exemplarily, when the GPU mounted on the target device supports indirect caching, it may be determined that the target processing mode is the first target processing mode. At the same time, since the CPU has constructed corresponding sub-computation pipelines for each network level of the CNN model as the target neural network model, as well as control pipelines that control the execution order of each network level, and at the same time, the above-mentioned computing pipelines are stored in the target cache. Therefore, When processing the data to be processed, the GPU can directly load the control pipeline in the cache to determine the execution order as input layer, convolution layer, pooling layer and fully connected layer. Further, based on the determined execution order, the GPU can use the input layer sub-computation pipeline, the convolution layer sub-computation pipeline, the pooling layer sub-computation pipeline, and the fully-connected layer sub-computation pipeline to colorize the corresponding to-be-used layers at each network level in turn. The processor sends program execution instructions to run each segment of the program to process the data to be processed.

Optionally, receive a program execution instruction sent based on the current sub-computing pipeline; process the data to be processed based on the program execution instruction, obtain a data processing result, and feed back the data processing result to the central processing unit, so that the central processing unit When the data processing result is received, the next sub-computation pipeline of the current sub-computation pipeline is used as the current sub-computation pipeline, and the program execution instruction is sent based on the current sub-computation pipeline until the current sub-computation pipeline is the last sub-computation pipeline, and the The target processing result corresponding to the processing data.

Specifically, when the GPU mounted on the target device does not support indirect caching, it may be determined that the target processing mode is the second target processing mode. In this case, the CPU only constructs the corresponding sub-computing pipelines for each network level of the target neural network model, and stores the above-mentioned computing pipelines in the target cache. Therefore, when processing the data to be processed, the GPU cannot determine based on itself. It is necessary to process the data to be processed according to the program execution instructions sent by the CPU. It can be understood that the program execution instructions are based on the sub-calculation of the central processor according to each neural network level. pipeline is determined.

Exemplarily, when the CPU only constructs corresponding sub-computing pipelines for each network level of the CNN model as the target neural network model, and stores the above-mentioned computing pipelines in the target cache, the GPU cannot determine when processing the data to be processed. To get the execution order of each shader to be used, it is necessary to receive the program execution instructions sent by the CPU and parse the instructions, so as to determine that the first shader to be used to be executed is the shader to be used corresponding to the input layer of the CNN , further, load the corresponding shader to be used based on the sub-computing pipeline corresponding to the input layer and execute it, so as to process the data to be processed; at the same time, the processing module on the CPU side can asynchronously query the execution result of the GPU, or receive the GPU According to the feedback information sent by the execution result, when the CPU determines that the shader to be used associated with the input layer executed by the GPU is correct and obtains valid data processing results, it can determine for the GPU that the next network layer to be executed is the convolution layer , and after the sub-computation pipeline corresponding to the convolution layer is used as the current sub-computation pipeline, the corresponding message is sent to the GPU again in the form of program execution instructions, and the GPU executes the pending sub-computation pipeline corresponding to the CNN convolution layer based on the current sub-computation pipeline. Use the shader and execute it to get the corresponding processing result again. Based on this method, the GPU executes the pooling layer and the shader to be used corresponding to the fully connected layer in sequence, and obtains the final target processing result.

The above process can be understood as, after the CPU constructs corresponding sub-computing pipelines for the four network layers of the target neural network model, it can send instructions to the GPU, so that the GPU executes the first-layer based on the sub-computing pipelines corresponding to the first layer. When the shader is to be used, when the GPU execution is completed, it sends a message that the program execution is complete to the CPU. After receiving the message, the CPU can send the instruction to execute the shader to be used in the second layer to the GPU, and so on, until the GPU until all shaders to be used at each network level are executed. After the GPU obtains the target processing result, the target processing result can be directly read by the CPU, or the target processing result can be stored on the GPU side for subsequent calls by other shaders to be used.

According to the technical solution of the embodiment of the present disclosure, when the data to be processed is received, the pre-determined parameters of the computing pipeline and neural network are loaded, and the target processing method of each shader to be used for processing the data to be processed is determined according to the parameters of the target device; The target processing method processes the data to be processed and obtains the target processing result, so that after receiving the data to be processed, the GPU performs coarse-grained calculations based on each shader in the cache, which greatly reduces the interaction between the CPU and the GPU. , which reduces the limitation of the GPU communication bandwidth on the model calculation process and improves the performance of the neural network model.

Embodiment 4

As an optional embodiment of the foregoing embodiment, FIG. 4 is a schematic flowchart of a data processing method provided in Embodiment 4 of the present disclosure. In order to clearly introduce the technical solution of this embodiment, the application scenario may be based on the graphics processor loading the computing pipeline in the cache, and the case where the data to be processed is processed as an example may be introduced. In a scenario where data related to the target neural network model needs to be processed.

Referring to Figure 4, in the process of calling the neural network on the GPU, the processing module on the CPU side first needs to load the built computing pipeline and the parameters of the neural network. At the same time, the processing module on the CPU side also needs to load the business data, which is loaded by the GPU based on the cache Calculate the pipeline.

Continue to refer to Figure 4, when the above data is loaded, it is necessary to judge whether the GPU supports indirect buffering. When the GPU supports indirect buffering, the GPU-side processing module can directly load the control pipeline into the indirect buffer, and determine based on the control pipeline. The execution order of each network level of the target neural network model, and further, according to the determined execution order, the corresponding shader programs to be used are executed based on the sub-computing pipelines corresponding to each network level in turn, so as to complete the scheduling of the entire process of the target neural network model ; When the GPU does not support indirect buffering, the CPU-side processing module needs to schedule each network level of the target neural network model. Under the control of the CPU, the GPU can determine the execution order of each layer of the target neural network model. The CPU executes the corresponding shaders to be used in the order given by the CPU.

Continuing to refer to Figure 4, when the GPU processes the data to be processed based on each shader program to be used, the CPU-side processing module can determine whether the target neural network model has been executed through an asynchronous query. Read the target processing result, or save the target processing result on the GPU side for use by other shaders to be used.

The technical solutions of the embodiments of the present disclosure enable the GPU to perform coarse-grained calculations based on each shader in the cache after receiving the data to be processed, so as to greatly reduce the interaction between the CPU and the GPU and reduce the communication bandwidth of the GPU. The limitation of the model calculation process improves the performance of the neural network model.

Embodiment 5

FIG. 5 is a schematic structural diagram of a data processing apparatus according to Embodiment 5 of the present disclosure. As shown in FIG. 5 , the apparatus includes: a network parameter determination module 310 , a shader determination module 320 , and a pipeline determination module 330 .

The network parameter determination module 310 is configured to acquire the neural network parameters corresponding to the target neural grid model.

The shader determination module 320 is configured to generate, according to the neural network parameters, a shader to be used corresponding to each network level; wherein, the target neural network model includes a plurality of network levels.

The pipeline determination module 330 is configured to determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so that when the data to be processed is received, according to the target device parameters , call the computing pipeline to process the data to be processed, and obtain the target processing result.

Optionally, the shader determination module 320 is further configured to call the shader generation module to process the neural network parameters to obtain the shader to be used for each network level in the target neural network model.

Based on the above technical solutions, the target device parameters include indirect buffer parameters.

Based on the above technical solutions, the pipeline determination module 330 includes a scheduling shader generation unit and a calculation pipeline determination unit.

The scheduling shader generating unit is configured to generate a scheduling shader corresponding to each shader to be used if the indirect buffer parameter of the device to which the target neural network model belongs is the first parameter.

a calculation pipeline determination unit, configured to obtain a calculation pipeline by compiling the scheduled shader and the to-be-used shader; wherein the calculation pipeline includes sub-calculation pipelines corresponding to each to-be-used shader, and Control pipelines corresponding to each sub-computing pipeline.

Based on the above technical solutions, the pipeline determination module 330 further includes a sub-computing pipeline determination unit.

The sub-computing pipeline determining unit is configured to determine the sub-computing pipeline corresponding to each shader to be used if the indirect buffer parameter of the device to which the target neural network model belongs is the second parameter.

The computing pipeline determining unit is further configured to determine the computing pipeline based on each sub computing pipeline.

In the technical solution provided by this embodiment, the neural network parameters corresponding to the target neural network model are obtained; according to the neural network parameters, the shaders to be used corresponding to each network level are generated, that is, for the most of the target neural network model Each network level generates the corresponding shader to be used; further, according to the target device parameters of the device to which the target neural network model belongs, determine the computing pipeline corresponding to the shader to be used, so that when the data to be processed is received, according to the target device Parameters, call the computing pipeline to process the data to be processed, get the target processing result, generate the shader corresponding to each network level of the neural network model through the CPU, and then perform coarse-grained calculation based on each shader after receiving the data to be processed , which greatly reduces the interaction between the CPU and the GPU, reduces the limitation of the GPU communication bandwidth on the model calculation process, and improves the performance of the neural network model.

The data processing apparatus provided by the embodiment of the present disclosure can execute the data processing method provided by any embodiment of the present disclosure, and has functional modules and beneficial effects corresponding to the execution method.

It is worth noting that the units and modules included in the above device are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only For the convenience of distinguishing from each other, it is not used to limit the protection scope of the embodiments of the present disclosure.

Embodiment 6

FIG. 6 is a schematic structural diagram of a data processing apparatus according to Embodiment 6 of the present disclosure. As shown in FIG. 6 , the apparatus includes: a network parameter loading module 410 , a processing mode determination module 420 , and a processing result determination module 430 .

The network parameter loading module 410 is used to load the pre-determined computing pipeline and neural network parameters when receiving the data to be processed; wherein, the computing pipeline is based on the neural network parameters of the target neural network model based on the central processing unit and the parameters of the target device to which the target neural network model belongs.

The processing mode determination module 420 is configured to determine, according to the target device parameters, the target processing mode of each shader to be used to process the data to be processed; wherein, the shader to be used is based on the central processing unit's adjustment of the neural network parameters obtained after processing.

The processing result determination module 430 is configured to process the data to be processed based on the target processing mode to obtain a target processing result.

Optionally, the processing mode determination module 420 is further configured to, if the target device parameter is the first parameter, determine the target processing mode of each shader to be used to process the data to be processed as the first target processing mode; if the If the target device parameter is the second parameter, the target processing mode is determined to be the second target processing mode.

Optionally, the processing result determination module 430 is further configured to determine the execution order of each sub-computation pipeline based on the control pipeline corresponding to the computation pipeline; based on the execution order and the corresponding sub-computation pipeline, color the corresponding to-be-used pipelines. The shader sends a program execution instruction, so that each shader to be used processes the data to be processed to obtain a processing result related to the target.

Optionally, the processing result determination module 430 is further configured to receive a program execution instruction sent based on the current sub-computing pipeline; wherein, the program execution instruction is determined based on the central processing unit according to the sub-computing pipeline of each neural network level; based on The program executes the instructions, processes the data to be processed, obtains an image processing result, and feeds back the image processing result to the central processing unit, so that when the central processing unit receives the image processing result, it will The next sub-computation pipeline of the current sub-computation pipeline is used as the current sub-computation pipeline, and a program execution instruction is sent based on the current sub-computation pipeline until the current sub-computation pipeline is the last sub-computation pipeline. The target processing result corresponding to the data to be processed.

In the technical solution provided by this embodiment, when the data to be processed is received, the pre-determined computing pipeline and neural network parameters are loaded, and the target processing mode of each shader to be used for processing the data to be processed is determined according to the target device parameters; The data to be processed is processed based on the target processing method, and the target processing result is obtained, so that the GPU can perform coarse-grained calculations based on the shaders in the cache after receiving the data to be processed, so as to greatly reduce the interaction between the CPU and the GPU. In this way, the limitation of the GPU communication bandwidth on the model calculation process is reduced, and the performance of the neural network model is improved.

The data processing apparatus provided by the embodiment of the present disclosure can execute the data processing method provided by any embodiment of the present disclosure, and has functional modules and beneficial effects corresponding to the execution method.

It is worth noting that the units and modules included in the above device are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only For the convenience of distinguishing from each other, it is not used to limit the protection scope of the embodiments of the present disclosure.

Embodiment 7

FIG. 7 is a schematic structural diagram of an electronic device according to Embodiment 7 of the present disclosure. Referring next to FIG. 7 , it shows a schematic structural diagram of an electronic device (eg, a terminal device or a server in FIG. 7 ) 500 suitable for implementing an embodiment of the present disclosure. Terminal devices in the embodiments of the present disclosure may include, but are not limited to, such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablets), PMPs (portable multimedia players), vehicle-mounted terminals (eg, mobile terminals such as in-vehicle navigation terminals), etc., and stationary terminals such as digital TVs, desktop computers, and the like. The electronic device shown in FIG. 7 is only an example, and should not impose any limitation on the function and scope of use of the embodiments of the present disclosure.

As shown in FIG. 7 , an electronic device 500 may include a processing device (eg, a central processing unit, a graphics processor, etc.) 501 that may be loaded into random access according to a program stored in a read only memory (ROM) 502 or from a storage device 506 Various appropriate actions and processes are executed by the programs in the memory (RAM) 503 . In the RAM 503, various programs and data necessary for the operation of the electronic device 500 are also stored. The processing device 501 , the ROM 502 , and the RAM 503 are connected to each other through a bus 504 . An edit/output (I/O) interface 505 is also connected to bus 504 .

In general, the following devices may be connected to the I/O interface 505: editing devices 506 including, for example, a touch screen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; including, for example, a liquid crystal display (LCD), speakers, vibration An output device 507 such as a computer; a storage device 506 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 509 . Communication means 509 may allow electronic device 500 to communicate wirelessly or by wire with other devices to exchange data. While FIG. 7 shows electronic device 500 having various means, it should be understood that not all of the illustrated means are required to be implemented or provided. More or fewer devices may alternatively be implemented or provided.

In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a non-transitory computer readable medium, the computer program containing program code for performing the method illustrated in the flowchart. In such an embodiment, the computer program may be downloaded and installed from the network via the communication device 509 , or from the storage device 506 , or from the ROM 502 . When the computer program is executed by the processing apparatus 501, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are executed.

The names of messages or information exchanged between multiple devices in the embodiments of the present disclosure are only for illustrative purposes, and are not intended to limit the scope of these messages or information.

The electronic device provided by the embodiment of the present disclosure and the data processing method provided by the above-mentioned embodiment belong to the same inventive concept. For the technical details not described in detail in this embodiment, please refer to the above-mentioned embodiment, and this embodiment has the same characteristics as the above-mentioned embodiment. beneficial effect.

Embodiment 8

Embodiments of the present disclosure provide a computer storage medium on which a computer program is stored, and when the program is executed by a processor, implements the data processing method provided by the foregoing embodiments.

It should be noted that the computer-readable medium mentioned above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In this disclosure, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present disclosure, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave with computer-readable program code embodied thereon. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device . Program code embodied on a computer readable medium may be transmitted using any suitable medium including, but not limited to, electrical wire, optical fiber cable, RF (radio frequency), etc., or any suitable combination of the foregoing.

In some embodiments, the client and server can communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and can communicate with digital data in any form or medium (eg, a communications network) interconnected. Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet (eg, the Internet), and peer-to-peer networks (eg, ad hoc peer-to-peer networks), as well as any currently known or future development network of.

The above-mentioned computer-readable medium may be included in the above-mentioned electronic device; or may exist alone without being assembled into the electronic device.

The above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by the electronic device, the electronic device:

Obtain the neural network parameters corresponding to the target neural grid model;

According to the neural network parameters, a shader to be used corresponding to each network level is generated; wherein, the target neural network model includes a plurality of network levels;

Determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so as to call the computing pipeline according to the target device parameters when receiving the data to be processed The data to be processed is processed to obtain a target processing result.

or,

When receiving the data to be processed, load the pre-determined computing pipeline and neural network parameters; wherein, the computing pipeline is based on the neural network parameters of the target neural network model by the central processing unit and the target neural network model belongs to The target device parameters are determined;

According to the target device parameters, determine the target processing mode of each shader to be used to process the data to be processed; wherein, the shader to be used is obtained by processing the neural network parameters based on the central processing unit;

The data to be processed is processed based on the target processing manner to obtain a target processing result.

Computer program code for performing operations of the present disclosure may be written in one or more programming languages, including but not limited to object-oriented programming languages—such as Java, Smalltalk, C++, and This includes conventional procedural programming languages - such as the "C" language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider through Internet connection).

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logical functions for implementing the specified functions executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations, can be implemented in dedicated hardware-based systems that perform the specified functions or operations , or can be implemented in a combination of dedicated hardware and computer instructions.

The units involved in the embodiments of the present disclosure may be implemented in a software manner, and may also be implemented in a hardware manner. Wherein, the name of the unit does not constitute a limitation of the unit itself under certain circumstances, for example, the first obtaining unit may also be described as "a unit that obtains at least two Internet Protocol addresses".

The functions described herein above may be performed, at least in part, by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on Chips (SOCs), Complex Programmable Logical Devices (CPLDs) and more.

In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in connection with the instruction execution system, apparatus or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media may include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media would include one or more wire-based electrical connections, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), fiber optics, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.

According to one or more embodiments of the present disclosure, [Example 1] provides a data processing method, which is applied to a central processing unit, and the method includes:

Obtain the neural network parameters corresponding to the target neural grid model;

According to the neural network parameters, a shader to be used corresponding to each network level is generated; wherein, the target neural network model includes a plurality of network levels;

Determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so as to call the computing pipeline according to the target device parameters when receiving the data to be processed The data to be processed is processed to obtain a target processing result.

According to one or more embodiments of the present disclosure, [Example 2] provides a data processing method, and the method further includes:

Optionally, generating a shader to be used corresponding to each network level according to the neural network parameters, including:

The shader generation module is called to process the neural network parameters to obtain the shader to be used for each network level in the target neural network model.

According to one or more embodiments of the present disclosure, [Example 3] provides a data processing method, and the method further includes:

Optionally, the target device parameters include indirect buffer parameters, and determining the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, including:

If the indirect buffer parameter of the device to which the target neural network model belongs is the first parameter, generating a scheduling shader corresponding to each shader to be used;

Obtaining a computing pipeline by compiling the scheduling shader and the shader to be used;

Wherein, the computing pipeline includes sub-computing pipelines corresponding to each shader to be used, and a control pipeline corresponding to each sub-computing pipeline.

According to one or more embodiments of the present disclosure, [Example 4] provides a data processing method, and the method further includes:

Optionally, the target device parameters include indirect buffer parameters, and the calculation pipeline corresponding to the shader to be used is determined according to the target device parameters of the device to which the target neural network model belongs, including:

If the indirect buffer parameter of the device to which the target neural network model belongs is the second parameter, determining the sub-computing pipeline corresponding to each shader to be used;

Based on each sub-computing pipeline, the computing pipeline is determined.

According to one or more embodiments of the present disclosure, [Example 5] provides a data processing method, and the method further includes:

Optionally, the neural network parameters and computing pipeline of the target neural network model are cached, so that when data to be processed is received, the neural network parameters and computing pipeline are retrieved to process the data to be processed.

According to one or more embodiments of the present disclosure, [Example 6] provides a data processing method, which is applied in a graphics processor, and the method includes:

When receiving the data to be processed, load the pre-determined computing pipeline and neural network parameters; wherein, the computing pipeline is based on the neural network parameters of the target neural network model by the central processing unit and the target neural network model belongs to The target device parameters are determined;

According to the target device parameters, determine the target processing mode of each shader to be used to process the data to be processed; wherein, the shader to be used is obtained by processing the neural network parameters based on the central processing unit;

The data to be processed is processed based on the target processing manner to obtain a target processing result.

According to one or more embodiments of the present disclosure, [Example 7] provides a data processing method, and the method further includes:

Optionally, according to the target device parameters, determining the target processing mode of each shader to be used to process the data to be processed includes:

If the target device parameter is the first parameter, determine that the target processing mode of each shader to be used to process the data to be processed is the first target processing mode;

If the target device parameter is the second parameter, it is determined that the target processing mode is the second target processing mode.

According to one or more embodiments of the present disclosure, [Example 8] provides a data processing method, and the method further includes:

Optionally, the target processing mode is a first target processing mode, and the data to be processed is processed based on the target processing mode to obtain a target processing result, including:

determining the execution order of each sub-computing pipeline based on the control pipeline corresponding to the computing pipeline;

Based on the execution order and corresponding sub-computing pipelines, program execution instructions are sent to the corresponding shaders to be used, so that each shader to be used processes the data to be processed to obtain a processing result corresponding to the target.

According to one or more embodiments of the present disclosure, [Example 9] provides a data processing method, the method further comprising:

Optionally, the processing of the data to be processed based on the target processing mode to obtain a target processing result, including:

Receive the program execution instruction sent based on the current sub-computing pipeline; wherein, the program execution instruction is determined based on the central processing unit according to the sub-computing pipelines of each neural network level;

Based on the program execution instructions, the data to be processed is processed to obtain a data processing result, and the data processing result is fed back to the central processing unit, so that when the central processing unit receives the data processing result, Taking the next sub-computing pipeline of the current sub-computing pipeline as the current sub-computing pipeline, and sending a program execution instruction based on the current sub-computing pipeline, until the current sub-computing pipeline is the last sub-computing pipeline, obtaining and The target processing result corresponding to the data to be processed.

According to one or more embodiments of the present disclosure, [Example 10] provides a data processing apparatus configured in a central processing unit, the apparatus comprising:

The network parameter determination module is used to obtain the neural network parameters corresponding to the target neural grid model;

a shader determination module, configured to generate shaders to be used corresponding to each network level according to the neural network parameters; wherein, the target neural network model includes a plurality of network levels;

A pipeline determination module, configured to determine the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs, so that when the data to be processed is received, according to the target device parameters, The computing pipeline is called to process the to-be-processed data to obtain a target processing result.

According to one or more embodiments of the present disclosure, [Example 11] provides a data processing apparatus configured in a graphics processor, the apparatus comprising:

The network parameter loading module is used to load the pre-determined calculation pipeline and neural network parameters when receiving the data to be processed; wherein, the calculation pipeline is based on the neural network parameters and the neural network parameters of the target neural network model based on the central processing unit. Determined by the parameters of the target device to which the target neural network model belongs;

A processing mode determination module, configured to determine, according to the target device parameters, a target processing mode for each shader to be used to process the data to be processed; wherein, the shader to be used is based on the processing of the neural network parameters by the central processing unit obtained after

A processing result determination module, configured to process the data to be processed based on the target processing manner to obtain a target processing result.

The above description is merely a preferred embodiment of the present disclosure and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the disclosure involved in the present disclosure is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, and should also cover, without departing from the above-mentioned disclosed concept, the technical solutions formed by the above-mentioned technical features or Other technical solutions formed by any combination of its equivalent features. For example, a technical solution is formed by replacing the above features with the technical features disclosed in the present disclosure (but not limited to) with similar functions.

Additionally, although operations are depicted in a particular order, this should not be construed as requiring that the operations be performed in the particular order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Likewise, although the above discussion contains several implementation-specific details, these should not be construed as limitations on the scope of the present disclosure. Certain features that are described in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination.

Although the subject matter has been described in language specific to structural features and/or logical acts of method, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.

Claims (13)

1. A data processing method is applied to a central processing unit and comprises the following steps:

acquiring neural network parameters corresponding to the target neural grid model;

generating shaders to be used corresponding to each network level according to the neural network parameters; wherein the target neural network model comprises a plurality of network hierarchies;

and determining a computing pipeline corresponding to the shader to be used according to target equipment parameters of the equipment to which the target neural network model belongs, and calling the computing pipeline to process the data to be processed according to the target equipment parameters when the data to be processed is received to obtain a target processing result.

2. The method of claim 1, wherein generating a to-be-used shader corresponding to each network level according to the neural network parameters comprises:

and calling a shader generating module to process the neural network parameters to obtain a shader to be used of each network level in the target neural network model.

3. The method of claim 1, wherein the target device parameter comprises an indirect buffering parameter, and wherein determining the computing pipeline corresponding to the shader to be used according to the target device parameter of the device to which the target neural network model belongs comprises:

if the indirect buffering parameter of the device to which the target neural network model belongs is a first parameter, generating scheduling shaders corresponding to the shaders to be used;

compiling and processing the scheduling shader and the to-be-used shader to obtain a computing pipeline;

the computing pipeline comprises sub-computing pipelines corresponding to all shaders to be used and control pipelines corresponding to all the sub-computing pipelines.

4. The method of claim 1, wherein the target device parameters comprise indirect buffer parameters, and wherein determining the computing pipeline corresponding to the shader to be used according to the target device parameters of the device to which the target neural network model belongs comprises:

if the indirect buffering parameter of the device to which the target neural network model belongs is a second parameter, determining a sub-computing pipeline corresponding to each shader to be used;

based on each sub-computation pipeline, the computation pipeline is determined.

5. The method of claim 1, further comprising:

caching the neural network parameters and the calculation pipelines of the target neural network model so as to call the neural network parameters and the calculation pipelines to process the data to be processed when the data to be processed is received.

6. A data processing method applied to a graphics processor includes:

loading predetermined calculation pipeline and neural network parameters when receiving data to be processed; wherein the calculation pipeline is determined based on the neural network parameters of a target neural network model and the target equipment parameters of the target neural network model by a central processing unit;

determining a target processing mode for processing the data to be processed by each shader to be used according to the target equipment parameters; the shader to be used is obtained after the neural network parameters are processed based on a central processing unit;

and processing the data to be processed based on the target processing mode to obtain a target processing result.

7. The method of claim 6, wherein determining the target processing mode for processing the data to be processed by each shader according to the target device parameters comprises:

if the target equipment parameter is a first parameter, determining that a target processing mode for processing the data to be processed by each shader to be used is a first target processing mode;

and if the target equipment parameter is the second parameter, determining that the target processing mode is the second target processing mode.

8. The method according to claim 7, wherein the target processing manner is a first target processing manner, and the processing the data to be processed based on the target processing manner to obtain a target processing result includes:

determining an execution order of each sub-compute pipeline based on a control pipeline corresponding to the compute pipeline;

and sending a program execution instruction to the corresponding to-be-used shaders based on the execution sequence and the corresponding sub-computing pipelines, so that each to-be-used shader processes the to-be-processed data to obtain the target processing result.

9. The method according to claim 8, wherein the processing the data to be processed based on the target processing manner to obtain a target processing result comprises:

receiving a program execution instruction sent based on a current sub-compute pipeline; wherein the program execution instructions are determined based on the central processor from sub-compute pipelines of each neural network hierarchy;

and processing the data to be processed based on the program execution instruction to obtain a data processing result, and feeding back the data processing result to the central processing unit, so that when the central processing unit receives the data processing result, the next sub-computing pipeline of the current sub-computing pipeline is used as the current sub-computing pipeline, and the program execution instruction is sent based on the current sub-computing pipeline until the current sub-computing pipeline is the last sub-computing pipeline, so as to obtain a target processing result corresponding to the data to be processed.

10. A data processing apparatus, configured in a central processing unit, comprising:

the network parameter determining module is used for acquiring neural network parameters corresponding to the target neural grid model;

the shader determining module is used for generating shaders to be used corresponding to each network level according to the neural network parameters; wherein the target neural network model comprises a plurality of network hierarchies;

and the pipeline determining module is used for determining a computing pipeline corresponding to the shader to be used according to the target equipment parameter of the equipment to which the target neural network model belongs, and calling the computing pipeline to process the data to be processed according to the target equipment parameter when the data to be processed is received to obtain a target processing result.

11. A data processing apparatus, configured in a graphics processor, comprising:

the network parameter loading module is used for loading predetermined calculation pipelines and neural network parameters when receiving data to be processed; wherein the calculation pipeline is determined based on the neural network parameters of a target neural network model and the target equipment parameters of the target neural network model by a central processing unit;

the processing mode determining module is used for determining a target processing mode for processing the data to be processed by each shader to be used according to the target equipment parameters; the shader to be used is obtained after the neural network parameters are processed based on a central processing unit;

and the processing result determining module is used for processing the data to be processed based on the target processing mode to obtain a target processing result.

12. An electronic device, characterized in that the electronic device comprises:

one or more processors;

a storage device for storing one or more programs,

when executed by the one or more processors, cause the one or more processors to implement a data processing method as claimed in any one of claims 1-5 or 6-9.

13. A storage medium containing computer-executable instructions for performing the data processing method of any one of claims 1-5 or 6-9 when executed by a computer processor.

Images