CN114021708B - Data processing method, device and system, electronic equipment and storage medium - Google Patents

Abstract

The application discloses a data processing method, device, system, electronic equipment, and computer-readable storage medium. The method includes: obtaining a setting instruction, and setting a computing network according to the setting instruction; the setting instruction is used to set each computing core in the computing network The data flow direction between them; obtain at least one eigenvalue, and input at least one eigenvalue into at least one initial computing core in the computing network; start from the initial computing core, and transmit the eigenvalue according to the data flow; use each computing core , to generate calculation results based on eigenvalues and corresponding weight values; this method sets the data flow direction between each calculation core in the calculation network by setting instructions, so that data flows between different levels, or between the same levels, so that the entire Computational networks can be 100% fully utilized regardless of the shape of the network model being processed.

Description

A data processing method, device, system, electronic equipment and storage medium

technical field

The present application relates to the technical field of deep learning, and in particular to a data processing method, a data processing device, a data processing system, electronic equipment, and a computer-readable storage medium.

Background technique

At present, the development of deep learning puts forward a very high demand for computing power. Various ASIC (Application Specific Integrated Circuit, application specific integrated circuit) architectures emerge in an endless stream. The more representative ones are: Google's tensor processing unit (TPU, Tensor Processing Unit), which is a custom ASIC chip designed for machine learning workloads. But the TPU is not a general-purpose processor, but a matrix processor dedicated to neural network workloads. Its main task is matrix processing, and the hardware designer of the TPU knows every step of the operation process. So they put tens of thousands of multipliers and adders and connect them directly to build a physical matrix of those operators, called a Systolic Array architecture. A systolic array is a series of processing elements (Processing Elements, PE) regularly arranged in a grid. Each PE in the systolic array will perform data transmission with its adjacent PEs in predetermined steps. In addition, other architecture schemes similar to TPU (such as Da Vinci architecture scheme) also adopt a regular two-dimensional or three-dimensional grid PE array structure. However, the shape of the neural network or the number of specification parameters vary greatly, while the hardware scale of TPU and the like is fixed. Only when the shape of the network fully matches the chip can the utilization of hardware resources be maximized. The neural network, its hardware resource utilization is related to the size of the network, but it cannot fully utilize the hardware resources, and there is a problem of insufficient resource utilization.

Contents of the invention

In view of this, the object of the present application is to provide a data processing method, device, system, electronic equipment, and computer-readable storage medium, which improves resource utilization.

In order to solve the above technical problems, the application provides a data processing method, including:

Obtaining a setting instruction, and setting a computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network;

obtaining at least one eigenvalue, and inputting the at least one eigenvalue into at least one initial computing core in the computing network;

Using the initial calculation core as a starting point, transmit the feature value according to the data flow direction;

using each of the calculation cores to generate calculation results based on the feature values and corresponding weight values;

Wherein, the computing network includes a first computing core, several second computing cores, and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and The third computing core, or includes the first computing core, the second computing core and the third computing core; any of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores , the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is a higher-level computing core of the target second computing core.

Optionally, also include:

Obtain configuration information, and store the configuration information in each of the computing cores; wherein, the configuration information includes several identification information and corresponding several data flows;

Correspondingly, setting the computing network according to the setting instruction includes:

Sending the setting instruction to each of the computing cores, using the target identification information and the configuration information in the setting instruction to determine the data flow direction corresponding to each of the computing cores.

Optionally, the setting instruction is used to set the computing network as one-way computing, then the data flow direction is the first flow direction from the upper level to the lower level, and the initial computing core is the first computing core;

Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes:

inputting the feature value into the first computing core, and using the first computing core to send the feature value to the target second computing core;

Based on the first flow direction, starting from the target second computing core, the feature values are sent to corresponding lower-level computing cores in sequence until the feature values are sent to the third computing core.

Optionally, the setting instruction is used to set the computing network to be at least three-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, the starting The initial calculation core includes the first calculation core and the initial third calculation core, the first calculation core corresponds to the first flow direction, and the initial third calculation core corresponds to the second flow direction;

Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes:

inputting a first feature value into the first computing core, and using the first computing core to send the first feature value to the target second computing core;

Based on the first flow direction, starting from the target second computing core, sending the first eigenvalues to the corresponding lower-level computing cores sequentially until sending the first eigenvalues to the end of the first flow direction the second computing core;

inputting the second eigenvalue into the starting third computing core, and starting from the starting third computing core, based on the second flow direction, sequentially sending the second eigenvalue to subsequent computing cores of the same level , until the second feature value is sent to the third computing core at the end of the second stream.

Optionally, the initial computing core further includes an initial second computing core, and the second computing core corresponds to the second flow direction;

Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes:

inputting the second eigenvalue into the starting second computing core, and starting from the starting second computing core, based on the second flow direction, sequentially sending the second eigenvalue to subsequent computing cores of the same level , until the second feature value is sent to the second computing core at the end of the second stream.

Optionally, also include:

Determining the weight values corresponding to each of the calculation cores;

Sending and storing the weight value to the corresponding computing core.

Optionally, the determining the weight values corresponding to each of the calculation cores includes:

Get the initial weight value;

Based on the data flow direction, determine the corresponding relationship between each of the calculation cores and the initial weight value, and complete the determination of the weight value.

Optionally, using each of the calculation cores to generate calculation results based on the feature values and corresponding weight values includes:

controlling each of the calculation cores, and multiplying the feature value and the weight value to obtain a target result;

The target result is added to the historical calculation result stored by the calculation core to obtain the calculation result.

Optionally, the calculation core includes an arithmetic logic unit, a weight value interface, a feature value interface, a control module and a storage unit, the weight value interface includes an external input port and an input port of the same level, and the feature value interface includes an external input port Port, upper-level input port and same-level input port, the control module is used to store configuration information, and determine the data flow direction according to the setting instruction, and the storage unit is used to store weight values, feature values and the calculation results.

The present application also provides a data processing device, including:

A setting module, configured to obtain a setting instruction, and set a computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network;

An input module, configured to acquire at least one eigenvalue, and respectively input the at least one eigenvalue into at least one initial calculation core in the calculation network;

a transmission module, configured to use the initial calculation core as a starting point, and transmit the characteristic value according to the data flow direction;

A calculation module, configured to use each of the calculation cores to generate calculation results based on the feature values and corresponding weight values;

Wherein, the computing network includes a first computing core, several second computing cores, and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and The third computing core, or includes the first computing core, the second computing core and the third computing core; any of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores , the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is a higher-level computing core of the target second computing core.

The present application also provides a data processing system, including a computing network, and the computing network includes a first computing core, several second computing cores, and several third computing cores; any of the second computing cores corresponds to One upper-level computing core and two lower-level computing cores, the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is The upper-level computing core of the target second computing core.

The present application also provides an electronic device, including a memory and a processor, wherein:

The memory is used to store computer programs;

The processor is configured to execute the computer program to implement the above data processing method.

The present application also provides a computer-readable storage medium for storing a computer program, wherein the above-mentioned data processing method is implemented when the computer program is executed by a processor.

The data processing method provided by the present application obtains setting instructions, and sets the computing network according to the setting instructions; the setting instructions are used to set the data flow direction between computing cores in the computing network; obtain at least one eigenvalue, and separate the at least one eigenvalue Input at least one initial computing core in the computing network; start from the initial computing core, transmit feature values according to the data flow direction; use each computing core to generate calculation results based on feature values and corresponding weight values; wherein, the computing network includes a The first computing core, several second computing cores and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and the third computing core, or includes the first computing core, the second computing core Computing core and third computing core; any second computing core corresponds to one upper-level computing core and two lower-level computing cores, the lower-level computing core is the second computing core or the third computing core, and the third computing core does not have a lower-level computing core , the first computing core is the superior computing core of the target second computing core.

It can be seen that the method adopts a computing network with a special architecture, and there are three types of computing cores in the computing network, and the computing cores have a preset hierarchical relationship. Among them, the first computing core is a single level, one target second computing core is the lower-level computing core of the first computing core, two second computing cores are the lower-level computing cores of the target second computing core, and four other second computing cores are The computing cores are the lower-level computing cores of the above two second computing cores, and so on, until all the third computing cores in pairs are the lower-level computing cores of the second computing cores of a certain level. All computing cores are arranged according to the level, and the computing cores of the same level are on the same plane, presenting a pyramid shape. In addition, the number of computing cores in the computing network is 1+1+2+4+...+2^n=2^(n+1), and the number of computing cores at each level is all Compute the sum of the number of cores. Since the number of channels of the neural network is usually 2^k, if it is not necessary to use all the computing cores to participate in the calculation, that is, when the shape of the network model does not match the size of the computing network, the first few stages can be used The calculation core of the calculation path is used as a calculation path to obtain 2^k calculation results, and the lower-level calculation core is split to obtain at least two other calculation paths to process other eigenvalues, and each path is independent of each other. ,Parallel Computing. Specifically, by setting instructions to set the data flow direction between each computing core in the computing network, so that the data flows between different levels, or between the same level, so that no matter what shape of the network model the entire computing network is processing, All of them can be fully utilized, which solves the problem of insufficient hardware utilization in related technologies.

In addition, the present application also provides a data processing device, an electronic device, and a computer-readable storage medium, which also have the above beneficial effects.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or related technologies. Obviously, the accompanying drawings in the following description are only For the embodiments of the application, those skilled in the art can also obtain other drawings according to the provided drawings without creative work.

FIG. 1 is a schematic structural diagram of a computing network provided by an embodiment of the present application;

FIG. 2 is a flow chart of a data processing method provided by an embodiment of the present application;

FIG. 3 is a schematic diagram of a level structure between a second computing core and a third computing core provided in an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a specific computing core provided by an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a data processing device provided in an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments It is only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

This application proposes a computing network with a new structure, please refer to FIG. 1 , which is a schematic structural diagram of a computing network provided by an embodiment of this application. The computing network includes a first computing core, several second computing cores and several third computing cores. Wherein, any one of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores, and the lower-level computing cores are specifically the second computing cores, or may be the third computing cores. The third computing core does not have a lower-level computing core, that is, it is the lowest-level computing core of the entire network, the first computing core is the upper-level computing core of the target second computing core, and the target second computing core is the middle level of the left and right second computing cores the tallest one.

As can be seen from Figure 1, the entire computing network is divided from the perspective of each computing core level, and the pyramid structure shown in Figure 1 can be obtained. The uppermost computing core is the first computing core (which can be regarded as being at level 0), and its corresponding lower-level computing core is the target second computing core (which can be regarded as being at level 1), and the number of target second computing cores for one. The target second computing core is also the second computing core, which also has the characteristics of the second computing core, that is, it has one upper-level computing core (level 0) and two lower-level computing cores (which can be regarded as being in the second level). The lower-level computing core (level 2) of the target second computing core also has two lower-level computing cores (level 3), and so on, until the third computing core serves as the lower-level computing core of the second computing core, and the third computing core A core does not have subordinate computing cores. In addition, the connected nodes in Figure 1 are computing cores, and the nodes connected by dotted lines are transmission nodes, which are used to transmit weight values.

It can be seen that starting from the first calculation core, if the third calculation core is at the nth level, there are 1+1+2+4+...+2^n=2^(n+1) calculations in the entire computing network Kernel, n is a positive integer. Moreover, the number of computing cores at each level is the sum of the numbers of all computing cores at levels higher than that level. Exemplarily, the number of computing cores at level 5 is the sum of the numbers of all computing cores corresponding to level 0, level 1, level 2, level 3, and level 4. Since the number of channels of the neural network is usually 2^k (k is a positive integer), when performing calculations, if k=n, all calculation cores of the entire calculation network can participate in the calculation. If you don’t need to use all the computing cores to participate in the calculation, that is, when the shape of the network model does not match the size of the computing network, when k<n, you can use the computing cores of the first k-1 levels as a computing path to get 2^k Calculate the result, and split the calculation cores of the kth level and lower levels to obtain at least two other calculation paths, and process other eigenvalues. Each path is independent of each other and calculated in parallel.

Exemplarily, if n=6, k=6, then the number of computing cores of the entire computing network is 2^(6+1)=2^7, and the number of channels of the network model is 2^6, then the computing The calculation cores before level 6 in the network are calculated as one way. For the calculation cores of level 7 (actually the third calculation core), the number is twice k, so it can be split into two independent channels. Computing path, three-way computing path for parallel computing.

On the basis of the above computing network, this embodiment also provides a processing method for data processing using it. Please refer to FIG. 2 , which is a flow chart of a data processing method provided by an embodiment of the present application. The method includes:

S101: Obtain a setting instruction, and set a computing network according to the setting instruction.

Wherein, the setting instruction is used to set the data flow direction between the various computing cores in the computing network, and the data flow direction refers to the transmission direction of the characteristic values between the computing cores. It can be understood that after the structure of the computing network is determined and the number of channels of the network model is determined, that is, after the sizes of n and k are determined, it can be determined whether the computing network and the network model are fully compatible. Therefore, before using the computing network for data processing, it needs to be specifically set up. By setting the data flow direction between computing cores, the computing network can be set as one or more parallel computing paths, so as to fully utilize the resource utilization of the computing network.

S102: Obtain at least one eigenvalue, and respectively input at least one eigenvalue to at least one initial calculation core in the calculation network.

The initial calculation kernel refers to the first calculation kernel of the eigenvalue input. It can be understood that each calculation path is calculated in parallel, and each path should have one and only one initial calculation core, so the number of initial calculation cores and calculation paths are the same. After obtaining the eigenvalues, input them into the initial calculation core, and the initial calculation core can use the eigenvalues for data processing, and can also transfer the eigenvalues according to the set data flow direction.

S103: Starting from the initial calculation core, transmit the feature value according to the data flow.

The data flow direction specifies the transmission direction of the eigenvalues, so the initial calculation core can determine the sender of the eigenvalues, and then send the eigenvalues to the next calculation core until all the calculation cores receive the eigenvalues.

S104: Using each calculation core, generate a calculation result based on the feature value and the corresponding weight value.

After obtaining the eigenvalues, each calculation core can use the eigenvalues and its corresponding weight values to perform calculations to obtain corresponding calculation results. The collective calculation process is not limited and can be set as required. The weight values are pre-stored in the computing cores, and their number and size are not limited.

It should be noted that the execution order of steps S102, S103 and S104 in this application is not limited. In one embodiment, the three steps are executed in series, that is, step S102 is executed first, then step S103 is executed, and finally S104 is executed. step. In another implementation manner, the three steps may be executed in parallel, that is, while step S102 is executed, steps S103 and/or S104 may be executed.

The application of the data processing method provided by the embodiment of the present application adopts a computing network with a special architecture. There are three types of computing cores in the computing network, and there is a preset hierarchical relationship among the computing cores. Among them, the first computing core is a single level, one target second computing core is the lower-level computing core of the first computing core, two second computing cores are the lower-level computing cores of the target second computing core, and four other second computing cores are The computing cores are the lower-level computing cores of the above two second computing cores, and so on, until all the third computing cores in pairs are the lower-level computing cores of the second computing cores of a certain level. All computing cores are arranged according to the level, and the computing cores of the same level are on the same plane, presenting a pyramid shape. In addition, the number of computing cores in the computing network is 1+1+2+4+...+2^n=2^(n+1), and the number of computing cores at each level is all Compute the sum of the number of cores. Since the number of channels of the neural network is usually 2^k, if it is not necessary to use all the computing cores to participate in the calculation, that is, when the shape of the network model does not match the size of the computing network, the first few stages can be used The calculation core of the calculation path is used as a calculation path to obtain 2^k calculation results, and the lower-level calculation core is split to obtain at least two other calculation paths to process other eigenvalues, and each path is independent of each other. ,Parallel Computing. Specifically, by setting instructions to set the data flow direction between each computing core in the computing network, so that the data flows between different levels, or between the same level, so that no matter what shape of the network model the entire computing network is processing, All of them can be fully utilized, which solves the problem of insufficient hardware utilization in related technologies.

Based on the foregoing embodiments, this embodiment further describes the steps of the data processing method. As for the specific content of the setting instruction, in an implementation manner, it may specify the data flow among all computing cores in the computing network. In another implementation manner, before setting the data flow, the following steps may also be included:

Step 11: Obtain configuration information, and store the configuration information to each computing core.

Wherein, the configuration information includes several identification information and corresponding several data flow directions.

Correspondingly, the process of setting the computing network according to the setting instruction may include:

Step 12: Send the setting command to each computing core, and use the target identification information and configuration information in the setting command to determine the data flow direction corresponding to each computing core.

Several data flow directions and unique identification information corresponding to these data flow directions are set in the configuration information. The configuration information is preset in each computing core, and after the setting instruction is obtained, the configuration information can be screened by using the target identification information in the setting instruction to obtain the corresponding data flow direction. Specifically, after determining the data flow direction, each calculation core can use its corresponding identity information to screen the data flow direction to obtain its own corresponding data flow direction, that is, the sending direction of the characteristic value.

There are many possibilities for the specific content of the data flow direction, wherein, in one embodiment, the setting instruction is used to set the computing network to one-way computing, then the data flow direction is the first flow direction from the upper level to the lower level, and the initial computing core is the first a computing core;

Correspondingly, at least one eigenvalue is respectively input into at least one initial calculation core in the calculation network; starting from the initial calculation core, the process of transmitting the eigenvalue according to the data flow may specifically include:

Step 21: Input the feature value into the first computing core, and use the first computing core to send the feature value to the target second computing core.

Step 22: Based on the first flow direction, starting from the target second computing core, sequentially sending the feature values to the corresponding lower-level computing cores until sending the feature values to the third computing core.

In this embodiment, the entire computing network is calculated as one computing path. In this case, there is only one initial computing core, that is, the first computing core, and the data flow direction is the first flow direction from the upper level to the lower level. Therefore, after inputting the feature value into the first computing core, the first computing core can be controlled to send the feature value to its corresponding lower-level computing core, that is, the target second computing core. Starting from the target second computing core, each computing core sends the obtained eigenvalues to the corresponding lower-level computing cores, until the eigenvalues are sent to all the third computing cores. The third calculation core does not have a lower-level calculation core, and the transfer of eigenvalues ends.

In the second embodiment, the setting instruction is used to set the computing network to at least three-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, and the initial computing core includes The first computing core and the starting third computing core, the first computing core corresponds to the first flow direction, and the starting third computing core corresponds to the second flow direction;

Correspondingly, at least one eigenvalue is respectively input into at least one initial computing core in the computing network; starting from the initial computing core, the process of transmitting the eigenvalue according to the data flow may include:

Step 31: Input the first feature value into the first computing core, and send the first feature value to the target second computing core by using the first computing core.

Step 32: Based on the first flow direction, starting from the target second computing core, sequentially send the first feature value to the corresponding lower-level computing core until sending the first feature value to the second computing core at the end of the first flow direction.

Step 33: Input the second eigenvalue into the starting third computing core, and from the starting third computing core, based on the second flow direction, sequentially send the second eigenvalue to subsequent computing cores of the same level until the second The eigenvalues are sent to the third computing core at the end of the second stream.

In this embodiment, the entire computing network can be divided into three computing paths, or more than three computing paths. In this case, in addition to the first flow direction from the upper level to the lower level, the data flow also includes the second flow direction between the same levels, that is, the feature values flow between the calculation cores of the same level. It can be understood that since there are multiple calculation paths, there are also multiple initial calculation cores, which include at least two initial third calculation cores while including the first calculation core.

The starting third computing core refers to the third computing core designated as the starting computing core, and the number of starting third computing cores can be changed according to the number of computing paths. However, it is understandable that when there are three computing paths, all the third computing cores are divided into two independent computing paths, so the initial number of third computing cores is at least two, and when there are more computing paths, The number of initial third computing cores is greater.

In this case, there is no difference between this embodiment and the above-mentioned first embodiment for the computation path where the first computing core is used as the initial computing core. The eigenvalue of the input first calculation core is the first eigenvalue, and after the first calculation core sends the first eigenvalue to the target second calculation core, the first eigenvalue is issued step by step according to the first flow direction until the The first feature to send to the second compute core as the end of the first flow. For the calculation path where the third calculation core is used as the initial calculation core, the eigenvalue of the input starting third calculation core is the second eigenvalue, and the second flow direction also specifies the order of data transmission between data cores, so according to In the second flow direction, subsequent computing cores of the same level corresponding to each third computing core starting from the initial third computing core can be determined, and each third computing core sends the second characteristic value to subsequent computing cores of the same level until the second The eigenvalues are sent to the third computing core at the end of the second stream.

On the basis of the second embodiment, there is also a third embodiment. Specifically, the initial calculation core also includes an initial second calculation core, and the second calculation core corresponds to the second flow direction;

Correspondingly, at least one eigenvalue is respectively input into at least one initial computing core in the computing network; starting from the initial computing core, the process of transmitting the eigenvalue according to the data flow may include:

Step 41: Input the second eigenvalue into the starting second computing core, and from the starting second computing core, based on the second flow direction, sequentially send the second eigenvalue to the subsequent computing core at the same level until the second The eigenvalues are sent to the second computational core at the end of the second stream.

In this embodiment, the initial computing core also includes the initial second computing core, that is, several levels higher than the third computing core except that the lowest-level third computing core is split into at least two computing paths. The second computing core is also split into at least two computing paths. Please refer to FIG. 3 , which is a schematic diagram of a level structure between a second computing core and a third computing core provided in an embodiment of the present application. Wherein, the calculation core of level 6 is the third calculation core, the calculation cores of level 1 to level 5 are the second calculation core, and there is also the first calculation core of level 0 (not shown in the figure). Exemplarily, the calculation path from the first calculation core of level 0 to the calculation core of level 4 can be set as one calculation path, that is, the number of channels of the network is 16. Based on this again, the second computing core at the fifth level can be divided into one computing path, and the third computing core at the fifth level can be divided into two computing paths.

When the second calculation core is used as the initial calculation core, its corresponding eigenvalue flow mode is the same as that of the third calculation core as the initial calculation core. For the calculation path where the second calculation core is used as the initial calculation core, the eigenvalue of the input starting second calculation core is the second eigenvalue, and the second flow direction also specifies the order of data transmission between data cores, so according to In the second flow direction, subsequent computing cores of the same level corresponding to each second computing core starting from the starting second computing core can be determined, and each second computing core sends the second feature value to subsequent computing cores of the same level until the second The eigenvalues are sent to the second computational core at the end of the second stream.

After obtaining the eigenvalues, each calculation core should use them and weight values to perform calculations to obtain calculation results. It can be understood that, before using the weight value to generate the calculation result, it is necessary to assign a corresponding weight value to each calculation core. Specifically, it may also include:

Step 51: Determine the weight values corresponding to each calculation core.

Step 52: Send and store the weight value to the corresponding computing core.

Each calculation core needs to store its own corresponding weight value, which is used to generate the output value of the corresponding channel together with the feature value, that is, the calculation result. According to the matching degree of the calculation network and the number of model output channels, the data flow direction can be determined, and then the weight value corresponding to each calculation core can be determined. Each weight value is sent to the calculation core and stored, so that the calculation core can directly call the weight value for calculation after obtaining the feature value.

Further, the process of determining the weight values corresponding to each calculation core may include:

Step 61: Get the initial weight value.

Step 62: Based on the data flow direction, determine the corresponding relationship between each calculation core and the initial weight value, and complete the determination of the weight value.

Wherein, the initial weight value refers to the weight value not allocated to the calculation core, and each initial weight value corresponds to different output channels of the model, which can be arranged according to the data flow direction. Therefore, according to the data flow direction corresponding to the computing network, including the first flow direction and the second flow direction, determine the corresponding relationship between the computing core and the initial weight value, and determine the initial weight value corresponding to the computing core as the weight value of the computing core, Complete the determination of the weight value.

This embodiment does not limit the specific method of calculation kernel calculation, and current research on deep learning mainly takes CNN (Convolutional Neural Networks, Convolutional Neural Networks) as the research object. Due to different processing scenarios, the performance requirements of CNN are also different, thus developing a variety of network structures. However, the basic composition of CNN is fixed, which are input layer, convolutional layer, activation layer, pooling layer and fully connected layer. The most computationally intensive part is the convolution layer, whose main function is to complete the convolution operation between the image (feature) and the neuron (filter). Therefore, in an implementation manner, a computing network may be used to perform convolution calculations. After the weight values are assigned, using each calculation core, the process of generating calculation results based on the feature values and corresponding weight values may include:

Step 71: Control each calculation core, multiply the feature value and the weight value to obtain the target result.

Step 72: Add the target result and the historical calculation result stored in the calculation core to obtain the calculation result.

It can be understood that convolution calculations are multiplication and addition calculations. Specifically, the target result obtained by multiplying the feature value and the weight value is added to the historical calculation result stored in the calculation core to obtain the final required calculation result. Correspondingly, the new calculation results should also be stored in the calculation core, so as to be used as the historical calculation results of subsequent calculations, or read out after the feature values of all input channels are facilitated.

The specific structure of the computing core is not limited in this embodiment. In an implementation manner, please refer to FIG. 4 , which is a schematic structural diagram of a specific computing core provided in an embodiment of the present application. The calculation core includes an arithmetic logic unit (Arithmetic&logical Unit, ALU), a weight value interface, a feature value interface, a control module and a storage unit. Among them, the weight value interface includes an external input port and an input port of the same level, the external input interface is used to obtain the weight value of the external input, and the input port of the same level is used to obtain the weight value input by the transmission node of the same level. The eigenvalue interface includes an external input port, an upper-level input port, and an input port of the same level. The external input port is used to obtain the eigenvalue of the external input, and the calculation core is used when the calculation core is used as the initial calculation core. ) is used to obtain the characteristic value sent by the computing core of the upper level, and the input port of the same level is used to obtain the characteristic value sent by the computing core of the same level. The control module is used to store configuration information and determine the data flow direction according to setting instructions, and the storage unit is used to store weight values, feature values and calculation results. In addition, it also includes a selector for selecting the information input by each input port as the feature value and weight value used for calculation, and the selection method of the selector is related to the data flow direction.

The data processing device provided in the embodiment of the present application is introduced below, and the data processing device described below and the data processing method described above may refer to each other correspondingly.

Please refer to FIG. 5. FIG. 5 is a schematic structural diagram of a data processing device provided in an embodiment of the present application, including:

The setting module 110 is used to obtain the setting instruction, and set the computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network;

The input module 120 is used to obtain at least one feature value, and input at least one feature value into at least one initial calculation core in the calculation network;

The transmission module 130 is used to transmit the feature value according to the data flow direction starting from the initial computing core;

Calculation module 140, is used for using each calculation kernel, based on feature value and corresponding weight value generation calculation result;

Wherein, the computing network includes a first computing core, several second computing cores, and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and the third computing core, or includes the first computing core A computing core, a second computing core, and a third computing core; any second computing core corresponds to an upper-level computing core and two lower-level computing cores, and the lower-level computing core is the second computing core or the third computing core, and the third computing core The core does not have a lower-level computing core, and the first computing core is the upper-level computing core of the target second computing core.

Optionally, also include:

The configuration storage module is used to obtain configuration information and store the configuration information in each computing core; wherein, the configuration information includes several identification information and corresponding several data flow directions;

Correspondingly, the setting module 110 includes:

The configuration instruction unit is configured to send a setting instruction to each computing core, and use the target identification information and configuration information in the setting instruction to determine the data flow direction corresponding to each computing core.

Optionally, the setting instruction is used to set the computing network to one-way computing, then the data flow direction is the first flow direction from the upper level to the lower level, and the initial computing core is the first computing core;

Correspondingly, the input module 120 includes:

The first input unit is used to input the feature value into the first calculation core;

Correspondingly, the transmission module 130 includes:

A first transmission unit, configured to use the first computing core to send the feature value to the target second computing core;

The second transmission unit is configured to, based on the first flow direction, start from the target second computing core, and sequentially send the feature values to corresponding lower-level computing cores until sending the feature values to the third computing core.

Optionally, the setting instruction is used to set the computing network to be at least three-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, and the initial computing core includes the first computing core and starting the third computing core, the first computing core corresponds to the first flow direction, and the starting third computing core corresponds to the second flow direction;

Correspondingly, the input module 120 includes:

a second input unit, configured to input the first feature value into the first calculation core;

Correspondingly, the transmission module 130 includes:

a third transmission unit, configured to use the first computing core to send the first feature value to the target second computing core;

The fourth transmission unit is configured to, based on the first flow direction, start from the target second computing core, and sequentially send the first eigenvalues to the corresponding lower-level computing cores, until sending the first eigenvalues to the second computing core at the end of the first flow direction nuclear;

The fifth transmission unit is configured to input the second eigenvalue into the starting third computing core, and from the starting third computing core, based on the second flow direction, sequentially send the second eigenvalue to subsequent computing cores of the same level, Until the second feature value is sent to the third computing core at the end of the second stream.

Optionally, the initial calculation core further includes an initial second calculation core, and the second calculation core corresponds to the second flow direction;

Correspondingly, the input module 120 includes:

The third input unit is used to input the second feature value into the starting second computing core;

Correspondingly, the transmission module 130 includes:

The sixth transmission unit is configured to start from the starting second computing core, and based on the second flow direction, sequentially send the second eigenvalues to subsequent computing cores at the same level until sending the second eigenvalues to the second eigenvalue at the end of the second flow direction Two computing cores.

Optionally, also include:

a weight determination module, configured to determine the weight values corresponding to each calculation core;

The weight storage module is used to send and store the weight value to the corresponding computing core.

Optionally, the weight determination module includes:

a weight acquisition unit, configured to acquire an initial weight value;

The corresponding relationship determination unit is configured to determine the corresponding relationship between each calculation core and the initial weight value based on the data flow direction, and complete the determination of the weight value.

Optionally, the computing module 140 includes:

The multiplication unit is used to control each calculation core, and obtain the target result by multiplying the feature value and the weight value;

The addition unit is used for adding the target result and the historical calculation result stored in the calculation core to obtain the calculation result.

Optionally, the calculation core includes an arithmetic logic unit, a weight value interface, a feature value interface, a control module and a storage unit, the weight value interface includes an external input port and an input port of the same level, and the feature value interface includes an external input port, an upper-level input port and The same-level input port, the control module is used to store configuration information, and determine the data flow direction according to the setting instructions, and the storage unit is used to store weight values, feature values and calculation results.

The data processing system provided by the embodiment of the present application is introduced below, and the data processing system described below and the data processing method described above may be referred to in correspondence.

The present application also provides a data processing system, including a computing network, and the computing network includes a first computing core, several second computing cores, and several third computing cores; any second computing core corresponds to a superior computing core and two lower-level computing cores, the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is the upper-level computing core of the target second computing core.

The electronic device provided by the embodiment of the present application is introduced below, and the electronic device described below and the data processing method described above may be referred to in correspondence.

Please refer to FIG. 6 , which is a schematic structural diagram of an electronic device provided in an embodiment of the present application. The electronic device 100 may include a processor 101 and a memory 102 , and may further include one or more of a multimedia component 103 , an information input/information output (I/O) interface 104 and a communication component 105 .

Among them, the processor 101 is used to control the overall operation of the electronic device 100, so as to complete all or part of the steps in the above-mentioned data processing method; the memory 102 is used to store various types of data to support the operation of the electronic device 100, these data For example, instructions for any application or method operating on the electronic device 100 may be included, as well as application-related data. The memory 102 can be implemented by any type of volatile or non-volatile memory device or their combination, such as Static Random Access Memory (Static Random Access Memory, SRAM), Electrically Erasable Programmable Read-Only Memory (Electrically Erasable Erasable Programmable Read-Only Memory, EEPROM), Erasable Programmable Read-Only Memory (Erasable Programmable Read-Only Memory, EPROM), Programmable Read-Only Memory (Programmable Read-Only Memory, PROM), Read-Only Memory (Read-Only Memory) One or more of OnlyMemory, ROM), magnetic memory, flash memory, magnetic disk or optical disk.

Multimedia components 103 may include screen and audio components. The screen can be, for example, a touch screen, and the audio component is used for outputting and/or inputting audio signals. For example, an audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 102 or sent via the communication component 105 . The audio component also includes at least one speaker for outputting audio signals. The I/O interface 104 provides an interface between the processor 101 and other interface modules, which may be a keyboard, a mouse, buttons, and the like. These buttons can be virtual buttons or physical buttons. The communication component 105 is used for wired or wireless communication between the electronic device 100 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (Near Field Communication, NFC for short), 2G, 3G or 4G, or one or a combination of them, so the corresponding communication component 105 may include: Wi-Fi parts, Bluetooth parts, NFC parts.

The electronic device 100 may be implemented by one or more application-specific integrated circuits (Application Specific Integrated Circuit, ASIC for short), digital signal processors (Digital Signal Processor, DSP for short), digital signal processing devices (Digital Signal Processing Device, DSPD for short), Programmable logic device (Programmable Logic Device, be called for short PLD), field programmable gate array (Field ProgrammableGate Array, be called for short FPGA), controller, microcontroller, microprocessor or other electronic components realize, be used to carry out the above-mentioned embodiment given data processing method.

The computer-readable storage medium provided by the embodiments of the present application is introduced below, and the computer-readable storage medium described below and the data processing method described above may refer to each other correspondingly.

The present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned data processing method are implemented.

The computer-readable storage medium may include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk, etc., which can store program codes. medium.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for relevant details, please refer to the description of the method part.

Those skilled in the art can further appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the hardware and software In the above description, the components and steps of each example have been generally described according to their functions. Whether these functions are executed by means of hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may implement the described functionality using different methods for each particular application, but such implementation should not be considered as exceeding the scope of the present application.

The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be directly implemented by hardware, software modules executed by a processor, or a combination of both. Software modules can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other Any other known storage medium.

Finally, it should also be noted that in this article, relationships such as first and second etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations, any such actual relationship or order exists. Moreover, the term comprises, comprises, or any other variation is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements but also other elements not expressly listed, or Yes also includes elements inherent to such a process, method, article, or device.

In this paper, specific examples are used to illustrate the principles and implementation methods of the application. The descriptions of the above embodiments are only used to help understand the method and core idea of the application; meanwhile, for those of ordinary skill in the art, according to the application There will be changes in the specific implementation and scope of application. In summary, the content of this specification should not be construed as limiting the application.

Claims (9)

1. A data processing method, characterized in that, comprising: Obtaining a setting instruction, and setting a computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network; obtaining at least one eigenvalue, and inputting the at least one eigenvalue into at least one initial computing core in the computing network; Using the initial calculation core as a starting point, transmit the feature value according to the data flow direction; using each of the calculation cores to generate calculation results based on the feature values and corresponding weight values; Wherein, the computing network includes a first computing core, several second computing cores, and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and The third computing core, or includes the first computing core, the second computing core and the third computing core; any of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores , the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is a higher-level computing core of the target second computing core; Wherein, the data processing method also includes: determining weight values corresponding to each of the calculation cores; sending and storing the weight values to the corresponding calculation cores; The determination of the weight values corresponding to each of the calculation cores includes: Acquiring an initial weight value; based on the data flow direction, determining the correspondence between each of the calculation cores and the initial weight value, and completing the determination of the weight value; The setting instruction is used to set the computing network to be at least three-way computing, and the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, and the initial computing core includes The first computing core and the initial third computing core, the first computing core corresponds to the first flow direction, and the initial third computing core corresponds to the second flow direction; Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes: inputting a first eigenvalue into the first calculation core, and using the first calculation core to send the first eigenvalue to the target second calculation core; based on the first flow direction, from the target second calculation core Starting from two calculation cores, the first eigenvalues are sent to the corresponding lower-level calculation cores in sequence until the first eigenvalues are sent to the second calculation core at the end of the first stream; the second eigenvalues are sent Input the initial third computing core, and start from the initial third computing core, based on the second flow direction, send the second feature value to the subsequent computing cores of the same level sequentially, until the sending the second eigenvalue to the third computing core at the end of the second stream; The initiating computing core also includes initiating a second computing core, the second computing core corresponding to the second flow direction; Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes: inputting the second eigenvalue into the starting second computing core, and starting from the starting second computing core, based on the second flow direction, sequentially sending the second eigenvalue to subsequent computing cores of the same level , until the second feature value is sent to the second computing core at the end of the second stream. 2. The data processing method according to claim 1, further comprising: Obtain configuration information, and store the configuration information in each of the computing cores; wherein, the configuration information includes several identification information and corresponding several data flows; Correspondingly, setting the computing network according to the setting instruction includes: Sending the setting instruction to each of the computing cores, using the target identification information and the configuration information in the setting instruction to determine the data flow direction corresponding to each of the computing cores. 3. The data processing method according to claim 1, wherein the setting instruction is used to set the computing network to one-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level, and the The initial computing core is the first computing core; Correspondingly, the inputting the at least one feature value into at least one initial computing core in the computing network respectively; starting from the initial computing core, transmitting the feature value according to the data flow direction includes: inputting the feature value into the first computing core, and using the first computing core to send the feature value to the target second computing core; Based on the first flow direction, starting from the target second computing core, the feature values are sent to corresponding lower-level computing cores in sequence until the feature values are sent to the third computing core. 4. The data processing method according to claim 1, wherein said utilizing each said calculation core to generate a calculation result based on said feature value and a corresponding weight value comprises: controlling each of the calculation cores, and multiplying the feature value and the weight value to obtain a target result; The target result is added to the historical calculation result stored by the calculation core to obtain the calculation result. 5. the data processing method as claimed in claim 1, is characterized in that, described calculation kernel comprises arithmetic logic unit, weight value interface, characteristic value interface, control module and storage unit, and described weight value interface comprises external input port and Same-level input ports, the characteristic value interface includes external input ports, upper-level input ports, and same-level input ports, the control module is used to store configuration information, and determine the data flow direction according to the setting instructions, and the storage unit is used for Weight values, feature values, and the calculation results are stored. 6. A data processing device, characterized in that, comprising: A setting module, configured to obtain a setting instruction, and set a computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network; An input module, configured to acquire at least one eigenvalue, and respectively input the at least one eigenvalue into at least one initial calculation core in the calculation network; a transmission module, configured to use the initial calculation core as a starting point, and transmit the characteristic value according to the data flow direction; A calculation module, configured to use each of the calculation cores to generate calculation results based on the feature values and corresponding weight values; Wherein, the computing network includes a first computing core, several second computing cores, and several third computing cores; the initial computing core includes the first computing core, or includes the first computing core and The third computing core, or includes the first computing core, the second computing core and the third computing core; any of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores , the lower-level computing core is the second computing core or the third computing core, the third computing core does not have a lower-level computing core, and the first computing core is a higher-level computing core of the target second computing core; Wherein, the data processing device is specifically used for: determining weight values corresponding to each of the calculation cores; sending and storing the weight values to the corresponding calculation cores; The data processing device is specifically used for: Acquiring an initial weight value; based on the data flow direction, determining the correspondence between each of the calculation cores and the initial weight value, and completing the determination of the weight value; Wherein, the setting instruction is used to set the computing network to be at least three-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, and the initial calculation The core includes the first computing core and an initial third computing core, the first computing core corresponds to the first flow direction, and the initial third computing core corresponds to the second flow direction; The data processing device is specifically used for: inputting a first eigenvalue into the first calculation core, and using the first calculation core to send the first eigenvalue to the target second calculation core; based on the first flow direction, from the target second calculation core Starting from two calculation cores, the first eigenvalues are sent to the corresponding lower-level calculation cores in sequence until the first eigenvalues are sent to the second calculation core at the end of the first stream; the second eigenvalues are sent Input the initial third computing core, and start from the initial third computing core, based on the second flow direction, send the second feature value to the subsequent computing cores of the same level sequentially, until the sending the second eigenvalue to the third computing core at the end of the second stream; The initiating computing core also includes initiating a second computing core, the second computing core corresponding to the second flow direction; The data processing device is specifically used for: inputting the second eigenvalue into the starting second computing core, and starting from the starting second computing core, based on the second flow direction, sequentially sending the second eigenvalue to subsequent computing cores of the same level , until the second feature value is sent to the second computing core at the end of the second stream. 7. A data processing system, characterized in that it comprises a computing network, the computing network includes at least one initial computing core, and the computing network includes a first computing core, several second computing cores and several first computing cores Three computing cores; the initial computing core includes the first computing core, or includes the first computing core and the third computing core, or includes the first computing core, the second computing core and The third computing core; any of the second computing cores corresponds to one upper-level computing core and two lower-level computing cores, and the lower-level computing cores are the second computing core or the third computing core, and the The third computing core does not have a lower-level computing core, and the first computing core is the upper-level computing core of the target second computing core; The data processing system is used to obtain a setting instruction, and set the computing network according to the setting instruction; the setting instruction is used to set the data flow direction between each computing core in the computing network; The data processing system is used to obtain at least one eigenvalue, and input the at least one eigenvalue to at least one of the initial computing cores in the computing network; The data processing system is configured to use the initial calculation core as a starting point, and transmit the characteristic value according to the data flow direction; The data processing system is configured to use each of the calculation cores to generate calculation results based on the feature values and corresponding weight values; The data processing system is used for: determining weight values corresponding to each of the calculation cores; sending and storing the weight values to the corresponding calculation cores; The data processing system is used for: Acquiring an initial weight value; based on the data flow direction, determining the correspondence between each of the calculation cores and the initial weight value, and completing the determination of the weight value; Wherein, the setting instruction is used to set the computing network to be at least three-way computing, then the data flow direction is the first flow direction flowing from the upper level to the lower level and the second flow direction flowing between the same levels, and the initial calculation The core includes the first computing core and an initial third computing core, the first computing core corresponds to the first flow direction, and the initial third computing core corresponds to the second flow direction; The data processing system is used for: inputting a first eigenvalue into the first calculation core, and using the first calculation core to send the first eigenvalue to the target second calculation core; based on the first flow direction, from the target second calculation core Starting from two calculation cores, the first eigenvalues are sent to the corresponding lower-level calculation cores in sequence until the first eigenvalues are sent to the second calculation core at the end of the first stream; the second The feature value is input into the starting third computing core, and starting from the starting third computing core, based on the second flow direction, the second feature value is sent to the subsequent computing core at the same level in sequence until the sending the second feature value to the third calculation core at the end of the second stream; The initiating computing core also includes initiating a second computing core, the second computing core corresponding to the second flow direction; The data processing system is used for: inputting the second eigenvalue into the starting second computing core, and starting from the starting second computing core, based on the second flow direction, sequentially sending the second eigenvalue to subsequent computing cores of the same level , until the second feature value is sent to the second computing core at the end of the second stream. 8. An electronic device, comprising a memory and a processor, wherein: The memory is used to store computer programs; The processor is configured to execute the computer program to implement the data processing method according to any one of claims 1 to 5. 9. A computer-readable storage medium, characterized by being used to store a computer program, wherein, when the computer program is executed by a processor, the data processing method according to any one of claims 1 to 5 is implemented.