WO2022028224A1

WO2022028224A1 - Data storage method and apparatus, and device and storage medium

Info

Publication number: WO2022028224A1
Application number: PCT/CN2021/106394
Authority: WO
Inventors: 牛昕宇; 李远超; 蔡权雄
Original assignee: 深圳鲲云信息科技有限公司
Priority date: 2020-08-03
Filing date: 2021-07-15
Publication date: 2022-02-10
Also published as: CN111737193B; CN111737193A

Abstract

A data storage method and apparatus, and a device and a storage medium. The data storage method comprises: acquiring a first neural network computation type corresponding to first data to be subjected to computation; configuring a first preset rule on the basis of the first neural network computation type (S120); and on the basis of the first preset rule, storing the first data to be subjected to computation, such that the first data to be subjected to computation is sent, according to a first data sequence that matches the first neural network computation type, to a computation module in a data stream network for computation (S130).

Description

Data storage method, apparatus, device and storage medium

The present disclosure claims the priority of a Chinese patent application with application number 202010764719.8 filed with the Chinese Patent Office on Aug. 3, 2020, the entire contents of which are incorporated into the present disclosure by reference.

technical field

The embodiments of the present application relate to the technical field of neural networks, for example, to a data storage method, apparatus, device, and storage medium.

Background technique

With the rapid development of neural network technology, the amount of computational data in neural network is increasing, and data storage is an important part of neural network technology.

As far as the neural network chip developed based on the data flow architecture is concerned, the neural network model usually includes multiple network layers, and the structure type of each layer is called the network computing type of each layer. The computing data of the neural network usually corresponds to the network computing type. Therefore, when the computing module is fixed, the neural network chip needs to adopt different data storage methods for different network computing types. Usually, after the data and parameters of the data flow architecture chip are transferred from the off-chip memory to the on-chip memory, they are directly and sequentially read out to the subsequent computing modules for use, that is, external software is required to arrange the data in an orderly manner and then transfer it to the chip.

Among them, the chip does not include off-chip memory.

However, in a chip with a data flow architecture, the above method cannot flexibly support multiple types of networks. If the network computing parameters change, the chip can only support the corresponding network computing type by supplementing data and increasing the amount of computation. The computing efficiency of the chip is reduced, and once the external real-time data is changed, the computing needs cannot be met.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a data storage method, apparatus, device, and storage medium, so as to achieve the effect of storing and calculating data of different network computing types of a data flow architecture chip in different storage manners.

In a first aspect, an embodiment of the present application provides a neural network computing type storage method, including:

Obtain the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated;

Configure a first preset rule based on the first neural network calculation type;

The first data to be calculated is stored based on the first preset rule, so that the first data to be calculated is sent to the data stream network in the order of the first data matching the calculation type of the first neural network. The calculation module performs calculations;

Wherein, the first data to be calculated flows in the data flow network according to a preset data flow direction.

In a second aspect, an embodiment of the present application provides a neural network computing type storage device, including:

an acquisition module, configured to acquire the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated;

a configuration module, configured to configure a first preset rule based on the first neural network calculation type;

an on-chip storage module, configured to store the first data to be calculated based on the first preset rule, so that the first data to be calculated is sent in the order of the first data matching the calculation type of the first neural network Perform calculations for the calculation modules in the data flow network;

In a third aspect, an embodiment of the present application provides a device, including:

one or more processors;

storage means arranged to store 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 storage method according to any embodiment of the present application.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the program is executed by a processor, the data storage method described in any embodiment of the present application is implemented.

Description of drawings

1 is a schematic flowchart of a data storage method provided in Embodiment 1 of the present application;

2 is a schematic flowchart of a data storage method provided in Embodiment 2 of the present application;

3 is a schematic structural diagram of a data storage device according to Embodiment 3 of the present application;

4 is a schematic structural diagram of another data storage device provided in Embodiment 3 of the present application;

FIG. 5 is a schematic structural diagram of a device provided in Embodiment 4 of the present application.

detailed description

The present application will be described below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are only used to explain the present application, but not to limit the present application. For convenience of description, the drawings only show some but not all structures related to the present application.

Some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowchart depicts the various steps as a sequential process, many of the various steps may be performed in parallel, concurrently, or concurrently. Furthermore, the order of the various steps can be rearranged. Processing may be terminated when multiple step operations are completed, but may also have additional steps not included in the figures. A process may correspond to a method, function, procedure, subroutine, subroutine, or the like.

The terms "first," "second," etc. may be used herein to describe various directions, acts, steps or elements, etc., but are not limited by these terms. These terms are only used to distinguish a first direction, act, step or element from another direction, act, step or element. For example, the first preset rule may be referred to as the second preset rule, and similarly, the second preset rule may be referred to as the first preset rule, without departing from the scope of this application. Both the first preset rule and the second preset rule are preset rules, but the first preset rule and the second preset rule are not the same preset rule. The terms "first", "second", etc. should not be understood as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature defined as "first" or "second" may expressly or implicitly include one or more of that feature. In the description of this application, "plurality" and "batch" mean at least two, such as two, three, etc., unless otherwise defined.

Example 1

1 is a schematic flowchart of a data storage method provided in Embodiment 1 of the present application, which can be applied to a scenario where data of different neural network calculation types is calculated. The method can be executed by a data storage device, and the device can use software and/or hardware, and can be integrated on the device. For example, the device may be a chip.

As shown in FIG. 1 , the data storage method provided in Embodiment 1 of the present application includes:

S110. Acquire the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated.

In this embodiment, there are multiple layers in the neural network for which calculation operations are to be performed. After the calculation of the current layer is completed, the data operation of the next layer is performed. The first layer of the neural network refers to the first layer in the calculation of the neural network model. The first data to be calculated refers to data to be calculated in the first layer of the neural network. The first neural network calculation type refers to the neural network calculation type corresponding to the first data to be calculated. In one embodiment, the first neural network computation type may be a feedforward neural network, radial basis neural network, deep feedforward neural network, or recurrent neural network, etc. The first neural network computation type is not limited herein.

In one embodiment, the neural network calculation type corresponding to the data to be calculated in each of the multiple layers of the neural network may be preset.

S120. Configure a first preset rule based on the first neural network calculation type.

In this embodiment, the first preset rule refers to a rule applicable to data transmission and calculation corresponding to the first neural network calculation type. In one embodiment, the first preset rule refers to a rule for performing data processing on the data to be calculated that needs to be transmitted to the computing module on the chip in a certain manner. In this embodiment, the first preset rule is to store the first data to be calculated according to the order required by the calculation module, so as to send the first data to be calculated to the calculation module in the order required by the calculation module, and provide the data to the calculation module The rules for the module to perform first-level operations. By arranging the data to be calculated by the calculation module in advance, the calculation efficiency of the chip is improved. In one embodiment, before the neural network model starts to calculate, preset rules corresponding to different neural network calculation types are configured in advance. When the neural network model starts to calculate, according to the relationship between the configured neural network calculation types and the preset rules relationship, and the first neural network calculation type in the first data to be calculated, configure the first preset rule.

S130. Store the first data to be calculated based on the first preset rule, so that the first data to be calculated is sent to a data stream network in the order of the first data matching the calculation type of the first neural network The calculation module in the calculation.

In this embodiment, the calculation module performs data calculation on the first layer of the neural network model according to the first preset rule. Wherein, the first data to be calculated flows in the data flow network according to the preset data flow direction, the data flow can also be obtained without relying on the instruction, and the calculation module only needs to wait for the data to arrive to perform the calculation. The first data sequence refers to the data sequence mapped to the operation sequence of the first neural network calculation type. For example, when the data is processed according to the first neural network calculation type, the first data to be calculated is processed according to the operation sequence corresponding to the first neural network calculation type. In this embodiment, the first data to be calculated flows to the calculation module according to a data stream. Therefore, the first data to be calculated needs to be sorted according to the first data sequence and then sent to the calculation module for calculation. After arranging the first data to be calculated according to the first data sequence, the calculation module can cyclically control the read operation and control the address addressing of the read operation to extract the data for calculation.

In this embodiment, by determining the neural network calculation type of the first data to be calculated, and configuring preset rules for different neural network calculation types, the first data to be calculated is stored based on the configured preset rules for calculation. The module calculates the first data to be calculated, and the calculation module is not limited to calculating for one type of neural network calculation. In one embodiment, the calculation module performs fixed addition and multiplication operations when performing data calculation, and the on-chip memory in the chip satisfies different calculation modes of the calculation module by storing different data arrangements.

In one embodiment, the first data to be calculated is stored in the on-chip memory. The on-chip memory is connected to the computing module; in addition, the on-chip memory can also be connected to the computing module through intermediate components.

In one embodiment, the data flow network is a framework of the chip, and the neural network is a type of application mode of the chip.

In this embodiment of the present application, the first data to be calculated in the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated are obtained; the first preset is configured based on the first neural network calculation type The first data to be calculated is stored based on the first preset rule, so that the first data to be calculated is sent to the data flow network in the order of the first data matching the calculation type of the first neural network. The calculation module in the device performs calculation, wherein the first data to be calculated flows in the data flow network according to a preset data flow direction. The neural network calculation provided in this embodiment is not limited to one type of neural network, but the preset rules are configured through the neural network calculation type, which can cope with the single type of network calculation in the related art or the inability to flexibly support multiple types. In the working condition of the network, the data of different network computing types are stored in different storage methods to achieve the technical effect of computing the data of different network computing types in the data flow architecture. In addition, the embodiments of the present application can also maintain high computing efficiency.

Embodiment 2

FIG. 2 is a schematic flowchart of a data storage method provided in Embodiment 2 of the present application. This embodiment describes the foregoing embodiments, and is applicable to scenarios in which data of different neural network calculation types are calculated. The method can be performed by a data storage device, which can be implemented in software and/or hardware, and can be integrated on a device.

As shown in FIG. 2 , the data transmission method of the neural network provided by the second embodiment of the present application includes:

S210. Obtain the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated.

In this embodiment, there are multiple layers to be calculated in the neural network. After the calculation of the current layer is completed, the data operation of the next layer is performed. The first layer of the neural network refers to the first layer in the calculation of the neural network model. The first data to be calculated refers to data to be calculated in the first layer of the neural network. The first neural network calculation type refers to the neural network calculation type corresponding to the first data to be calculated. In one embodiment, the neural network computation type may be a feedforward neural network, a radial basis neural network, a deep feedforward neural network, or a recurrent neural network, etc. The computation type of the neural network is not limited herein.

In this embodiment, the first data to be calculated includes at least one calculation parameter. In one embodiment, the calculation parameters include a convolution kernel size (kernel size), a stride (stride), a convolution kernel channel (channel) and a filter (filter), etc. The types of calculation parameters are not limited herein. For example, kernel size, stride, channel, and filter of different neural network computation types will have at least one computation parameter different.

S220. Configure a first preset rule based on the first neural network calculation type.

In this embodiment, the first preset rule refers to a rule applicable to data transmission and calculation corresponding to the first neural network calculation type. In one embodiment, the first preset rule refers to a rule for performing data processing on the data to be calculated that needs to be transmitted to the computing module on the chip in a certain manner. In this embodiment, the first preset rule is to store the first data to be calculated according to the order required by the calculation module, so as to send the first data to be calculated to the calculation module in the order required by the calculation module, and provide the data to the calculation module The rules for the module to perform first-level operations. In one embodiment, before the neural network model starts to calculate, preset rules corresponding to different neural network calculation types are configured in advance. When the neural network model starts to calculate, according to the relationship between the configured neural network calculation types and the preset rules relationship, and the first neural network calculation type in the first data to be calculated, configure the first preset rule.

S230. Store the first data to be calculated based on the first preset rule, so that the first data to be calculated is sent to the data stream network in the order of the first data matching the calculation type of the first neural network The calculation module in the calculation.

In one embodiment, the calculation module performs data calculation on the first layer of the neural network model according to the first preset rule.

In this embodiment, by determining the neural network calculation type of the first data to be calculated, and configuring preset rules for different neural network calculation types, the first data to be calculated is stored based on the configured preset rules for calculation. The module calculates the first data to be calculated, and the calculation module is not limited to calculating for one type of neural network calculation. Wherein, the first data to be calculated flows in the data flow network according to a preset data flow direction.

In S230, the first preset rule includes a preset storage rule and a preset calculation rule, and S230 includes: storing the first data to be calculated based on the preset storage rule. After S230, the method further includes: transmitting the preset calculation rule to the calculation module, so that the calculation module can calculate the first data to be calculated according to the preset calculation rule.

In this embodiment, the preset storage rule refers to a rule for sorting data in a certain order. For example, the preset storage rule is a rule for storing the first data to be calculated according to the order of the first data. In one embodiment, the order of data storage is determined according to the order in which the computing module performs data computing, and the data is ordered in advance, which can improve the computing efficiency of the computing module. In one embodiment, the preset storage rule may be the sorting of different data, such as data, weight, and paranoia; it may also be the sorting of the same data, such as the pixel size of the picture or the three colors of red, blue, and yellow, which are not limited herein. . The preset calculation rule refers to the rule by which the calculation module calculates data. For example, the preset calculation rule is a rule for calculation according to the first neural network calculation type. In one embodiment, the preset calculation rule is to pull data in sequence for calculation.

In this embodiment, storing the first data to be calculated based on the preset storage rule includes: storing the first data to be calculated in the on-chip in the order of the first data based on the preset storage rule and a memory, so as to transmit the sorted first data to be calculated to the calculation module after sorting the first data to be calculated according to the order of the first data through the on-chip memory.

In this step, for the first data to be calculated that needs to be calculated in the neural network model, it is not necessary to identify the neural network calculation type of the first data to be calculated to transmit the data to different on-chip memories. After all neural network calculation types are transferred to the same on-chip memory, and the data is sorted according to the operation sequence of the neural network calculation type, the preset rules of the neural network calculation type are configured, which can reduce the layout space on the chip. As well as increasing the flexibility of data storage.

S240. Acquire the second data to be calculated of the second layer of the neural network and the second neural network calculation type corresponding to the second data to be calculated.

In this embodiment, the second layer of the neural network refers to the next layer of the first layer of the neural network. The second data to be calculated refers to data to be calculated in the second layer of the neural network. The second neural network calculation type refers to the neural network calculation type of the second data to be calculated. In one embodiment, the calculation type of the second neural network may be a feedforward neural network, a radial basis neural network, a deep feedforward neural network or a recurrent neural network, etc. The calculation type of the second neural network is not limited herein.

In this embodiment, the second data to be calculated includes at least one calculation parameter. In one embodiment, the calculation parameters include kernel size, stride, channel, filter, etc., and the types of the calculation parameters are not limited here.

S250. Determine whether the calculation type of the first neural network is the same as the calculation type of the second neural network.

Exemplarily, if the first neural network calculation type is a deep feedforward neural network, and the second neural network calculation type is also a deep feedforward neural network, the first neural network calculation type and the second neural network calculation type are the same. If the first neural network computation type is a deep feedforward neural network and the second neural network computation type is a recurrent neural network, the first neural network computation type and the second neural network computation type are not the same.

S260. In response to a judgment result that the first neural network calculation type is the same as the second neural network calculation type, store the second data to be calculated based on the first preset rule.

In one embodiment, the calculation type of the first neural network is the same as the calculation type of the second neural network, the first preset rule is also applicable to the calculation of the neural network calculation type of the second layer, and the second layer of the neural network can be directly activated The second data to be calculated is calculated in the second layer of the neural network.

S270. In response to the judgment result that the first neural network calculation type is different from the second neural network calculation type, configure a second preset rule based on the second neural network calculation type, and based on the second preset The rule stores the second data to be calculated.

Wherein, the second data to be calculated stored according to the second preset rule is used to be sent to the calculation module in the data flow network for calculation according to the second data sequence matching the calculation type of the second neural network .

For example, a second preset rule is configured based on the second neural network calculation type, and the second data to be calculated is stored based on the second preset rule, so that the second data to be calculated is calculated according to the first The second data matching the calculation types of the two neural networks is sequentially sent to the calculation module in the data flow network for calculation.

In this embodiment, the second preset rule refers to a rule applicable to data transmission and calculation corresponding to the second neural network calculation type. The second data order refers to the data order mapped to the operation order of the second neural network calculation type. For example, when the calculation type of the second neural network is different from the calculation type of the first neural network, for example, when the size of the convolution kernel is different, the required data order is different, so when the calculation type of the first neural network and the calculation type of the second neural network are different At different times, the second data to be calculated needs to be stored in the on-chip memory in the order of the second data corresponding to the second neural network calculation type, so as to be used by the calculation module for calculation when flowing to the calculation module. In one embodiment, the second preset rule refers to a rule for performing data processing on the data to be calculated that needs to be transmitted to the computing module on the chip in a certain manner. In this embodiment, the second preset rule is to store the second data to be calculated according to the order required by the calculation module, so as to send the second data to be calculated to the calculation module in the order required by the calculation module, and provide the data to the calculation module. The rules for the module to perform the second-level operations.

In an alternative embodiment, it may not be necessary to determine whether the calculation type of the second neural network is the same as the calculation type of the first neural network. The second preset rule is configured directly according to the second neural network calculation type. If the calculation type of the second neural network is the same as the calculation type of the first neural network, the second preset rule and the first preset rule are also the same.

In another alternative embodiment, the data to be calculated of the third layer or even more layers of the neural network is acquired, and the level of the acquired data to be calculated needs to be determined according to the number of levels of the computational model of the neural network.

For example, when the data is image data, in the field of artificial intelligence, convolutional networks are generally used to process image data. For a convolutional network, there will be multiple convolutional layers, and different convolutional layers have different convolution kernel sizes, step sizes and filters, that is, different convolutional layers correspond to different types of neural network computations. When the image data enters the start processing of the first convolutional layer, first determine the calculation type of the first neural network of the first convolutional layer, and then according to the channel/width (width)/height characteristics of the image data, the first The first preset rule corresponding to the neural network calculation type sorts the image data according to the operation sequence required by the first neural network calculation type, stores it in the on-chip memory, and directly sends the data stream in the first data sequence to the calculation module Calculate, when the first convolutional layer calculation is completed, output the first feature map (feature map). The image data entering the first convolutional layer is also a feature map. When the first feature map starts to enter the second convolutional layer for processing, first determine whether the second neural network calculation type of the second convolutional layer is the same as the first neural network calculation type. When the calculation type of the second neural network is different from the calculation type of the first neural network, according to the channel/width/height characteristics of the second data to be calculated, according to the second preset rule corresponding to the second neural network calculation type After the feature map is sorted according to the operation sequence required by the second neural network calculation type and stored in the on-chip memory, it is directly sent to the calculation module for calculation according to the data stream of the second data sequence. And so on, the image processing is done in the convolutional network.

It is understandable that when the neural network calculation types are different, the data is linearly transformed to meet the calculation needs of the calculation module, which can ensure that the convolution operation of different features of multiple neural networks is compatible with the data flow architecture. Adding extra time consumption ensures the efficiency and flexibility of the operation of the data flow structure, and can satisfy the convolution operation of multiple neural networks in the data flow structure.

In this embodiment of the present application, the first data to be calculated in the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated are obtained; the first preset is configured based on the first neural network calculation type The first data to be calculated is stored based on the first preset rule, so that the first data to be calculated is sent to the data flow network in the order of the first data matching the calculation type of the first neural network. The calculation module in the calculation module performs calculation, after the calculation module calculates the first to-be-calculated data, obtains the second to-be-calculated data of the second layer of the neural network and the second neural network calculation type corresponding to the second to-be-calculated data, By judging whether the calculation type of the first neural network is the same as the calculation type of the second neural network, the data storage mode for the second-layer neural network is determined. The neural network calculation provided in this embodiment is not limited to one type of neural network, but configures preset rules through the neural network calculation type, which can cope with the network calculation type in the related art that supports a single type or cannot flexibly support multiple types of networks The data of different network computing types are stored in different storage methods to achieve the technical effect of computing data of different network computing types. In addition, the embodiments of the present application can also maintain high computing efficiency.

In this embodiment of the present application, the first data to be calculated in the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated are obtained; the first preset rule is configured based on the first neural network calculation type ; Based on the first preset rule, the first data to be calculated is stored, so that the first data to be calculated is sent to the data flow network in the order of the first data matching the calculation type of the first neural network. The calculation module performs calculation; wherein, the first data to be calculated flows in the data flow network according to the preset data flow direction, and when the chip of the data flow architecture supports a variety of neural network calculation types, it is necessary to pass the supplementary data It is realized by increasing the amount of calculation, reducing the computing efficiency of the chip, and realizing the data flow architecture of different network computing types. Effect.

Embodiment 3

FIG. 3 is a schematic structural diagram of a data storage device provided in Embodiment 3 of the present application. This embodiment is applicable to a scenario in which data of different neural network calculation types is calculated, and the device can be implemented in software and/or hardware. , and can be integrated on the device.

As shown in FIG. 3 , the data storage device provided in this embodiment may include an acquisition module 310, a configuration module 320, and an on-chip storage module 330, wherein: the acquisition module 310 is configured to acquire the first data to be calculated of the first layer of the neural network the first neural network calculation type corresponding to the first data to be calculated; the configuration module 320 is configured to configure a first preset rule based on the first neural network calculation type; the on-chip storage module 330 is configured to be based on the first neural network calculation type. A preset rule stores the first data to be calculated, so that the first data to be calculated is sent to the calculation module in the data flow network for calculation according to the first data sequence matching the calculation type of the first neural network ; wherein, the first data to be calculated flows in the data flow network according to a preset data flow direction.

Referring to FIG. 4 , in one embodiment, the obtaining module 310 is further configured to obtain the second data to be calculated of the second layer of the neural network and the second neural network calculation type corresponding to the second data to be calculated.

The device further includes: a judgment module 340, the judgment module 340 is configured to judge whether the calculation type of the first neural network is the same as the calculation type of the second neural network;

The on-chip storage module 330 is further configured to, in response to the judgment result that the first neural network calculation type is the same as the second neural network calculation type, perform the calculation on the second data to be calculated based on the first preset rule. storing; and in response to a judgment result that the first neural network calculation type is different from the second neural network calculation type, configure a second preset rule based on the second neural network calculation type, and configure a second preset rule based on the second neural network calculation type. The preset rule stores the second data to be calculated, wherein the second data to be calculated stored according to the second preset rule is used for the second data sequence matched according to the second neural network calculation type, It is sent to the calculation module in the data flow network for calculation.

In one embodiment, the first preset rule includes a preset storage rule and a preset calculation rule.

The on-chip storage module 330 includes: a storage unit and a sending unit;

A storage unit configured to store the first data to be calculated based on the preset storage rule, wherein the preset storage rule is to store the first data to be calculated in the order of the first data a rule; a sending unit configured to transmit the preset calculation rule to the calculation module, so that the calculation module can calculate the first data to be calculated according to the preset calculation rule.

Wherein, the preset calculation rule is a rule for calculation according to the first neural network calculation type.

In one embodiment, the first data to be calculated includes at least one calculation parameter, and the second data to be calculated includes at least one calculation parameter.

In one embodiment, the calculation parameters include: at least one of kernel size, stride, channel, and filter.

In one embodiment, the storage unit is further configured to store the first data to be calculated in the on-chip memory in the order of the first data based on the preset storage rule, so as to store the first data in the on-chip memory through the on-chip memory. After the data to be calculated is sorted according to the first data sequence, it is transmitted to the calculation module.

The data storage device provided by the embodiment of the present application can execute the data storage method provided by any embodiment of the present application, and has functional modules and beneficial effects corresponding to the execution method. For content not described in this embodiment, reference may be made to the description in any method embodiment of this application.

Embodiment 4

FIG. 5 is a schematic structural diagram of a device provided in Embodiment 4 of the present application. FIG. 5 shows a block diagram of an exemplary apparatus 612 suitable for implementing embodiments of the present application. The device 612 shown in FIG. 5 is only an example, and should not impose any limitations on the functions and scope of use of the embodiments of the present application.

As shown in FIG. 5, device 612 takes the form of a generic device. Components of device 612 may include, but are not limited to, one or more processors 616 , storage 628 , and bus 618 connecting various system components, such as storage 628 and processor 616 .

Bus 618 represents one or more of several types of bus structures, including a storage device bus or storage device controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of a variety of bus structures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, enhanced ISA bus, Video Electronics Standards Association (Video Electronics Standards) Association, VESA) local bus and Peripheral Component Interconnect (PCI) bus.

In one embodiment, device 612 includes various computer system readable media. These media can be any available media that can be accessed by device 612, including volatile and non-volatile media, removable and non-removable media.

Storage 628 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 630 and/or cache 632 . In one embodiment, terminal 612 may include other removable/non-removable, volatile/non-volatile computer system storage media. For example only, storage system 634 may be used to read and write to non-removable, non-volatile magnetic media, not shown in FIG. 5, which are commonly referred to as hard disk drives. Although not shown in FIG. 5, a magnetic disk drive for reading and writing to removable non-volatile magnetic disks, such as floppy disks, and removable non-volatile optical disks, such as Compact Disc Read-Only Memory, may be provided. CD-ROM), Digital Video Disc (Digital Video Disc-Read Only Memory, DVD-ROM) or other optical media read and write optical disc drives. In these cases, each drive may be connected to bus 618 through one or more data media interfaces. The storage device 628 may include at least one program product having a set of, eg, at least one program module configured to perform the functions of the embodiments of the present application.

Program/utility 640 having, for example, a set of at least one program module 642, which may be stored, for example, in storage device 628, such program module 642 including, but not limited to, an operating system, one or more application programs, other program modules, and program data , each or a combination of these examples may include an implementation of a network environment. Program modules 642 generally perform the functions and/or methods of the embodiments described herein.

Device 612 may also communicate with one or more external devices 614, such as a keyboard, pointing terminal, display 624, etc., and one or more terminals that enable a user to interact with the device 612, and/or with Any terminal (eg, network card, modem, etc.) that enables the device 612 to communicate with one or more other computing terminals. Such communication may take place through an input/output (I/O) interface 622 . Also, the device 612 may communicate with one or more networks, such as a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network, such as the Internet, through a network adapter 620. As shown in FIG. 5 , network adapter 620 communicates with other modules of device 612 via bus 618 . Although not shown, other hardware and/or software modules may be used in conjunction with device 612, including but not limited to: microcode, terminal drivers, redundant processors, external disk drive arrays, Redundant Arrays of Independent Disks, RAID) systems, tape drives, and data backup storage systems.

The processor 616 executes a variety of functional applications and data processing by running the programs stored in the storage device 628, for example, implementing a data storage method provided by any embodiment of the present application, the method may include: obtaining the first data of the neural network. The first data to be calculated in one layer and the first neural network calculation type corresponding to the first data to be calculated; configure a first preset rule based on the first neural network calculation type; based on the first preset rule The first data to be calculated is stored, so that the first data to be calculated is sent to the calculation module in the data flow network for calculation according to the first data sequence matching the calculation type of the first neural network; wherein, the first The data to be calculated flows in the data flow network according to a preset data flow direction.

An embodiment of the present application further provides a system, including: a configuration device, a direct memory access (Direct Memory Access, DMA) device, an external storage device, and an on-chip storage device. The configuration device can be a processor; the external storage device can be a double-rate (Double Data Rate, DDR) memory; the DMA device connects the external storage device and the on-chip storage device, and the DMA device is controlled by the configuration device; the DMA device is configured according to the processor , the data in the external storage device is moved to the on-chip storage device, and the on-chip storage device stores the data according to the processor configuration. When the DMA device is executed by the processor, the DMA device transfers the data from the external storage device to the on-chip storage device according to the processor configuration, and when the on-chip storage device is configured by the processor, the on-chip storage device transfers the data from the DMA device to the on-chip storage device according to the processor configuration. The configuration mode is stored in on-chip storage, and implements the data storage method provided by any embodiment of the present application.

External storage devices may be referred to as off-chip memory, and on-chip storage devices may be referred to as on-chip memory.

Embodiment 5

Embodiment 5 of the present application further provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the program is executed by a processor, implements a data storage method as provided in any embodiment of the present application, the The method may include: acquiring first data to be calculated in the first layer of the neural network and a first neural network calculation type corresponding to the first data to be calculated; configuring a first preset rule based on the first neural network calculation type; The first preset rule stores the first data to be calculated, so that the first data to be calculated is sent to the calculation module in the data flow network according to the first data sequence matching the calculation type of the first neural network Calculate; wherein, the first data to be calculated flows in the data flow network according to a preset data flow direction.

The computer-readable storage medium of the embodiments of the present application may adopt any combination of one or more computer-readable mediums. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. 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. Examples (non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, RAM, Read-Only Memory (ROM), erasable Erasable Programmable Read-Only Memory (EPROM or flash memory), optical fiber, CD-ROM, optical storage device, magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can 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.

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 in the computer-readable signal medium. 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 .

The program code contained on the storage medium can be transmitted by any suitable medium, including - but not limited to wireless, wire, optical fiber cable, radio frequency (RF), etc., or any suitable combination of the above.

Computer program code for performing the operations of the present application may be written in one or more programming languages, including object-oriented programming languages—such as Java, Smalltalk, C++, but also conventional Procedural programming language - such as the "C" language or similar programming language. 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 terminal. Where a remote computer is involved, the remote computer may be connected to the user's computer through any kind of network, including a LAN or WAN, or may be connected to an external computer, such as through the Internet using an Internet service provider.

In this embodiment of the present application, the first data to be calculated in the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated are obtained; the first preset is configured based on the first neural network calculation type The first data to be calculated is stored based on the first preset rule, so that the first data to be calculated is sent to the data flow network in the order of the first data matching the calculation type of the first neural network. The calculation module in the device performs calculation, wherein the first data to be calculated flows in the data flow network according to a preset data flow direction. The neural network calculation provided in this embodiment is not limited to one type of neural network, but configures preset rules through the neural network calculation type, which can cope with the related art where the network calculation type is single or cannot flexibly support multiple types of networks. Working conditions, the data of different network computing types are stored in different storage methods to achieve the technical effect of computing the data of different network computing types in the data flow architecture. In addition, the embodiments of the present application can also maintain high computing efficiency.

The embodiment of the present application also provides a storage medium with a processor-configurable storage arrangement, in which external storage data can be stored in different arrangement sequences, and the data is transmitted by the processor through the DMA device to the storage device for execution. The data storage method provided by any embodiment of the present application.

Claims

A data storage method comprising:

Obtain the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated;

Configure a first preset rule based on the first neural network calculation type;

The first data to be calculated is stored based on the first preset rule, so that the first data to be calculated is sent to the data stream network in the order of the first data matching the calculation type of the first neural network. The calculation module performs calculation; wherein, the first data to be calculated flows in the data flow network according to a preset data flow direction.
The method according to claim 1, after storing the first data to be calculated based on the first preset rule, further comprising:

Obtain the second data to be calculated of the second layer of the neural network and the second neural network calculation type corresponding to the second data to be calculated;

judging whether the calculation type of the first neural network is the same as the calculation type of the second neural network;

In response to the judgment result that the calculation type of the first neural network is the same as the calculation type of the second neural network, the second data to be calculated is stored based on the first preset rule;

In response to the judgment result that the calculation type of the first neural network is different from the calculation type of the second neural network, a second preset rule is configured based on the calculation type of the second neural network, and based on the second preset rule The second data to be calculated is stored, wherein the second data to be calculated stored according to the second preset rule is used for the second data sequence matching the calculation type of the second neural network to be sent to the The calculation module in the data flow network described above performs the calculation.
The method according to claim 1 or 2, wherein the first preset rule includes a preset storage rule and a preset calculation rule;

The storing the first data to be calculated based on the first preset rule includes:

The first data to be calculated is stored based on the preset storage rule, wherein the preset storage rule is a rule for storing the first data to be calculated in the order of the first data;

After the first preset rule is configured based on the first neural network calculation type, the method further includes:

The preset calculation rule is transmitted to a calculation module for the calculation module to calculate the first data to be calculated according to the preset calculation rule.
The method of claim 2, wherein the first data to be calculated includes at least one calculation parameter, and the second data to be calculated includes at least one calculation parameter.
The method of claim 4, wherein the calculation parameters include at least one of: convolution kernel size kernel size, stride stride, convolution kernel channel channel and filter filter.
The method of claim 3, wherein the storing the first data to be calculated based on the preset storage rule comprises:

The first data to be calculated is stored in the on-chip memory in the order of the first data based on the preset storage rule, so that the first data to be calculated is stored in the on-chip memory in the order of the first data After sorting, the sorted first data to be calculated is transmitted to the calculation module.
A data storage device, comprising:

an acquisition module, configured to acquire the first data to be calculated of the first layer of the neural network and the first neural network calculation type corresponding to the first data to be calculated;

a configuration module, configured to configure a first preset rule based on the first neural network calculation type;

An on-chip storage module, configured to store the first data to be calculated based on the first preset rule, so that the first data to be calculated is sent in the order of the first data matching the calculation type of the first neural network Perform calculations for the calculation modules in the data flow network;

The first data to be calculated flows in the data flow network according to a preset data flow direction.
The device according to claim 7, wherein the obtaining module is further configured to obtain the second data to be calculated of the second layer of the neural network and the second neural network calculation type corresponding to the second data to be calculated;

The device further includes: a judgment module, the judgment module is configured to judge whether the calculation type of the first neural network is the same as the calculation type of the second neural network;

The on-chip storage module is further configured to, in response to a judgment result that the calculation type of the first neural network is the same as the calculation type of the second neural network, perform a calculation on the second data to be calculated based on the first preset rule. storing; and in response to a judgment result that the first neural network calculation type is different from the second neural network calculation type, configure a second preset rule based on the second neural network calculation type, and based on the second preset Set the rules to store the second data to be calculated, wherein the second data to be calculated stored according to the second preset rule is used for the second data sequence matched according to the second neural network calculation type, is It is sent to the calculation module in the data flow network for calculation.
The apparatus according to claim 7 or 8, wherein the first preset rule includes a preset storage rule and a preset calculation rule;

The on-chip storage module includes a storage unit and a sending unit;

The storage unit is configured to store the first data to be calculated based on the preset storage rule, wherein the preset storage rule is to store the first data to be calculated in the order of the first data storage rules;

The sending unit is configured to transmit the preset calculation rule to a calculation module, so that the calculation module can calculate the first data to be calculated according to the preset calculation rule.
The apparatus of claim 8, wherein the first data to be calculated includes at least one calculation parameter, and the second data to be calculated includes at least one calculation parameter.
The apparatus of claim 10, wherein the calculation parameters include at least one of: a convolution kernel size kernel size, a stride stride, a convolution kernel channel channel, and a filter filter.
The apparatus according to claim 9, wherein the storage unit is further configured to store the first data to be calculated in the on-chip memory in the order of the first data based on the preset storage rule, so as to pass the After the on-chip memory sorts the first data to be calculated according to the order of the first data, the sorted first data to be calculated is transmitted to the calculation module.
A device comprising:

at least one processor;

a storage device configured to store at least one program;

When the at least one program is executed by the at least one processor, the at least one processor is caused to implement the data storage method according to any one of claims 1-6.
A computer-readable storage medium storing a computer program, when the computer program is executed by a processor, implements the data storage method according to any one of claims 1-6.