CN116325703A - A data format processing method and device - Google Patents

Abstract

The embodiment of the application discloses a data format processing method and device, relates to the technical field of artificial intelligence, solves the problem of high performance overhead caused by excessive format conversion processing in the prior art, and improves user experience. The specific scheme is as follows: the method is applied to a computing network, and the computing network sequentially comprises a first computing unit and a second computing unit, and comprises the following steps: acquiring data formats supported by a first computing unit and a second computing unit; if the first computing unit is determined not to support the format of the input data of the first computing unit, adding a first format conversion unit before the first computing unit, wherein the first format conversion unit is used for converting the format of the input data of the first computing unit into a data format supported by the first computing unit; if it is determined that the second computing unit supports the format of the output data of the first computing unit, it is determined that the format conversion process is not performed between the first computing unit and the second computing unit.

Description

A data format processing method and device technical field

The embodiments of the present application relate to the technical field of artificial intelligence, and in particular, to a data format processing method and device.

Background technique

In an artificial intelligence (AI) chip, in order to give full play to the computing power of the chip, it is usually required that the data be arranged in a specific format. Since different operations require different data formats, it may be necessary to process data in different formats. convert.

Fig. 1 is a schematic structural diagram of a computing network. As shown in Fig. 1, the operator sequence in the computing network includes convolution Conv1 operator, activation function unit Relu1 operator, Conv2 operator and Relu2 operator in turn, wherein, The Conv operator requires the format of the input data to be arranged in the NC1HWC0 format, and the Relu operator requires the format of the input data to be arranged in the ND format. Since the data format of the original image is NHWC format, which is different from the format of the input data required by the Conv1 operator, the format of the input data needs to be converted.

Figure 2 is a schematic diagram of data format conversion processing, as shown in Figure 2, the format conversion processing can be added by adding format conversion processing logic at the start position and end position in the Conv operator, and the start position in the Conv operator The logic can convert the format of the input data from the NHWC format to the NC1HWC0 format, and the format conversion processing logic added to the end position in the Conv operator can convert the format of the output data from the NC1HWC0 format to the NHWC format. However, this method increases the time consumption of the operator and reduces the performance of the operator after the format conversion processing logic is added inside the operator. If there are multiple Conv operators in the network, each Conv operator has conversion logic inside, so there are more format conversion processes, resulting in a large network performance overhead.

Contents of the invention

Embodiments of the present application provide a data format processing method and device, which can reduce format conversion processing and improve network performance.

In order to achieve the above purpose, the embodiment of the present application adopts the following technical solutions:

According to the first aspect of the embodiments of the present application, a data format processing method is provided, which is applied to a computing network, and the computing network includes a first computing unit and a second computing unit in sequence, and the method includes: obtaining the first computing unit and the The data format supported by the second calculation unit; if it is determined that the first calculation unit does not support the format of the input data of the first calculation unit, a first format conversion unit is added before the first calculation unit, and the first format conversion unit is used to convert the first calculation unit The format of the input data of a computing unit is converted to the data format supported by the first computing unit; if it is determined that the second computing unit supports the format of the output data of the first computing unit, it is determined not to perform between the first computing unit and the second computing unit Format conversion processing. Based on this scheme, when the first computing unit does not support the format of the input data of the first computing unit, a first format conversion unit is added before the first computing unit to convert the format of the input data of the first computing unit into the first For the data format supported by the computing unit, when the second computing unit supports the format of the output data of the first computing unit, it is determined not to perform format conversion processing between the first computing unit and the second computing unit, that is, it is determined not to perform format conversion processing between the first computing unit and the second computing unit. The format conversion operator is inserted between the second calculation units. Therefore, the number of format conversion units in the network is reduced, the format processing in the network is reduced, and the network performance is improved.

Optionally, when the computing network includes three or more computing units, the above-mentioned first computing unit and the second computing unit may be any consecutive two computing units in the computing network. For example, when the first computing unit is the first computing unit of the computing network, the second computing unit is the computing unit after the first computing unit. For another example, when the first computing unit is a computing unit located in the middle of the computing network, the second computing unit is a computing unit after the first computing unit.

It can be understood that when the application performs format processing, when the second operator does not support the format of the output data of the first operator, a data conversion unit is inserted between the first operator and the second operator , when the third operator supports the output data format of the second operator, it is determined not to insert a data conversion unit between the second operator and the third operator, when the fourth operator supports the third When determining the format of the output data of an operator, it is determined not to insert a data conversion unit between the third operator and the fourth operator until the Nth operator does not support the format of the output data of the N-1th operator , the data conversion unit will be inserted between the N-1th operator and the Nth operator, that is, the application considers the continuity of the data format supported by multiple operators when judging whether to perform data format conversion processing , compared with the prior art, the format conversion process can be greatly reduced, and the network performance can be improved.

With reference to the first aspect, in a possible implementation manner, when the first computing unit is the first computing unit in the computing network, the format of the input data of the first computing unit is the original The format of the input data. Based on this scheme, when the first computing unit does not support the format of the original input data, a format conversion processing unit is inserted before the first computing unit to convert the format of the original input data into a data format supported by the first computing unit, thereby It can make the operation of the first computing unit go on normally.

In combination with the first aspect and the above-mentioned possible implementation, in another possible implementation, when the above-mentioned first computing unit is a computing unit located in the middle of the above-mentioned computing network, the input data of the above-mentioned first computing unit The format of is the format of the output data of the calculation units preceding the first calculation unit. Based on this scheme, when the first computing unit in the middle does not support the format of the output data of the previous computing unit, a format conversion processing unit is inserted between the previous computing unit of the first computing unit and the first computing unit , to convert the output data format of the previous computing unit of the first computing unit into a data format supported by the first computing unit, so that the operation of the first computing unit can be performed normally.

In combination with the first aspect and the above-mentioned possible implementation, in another possible implementation, in the case that the above-mentioned second computing unit supports multiple data formats, it is determined that the above-mentioned second computing unit supports the output of the above-mentioned first computing unit The data format includes: if the multiple data formats supported by the second computing unit include the format of the output data of the first computing unit, determining that the second computing unit supports the format of the output data of the first computing unit. Based on this scheme, when the multiple data formats supported by the second computing unit include the format of the output data of the first computing unit, it can be determined that the second computing unit supports the format of the output data of the first computing unit, so that it is determined that the format of the output data of the first computing unit is not Format conversion processing is performed between the first computing unit and the second computing unit, thereby reducing the number of format conversion processing times in the network and improving network performance. Optionally, in a case where a data format supported by the second computing unit is the same as the output data format of the first computing unit, it may be determined that the second computing unit supports the format of the output data of the first computing unit.

In combination with the first aspect and the above possible implementation manner, in another possible implementation manner, the acquisition of the data format supported by the first computing unit and the second computing unit includes: acquiring the data format supported by the first computing unit and the second computing unit 2. The preset mode type of the computing unit; the preset mode type includes format-independent class, reduction Reduce class and broadcast Broadcast class; according to the preset mode types of the first computing unit and the second computing unit, determine the first The data format supported by the computing unit and the above-mentioned second computing unit. Based on this scheme, the mode type can be divided for each operator in the operator information base, so that the data format supported by the operator can be determined according to the mode type information of the operator, which can reduce the amount of information in the operator information base. Each format supported by the operator needs to be configured. Moreover, configuring the preset mode type of the operator through the operator information library can enable the implementation of the operator without paying attention to the specific format, and can directly perform operations according to the input shape and attributes of the operator, reducing the burden of format processing in operator development. .

In combination with the first aspect and the above possible implementation manner, in another possible implementation manner, the above method further includes: if it is determined that the format of the output data of the last computing unit of the above computing network is different from the format of the original output data of the above computing network The formats are different, and a second format conversion unit is added after the last calculation unit of the above-mentioned calculation network, and the second format conversion unit is used to convert the format of the output data of the above-mentioned last calculation unit into the format of the original output data of the above-mentioned calculation network . Based on this solution, when the format of the output data of the last computing unit of the computing network is different from that of the original output data, format conversion is performed to keep the format of the output data of the computing network consistent with the format of the original output data. Optionally, when the second computing unit is the last computing unit of the computing network, if the format of the output data of the second computing unit is different from that of the original output data, a format conversion operator is inserted after the second computing unit to Converting the format of the output data of the second computing unit to the format of the original output data.

The second aspect of the embodiment of the present application provides a data format processing device, which is applied to a computing network, and the computing network includes a first computing unit and a second computing unit in sequence, and the device includes: a processing unit and an acquisition unit; The obtaining unit is used to obtain the data format supported by the first computing unit and the second computing unit; the processing unit is used to determine that the first computing unit does not support the format of the input data of the first computing unit, in the first computing unit A first format conversion unit is added before, and the first format conversion unit is used to convert the format of the input data of the first computing unit into a data format supported by the first computing unit; the processing unit is also used to determine that the second computing unit supports The format of the output data of the first computing unit is determined not to perform format conversion processing between the first computing unit and the second computing unit.

With reference to the second aspect, in a possible implementation manner, when the first computing unit is the first computing unit in the computing network, the format of the input data of the first computing unit is the original The format of the input data.

With reference to the second aspect and the above possible implementation manner, in another possible implementation manner, when the above-mentioned first computing unit is a computing unit at an intermediate position in the above-mentioned computing network, the input data of the above-mentioned first computing unit The format of is the format of the output data of the calculation units preceding the first calculation unit.

In combination with the second aspect and the above possible implementation, in another possible implementation, the processing unit is specifically configured to determine that the multiple data formats supported by the second computing unit include the output data of the first computing unit The format of the output data of the first computing unit supported by the second computing unit is determined.

With reference to the second aspect and the above possible implementation manner, in another possible implementation manner, the acquisition unit is further configured to acquire the preset mode types of the first calculation unit and the second calculation unit; the preset mode The types include format-independent class, reduction class, and broadcast broadcast class; the above-mentioned processing unit is also used to determine the above-mentioned first computing unit and the above-mentioned second computing unit according to the preset mode types of the above-mentioned first computing unit and the above-mentioned second computing unit The data format supported by the computing unit.

In combination with the second aspect and the above possible implementation manner, in another possible implementation manner, the above processing unit is further configured to: if it is determined that the output data format of the last computing unit of the above computing network is the same as the original The format of the output data is different, and a second format conversion unit is added after the last calculation unit of the above-mentioned calculation network, and the second format conversion unit is used to convert the format of the output data of the above-mentioned last calculation unit into the original output of the above-mentioned calculation network The format of the data.

The third aspect of the embodiments of the present application provides an electronic device, which can implement the data format processing method described in the first aspect, and implement the above method through software, hardware, or by executing corresponding software through hardware. In a possible implementation manner, the electronic device may include a processor and a memory. The processor is configured to support the electronic device to execute corresponding functions in the method described in the first aspect above. The memory is used to be coupled with the processor, and it stores necessary program instructions and data of the electronic device.

The fourth aspect of the embodiments of the present application provides a computer-readable storage medium, the computer-readable storage medium includes computer instructions, and when the computer instructions are run on the electronic device, the electronic device executes the above-mentioned first aspect and The data format processing method described in its possible implementation manner.

In the fifth aspect of the embodiments of the present application, a computer program product is provided. When the computer program product is run on a computer, the computer executes the data format as described in the above first aspect and its possible implementation manners. Approach.

A sixth aspect of the embodiments of the present application provides an electronic device, the electronic device includes a central processing unit (Central Process Unit, CPU) and a neural network processor (Neural-network Processing Unit, NPU), the CPU is used to execute According to the data format processing method described in the first aspect and possible implementations thereof, the NPU is configured to execute the computing network obtained by the CPU based on the data format processing method.

For the description of the effects of the second aspect, the third aspect, the fourth aspect, the fifth aspect and the sixth aspect, reference may be made to the description of the corresponding effects of the first aspect, which will not be repeated here.

Description of drawings

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

FIG. 2 is a schematic diagram of a data format conversion process provided by an embodiment of the present application;

Fig. 3 is a schematic diagram of the data composition of a picture provided by the embodiment of the present application;

FIG. 4 is a schematic diagram of a data format provided by an embodiment of the present application;

FIG. 5 is a schematic diagram of another data format conversion process provided by the embodiment of the present application;

FIG. 6 is a schematic structural diagram of a system architecture provided by an embodiment of the present application;

FIG. 7 is a schematic flowchart of a data format processing method provided by an embodiment of the present application;

FIG. 8 is an application schematic diagram of a data format processing method provided in the embodiment of the present application;

FIG. 9 is an application schematic diagram of another data format processing method provided by the embodiment of the present application;

FIG. 10 is a schematic diagram of intermediate logic processing of a data format processing method provided by an embodiment of the present application;

FIG. 11 is a schematic diagram of a composition of a data format processing device provided by an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. In this application, "at least one" means one or more, and "multiple" means two or more. "And/or" describes the association relationship of associated objects, indicating that there may be three types of relationships, for example, A and/or B, which can mean: A exists alone, A and B exist at the same time, and B exists alone, where A, B can be singular or plural. The character "/" generally indicates that the contextual objects are an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one item (piece) of a, b or c can represent: a, b, c, a and b, a and c, b and c, or, a and b and c, wherein a, b and c can be single or multiple. In addition, in order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, words such as "first" and "second" are used to distinguish the same or similar items with basically the same function and effect, Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and execution order. For example, "first" in the first computing unit and "second" in the second computing unit in the embodiment of the present application are only used to distinguish different computing units. The first, second, etc. descriptions that appear in the embodiments of this application are only for illustration and to distinguish the description objects, and there is no order, nor does it represent a special limitation on the number of devices in the embodiments of this application, and cannot constitute a limitation on the number of devices in this application. Any limitations of the examples.

It should be noted that, in this application, words such as "exemplary" or "for example" are used as examples, illustrations or illustrations. Any embodiment or design described herein as "exemplary" or "for example" is not to be construed as preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "such as" is intended to present related concepts in a concrete manner.

First, the meanings of various data formats involved in this application are explained.

1. NCHW

Among them, N represents the number of batches, C represents the channel channel, H represents the height, and W represents the width. The arrangement order of the NCHW format is [batch, channels, height, width], which means that the arrangement order of the data is in order of W dimension, H dimension, C dimension, and N dimension. For example, taking one input picture and the data of the picture as shown in Figure 3 as an example, if the NCHW format is used to store, then the shape is (1, 3, 10, 10). (a) in Figure 4 is the data arrangement in NCHW format. As shown in (a) in Figure 4, the first element is 000, and the second element is along the W direction, that is, 001, so that Follow the W direction until 009, then follow the H direction, that is, 010, 011, 012...019...090... until 099, then follow the C direction, 100, 101...109, 110, 111...119... 200, 201... until 299. If N is an integer greater than 1, then continue to arrange sequentially along the N direction.

2. NHWC

The arrangement order of NHWC is [batch, height, width, channels], which means that the arrangement order of the data is arranged according to the C dimension, W dimension, H dimension, and N dimension. For example, taking one input picture and the data of the picture as shown in Figure 3 as an example, if the NHWC format is used to store, then the shape is (1, 10, 10, 3). (b) in Figure 4 is the data arrangement in the NHWC format. As shown in (b) in Figure 4, the first element is 000, and the second element is along the C direction, that is, 100, so in turn Follow the C direction to 200, then follow the W direction, that is, 001, 101, 201...009... until 209, and follow the H direction, 010, 110, 210, 011, 111, 211... until 299. If N is an integer greater than 1, then continue to arrange sequentially along the N direction.

3. NC1HWC0

The NC1HWC0 format is to split the C axis, which can be split into C1 and C0. In the case where N is 1, the data is arranged in the order of the C dimension, W dimension, H dimension of the first C0, the C dimension, W dimension, H dimension of the second C0... until the C dimension, W dimension of the C1th C0 Dimension and H dimension are arranged in sequence. If N is an integer greater than 1, then continue to arrange sequentially along the N direction.

Optionally, when using the NC1HWC0 format, if the number of C-axis is not a multiple of C0, you need to first expand the C-axis dimension to a multiple of C0 with a value of 0, and then split the C-axis into two dimensions, C1 and C0, and convert into NC1HWC0 format. If the number of C axes is a multiple of C0, no expansion is required, and the C axes can be split into two dimensions, C1 and C0.

For example, take the input picture as an example, and the data of the picture is the data shown in Figure 3. If the NC1HWC0 format is used to store, then the shape is (1, 1, 10, 10, 16), where C0 is 16. Since the C axis of the picture is 3, it is necessary to expand the C dimension to a multiple of 16 with a value of 0. Then split the C dimension into two dimensions, C1 and C0, and finally transpose it into NC1HWC0 format. (c) in Figure 4 is the data arrangement in NC1HWC0 format. As shown in (c) in Figure 4, the first element is 000, and the second element is along the C direction of the first C0. That is 100, so follow the C direction to 200, 0, 0, 0, 0...0, and then follow the W direction, that is, 001, 101, 201, 0...0, 009, 109, 209, 0. After ..0, along the H direction, 010, 110, 210, 0...0, 011, 111, 211... until the last 0. If N is an integer greater than 1, then continue to arrange sequentially along the N direction.

It should be noted that the value of C0 in the NC1HWC0 format is related to hardware, and the value of C0 may be different for different chips. The embodiment of this application does not limit the value of C0. In the embodiment of this application and the drawings, only C0 16 is used as an example for illustration.

Optionally, the NHWC format can be converted to the NC1HWC0 format through format conversion processing, that is, the initial C axis is split into C1 and C0 axes, and then the order of the axes is adjusted. For example, the initial shape is (32, 114, 114, 32), and after the format is adjusted to NC1HWC0, the shape is adjusted to (32, 2, 114, 114, 16).

Optionally, the NCHW format can be converted to the NC1HWC0 format through format conversion processing, that is, the initial C axis is split into C1 and C0 axes, and then the order of the axes is adjusted. For example, the initial shape is (32, 32, 114, 114), after being adjusted to NC1HWC0, the shape is adjusted to (32, 2, 114, 114, 16).

In AI chips, data is usually required to be arranged in a specific format. Since different operations require different data formats, it may be necessary to convert data in different formats.

For example, in the computing network shown in Figure 1, the sequence of operators in the computing network includes Conv1 operator, Relu1 operator, Conv2 operator, and Relu2 operator in turn, where the Conv operator requires the format of input data to follow the NC1HWC0 format Arrangement, the Relu operator requires the format of the input data to be arranged in the ND format. Since the data format of the original image is NHWC format, which is different from the format of the input data required by the Conv1 operator, the format of the input data needs to be converted. It should be noted that ND indicates that operations are performed in the original format. Optionally, the original format may be NHWC or NCHW. That is, the ND format can be NHWC format or NCHW format.

For example, as shown in Figure 2, a schematic diagram of data format conversion processing, by adding format conversion processing logic at the start position and end position of the convolution Conv operator, the format conversion process of increasing the start position in the Conv operator The logic can convert the format of the input data from the NHWC format to the NC1HWC0 format, and the format conversion processing logic added to the end position in the Conv operator can convert the format of the output data from the NC1HWC0 format to the NHWC format. However, this method increases the time consumption of the operator and reduces the performance of the operator after the format conversion processing logic is added inside the operator. If there are multiple Conv operators in the network, each Conv operator has conversion logic inside, and the network performance overhead is very high.

For another example, as shown in Figure 5, another schematic diagram of data format conversion processing, by inserting a format conversion operator (for example, TransData) before and after each Conv operator, the format conversion operator inserted before the Conv operator The operator can convert the format of the original input data from NHWC format to NC1HWC0 format, and the format conversion operator inserted after the Conv operator can convert the format of the output data of the Conv operator from NC1HWC0 format to NHWC format. However, the format conversion operators inserted by this method will bring additional performance overhead. If in a large network with 20 layers of this small network structure, the number of format conversion operators will increase by 20 times, and the performance overhead will be large, reducing network performance.

It can be understood that the difference between the data format conversion processing method shown in Figure 2 and the data format conversion processing method shown in Figure 5 is that the method shown in Figure 2 performs format conversion inside the operator, while the method shown in Figure 5 The method is to perform format conversion outside the operator. Both of these two data format conversion processing methods have the problem of large network performance overhead when performing format conversion processing.

In order to alleviate the problem of high network performance overhead due to excessive format conversion processing, the embodiment of the present application provides a data format processing method, which can reduce format conversion processing and improve network performance.

Exemplarily, the data format processing method provided in the embodiment of the present application may be applied to the system architecture shown in FIG. 6 . As shown in Figure 6, the system architecture includes a graph engine (Graph Engine, GE), a fusion engine (Fusion Engine, FE) and an operator library (Operator, Op).

Among them, GE and FE call the operator in Op. Op provides operator prototype library, operator information library, and operator implementation library. The operator prototype library provides the definition of the operator, such as the input, output, and attributes of the operator. The operator information library provides the capabilities implemented by the operator, such as the data types supported by the operator, and the supported input data format (format) and output data format. The operator implementation library provides operator implementation, and FE compiles and generates operators by calling the operator implementation library.

FE is used to handle hardware-related graph optimizations. FE includes an Op Judge module (Op Judge), a format selection module (Format Selector) and an operator information library loading module (Op Store). Among them, the Op judgment module in FE is responsible for adjusting the operator, so that the operator is adapted to the hardware and can run well on the corresponding chip. For example, when the initial data format is different from the data format supported by the operator, the Op judgment module in FE can perform format conversion processing on the initial data format, so that the converted data format is the data format supported by the operator. The operator information library loading module loads the operator information library during initialization to obtain the capabilities and supported data formats of each operator. The format selector module (Format Selector) is responsible for selecting which data format among multiple data formats to adopt when an operator supports multiple data formats. For example, when an operator supports multiple data formats, the format selection module may select the same data format as the output data format of the previous operator of the operator among the multiple data formats, so as to maintain the continuity of the data format.

The operator prototype library is loaded when GE is initialized, and the operator information library is loaded when FE is initialized. GE calls the derivation logic in the operator prototype library to derive the initial shape and initial data format of each operator node in the network graph, and GE calls FE to optimize the network graph (graph) related to hardware.

Exemplarily, the above-mentioned operator information base can provide the data format supported by the operator in two ways. The first way is to classify operators, and configure the preset pattern type that the operator belongs to through op.pattern. The data formats supported by different preset pattern types can be different. The second is to configure the input data format supported by the operator through input0.format.

Optionally, the preset mode types of the operator may include: format-independent type, reduction type and broadcast type.

Among them, the format-independent class can support any data format, and is suitable for adjusting pure ElementWise operators with single input and single output. For example, operators of the format-independent class can include Sqrt operator, Square operator, Sin operator, Neg operator, etc. operator. Operators of the format-independent class support adjustment between arbitrary formats. For example, the formats supported by format-independent operators include NCHW, NHWC, NC1HWC0, FRACTAL_Z, FRACTAL_Z_3D, FRACTAL_NZ, C1HWNCoC0 and other formats. The embodiment of the present application does not limit the type of data format supported by the format-independent operator, and this is only an example.

The data formats supported by Reduce operators are related to constraints. Reduce operators can include operators such as ReduceSum, ReduceMax, and ReduceMin. For example, for a Reduce operator, in the scene where the Reduce axis includes the C axis, the input data format is NC1HWC0, and the output data format is ND. In the scenario where the reduce axis does not include the C axis, the input data format NC1HWC0 and the output data format NC1HWC0 are supported. Reduce operators support input data in ND format and output data in ND format in any scenario.

For example, the input initial Shape is (32, 32, 114, 114), the initial format is NCHW, and the initial Reduce axis attribute is [1,], after the format is adjusted, the input format is NC1HWC0 format, and the adjusted Shape is (32 , 2, 114, 114, 16), the adjusted Reduce axis attribute is [1, 4].

The data formats supported by Broadcast operators are related to constraints. Broadcast operators include operators such as Add, RealDiv, and Minimum. For example, for a Broadcast operator, if the Broadcast axis does not contain C, the supported input data format is NC1HWC0, and the output data format is also NC1HWC0. If the axis of Broadcast includes the scene of C axis, the input data format supported by the Broadcast class operator is ND, and the output data format is also ND.

Optionally, when the operator information base configures the preset pattern type of the operator through op.pattern, the specific configuration method is as follows:

[Square]

input0.name=x

input0.dtype=float16, float, int32

output0.name=y

output0.dtype = float16, float, int32

op.pattern=formatAgnostic

Optionally, when the operator information base configures the data format supported by the operator through input0.format, the specific configuration method is as follows:

[Square]

input0.name=x

input0.dtype=float16, float, int32

input0.format=ND,NC1HWC0

output0.name=y

output0.dtype = float16, float, int32

output0.format=ND,NC1HWC0

Optionally, when the operator does not belong to any of the preset mode types, the data format supported by the operator can be configured through input0.format. However, when the operator supports multiple data formats, each data format supported by the operator needs to be configured through input0.format.

Understandably, the operator information base classifies operators and uses op.pattern to configure the preset pattern type of the operator. According to the preset pattern type, the data format supported by the operator can be determined. This configuration method is compared with the input0.format configures the data format supported by the operator, which can greatly reduce the amount of information in the operator information database, and does not need to configure each format supported by the operator. Moreover, configuring the preset mode type of the operator through the operator information library can enable the implementation of the operator without paying attention to the specific format, and can directly perform operations according to the input shape and attributes of the operator, reducing the burden of format processing in operator development. .

An embodiment of the present application provides a data format processing method, the method is applied to a computing network, the computing network includes a first computing unit and a second computing unit in sequence, as shown in Figure 7, the data format processing method may include The following steps:

S701. Acquire data formats supported by the first computing unit and the second computing unit.

It can be understood that in step S701 above, the processor may call the interface of the operator information library loading module to obtain the data formats supported by the first computing unit and the second computing unit.

Optionally, the first calculation unit and the second calculation unit in this embodiment of the application may be operators of the same type, or operators of different types. This embodiment of the application does not limit the first calculation unit and the second calculation unit. The specific operator type of the computing unit.

Exemplarily, the data format processing method provided by the embodiment of the present application can be applied to the computing network shown in FIG. son. The embodiment of the present application does not limit the specific structure of the operator sequence in the computing network, and only the operator sequence shown in FIG. 1 is taken as an example for illustration.

Optionally, the above-mentioned first computing unit may be the first computing unit in the computing network, or may be a computing unit in an intermediate position in the computing network. For example, when the first computing unit is the first computing unit in the computing network, the first computing unit can be the Conv1 operator in Figure 1, and the second computing unit can be the Relu1 operator in Figure 1, the Conv1 operator The input data of is the original input data of the computing network. For another example, when the first computing unit is a computing unit in the middle of the computing network, the first computing unit can be the Conv2 operator in Figure 1, and the second computing unit can be the Relu2 operator in Figure 1, the Conv2 The input data of the operator is the output data of the Relu1 operator. It can be understood that the embodiment of the present application does not limit the specific position of the first computing unit in the computing network, and the first computing unit is a computing unit before the second computing unit.

Exemplarily, obtaining the data formats supported by the first computing unit and the second computing unit in the above step S701 may include: obtaining the preset mode types of the first computing unit and the second computing unit; The preset mode type of the computing unit determines the data format supported by the first computing unit and the second computing unit. The preset mode types may include format-independent class, reduce-reduce class and broadcast-broadcast class. Optionally, the FE may acquire the preset mode types of the first computing unit and the second computing unit from the operator information library. Optionally, when the FE acquires the data formats supported by the first computing unit and the second computing unit, it may determine the data formats supported by the first computing unit and the second computing unit according to the preset mode type and specific constraints.

It is understandable that configuring the preset mode type of an operator through the operator information base can make the implementation of the operator do not need to pay attention to the specific format, and can directly perform operations according to the input shape and attributes of the operator, reducing the format of the operator development processing burden.

Optionally, when the first calculation unit in the operator information base configures the data format supported by the operator through input0.format, the FE can directly obtain the data format supported by the first calculation unit from the operator information base.

Optionally, the data format supported by the computing unit in this embodiment of the present application may be the input data format supported by the computing unit.

S702. If it is determined that the first computing unit does not support the format of the input data of the first computing unit, add a first format converting unit before the first computing unit.

It can be understood that in the above step S702, the processor may determine whether the first computing unit supports the format of the input data of the first computing unit, and when the processor determines that the first computing unit does not support the format of the input data of the first computing unit, A first format conversion unit is added before the first calculation unit. Optionally, if the processor determines that the first computing unit supports the format of the input data of the first computing unit, the processor determines not to perform format conversion processing before the first computing unit.

The first format converting unit is used for converting the format of the input data of the first computing unit into a data format supported by the first computing unit.

Exemplarily, in the case that the first calculation unit supports a data format, if the data format supported by the first calculation unit is different from the format of the input data of the first calculation unit, the processor determines that the first calculation unit does not support the first calculation The format of the cell's input data.

Exemplarily, in the case that the first computing unit supports multiple data formats, if the multiple data formats supported by the first computing unit do not include the format of the input data of the first computing unit, the processor determines that the first computing unit does not support The format of the input data for the first computing unit.

In the case that the first calculation unit does not support the format of the input data of the first calculation unit, by adding the first format conversion unit before the first calculation unit, the format of the input data of the first calculation unit can be converted into the first calculation unit The data formats supported by the cell.

In an implementation manner, if the first computing unit is the first computing unit in the computing network, then the input data of the first computing unit is the original input data of the computing network. That is, the format of the input data of the first computing unit is the format of the original input data of the computing network.

Optionally, the original input data of the above computing network may be the initial input data of the network. For example, the original input data of the computing network may be real picture data, voice data or text data.

Exemplarily, when the first computing unit supports a data format, and the data format supported by the first computing unit is different from the original input data format of the computing network, the processor determines that the first computing unit does not support the original input data of the computing network The format of the data. When the first computing unit supports multiple data formats, and the multiple data formats supported by the first computing unit do not include the format of the original input data of the computing network, the processor determines that the first computing unit does not support the original input data of the computing network format.

If the processor determines that the first computing unit does not support the format of the original input data of the computing network, the processor adds a first format conversion unit before the first computing unit to convert the format of the original input data of the computing network into the first computing unit Supported data formats.

For example, as shown in Figure 8, taking the first computing unit as the Conv1 operator, the second computing unit as the Relu1 operator, and the format of the original input data of the computing network as NHWC as an example, since the data format supported by the Conv operator is NC1HWC0 , that is, the data format supported by the Conv1 operator is different from the format NHWC of the original input data (the Conv1 operator does not support the format of the original input data), so a format conversion operator is inserted before the Conv1 operator, which is used to convert The original input data format NHWC is converted to the data format NC1HWC0 supported by the Conv1 operator.

In another implementation manner, if the first computing unit is a computing unit located in the middle of the computing network, then the input data of the first computing unit is the output data of the computing unit before the first computing unit. That is, the format of the input data of the first computing unit is the format of the output data of the computing units before the first computing unit.

Exemplarily, it is taken as an example that the sequence of operators in the computing network includes a third computing unit, a first computing unit, and a second computing unit in sequence, that is, the computing unit before the first computing unit is the third computing unit. When the first computing unit supports a data format, and the data format supported by the first computing unit is different from the output data format of the third computing unit, the processor determines that the first computing unit does not support the output data format of the third computing unit Format. When the first computing unit supports multiple data formats, and the multiple data formats supported by the first computing unit do not include the format of the output data of the third computing unit, the processor determines that the first computing unit does not support the format of the third computing unit The format of the output data.

If the processor determines that the first computing unit does not support the format of the output data of the third computing unit, the processor adds a first format conversion unit between the third computing unit and the first computing unit, so that the output data of the third computing unit The format is converted into a data format supported by the first calculation unit.

S703. If it is determined that the second computing unit supports the output data format of the first computing unit, determine not to perform format conversion processing between the first computing unit and the second computing unit.

It can be understood that in the above step S703, the processor may determine whether the second computing unit supports the format of the output data of the first computing unit, and when the processor determines that the second computing unit supports the format of the output data of the first computing unit, process The processor determines not to perform format conversion processing between the first computing unit and the second computing unit. Optionally, if the processor determines that the second computing unit does not support the format of the output data of the first computing unit, the processor adds a format conversion processing unit between the first computing unit and the second computing unit, so that the first computing unit The format of the output data of the unit is converted into a data format supported by the second computing unit.

Exemplarily, in the case that the second computing unit supports a data format, if the data format supported by the second computing unit is the same as the output data format of the first computing unit, the processor determines that the second computing unit supports the first computing unit The format of the output data.

Exemplarily, in the case that the second computing unit supports multiple data formats, if the multiple data formats supported by the second computing unit include the output data format of the first computing unit, the processor determines that the second computing unit supports the first The format of the output data of the compute unit.

In case the second computing unit supports the format of the output data of the first computing unit, the processor determines not to perform format conversion processing between the first computing unit and the second computing unit. It can be understood that, in the case that the second computing unit supports the format of the output data of the first computing unit, the second computing unit may directly use the output data of the first computing unit as the input data of the second computing unit, and For processing, it is not necessary to perform format conversion processing on the output data of the first computing unit, thereby reducing format processing in the computing network and improving network performance. That is to say, in the embodiment of the present application, when considering whether to perform data format conversion processing, the continuity of the data formats supported by multiple operators is considered, so that the latter operator can support the output data of the previous operator. format, it is determined not to perform format conversion processing between these two operators, thereby reducing the number of format processing in the computing network and improving network performance.

For example, as shown in Figure 8, taking the first computing unit as the Conv1 operator and the second computing unit as the Relu1 operator as an example, if the mode type of the Relu operator configured in the operator information base is a format-independent class (for example, op.pattern=formatAgnostic), the processor determines that the Relu1 operator supports any data format, that is, the Relu1 operator supports the output data format of the Conv1 operator, then the processor determines not to perform format conversion between the Conv1 operator and the Relu1 operator , that is, the format conversion operator is not inserted between the Conv1 operator and the Relu1 operator.

For another example, as shown in Figure 8, taking the first computing unit as the Relu1 operator and the second computing unit as the Conv2 operator as an example, since the data format supported by the Conv operator is NC1HWC0, the format of the output data of the Relu1 operator is also For NC1HWC0, the processor determines that the Conv2 operator supports the format of the output data of the Relu1 operator, then the processor determines not to perform format conversion between the Relu1 operator and the Conv2 operator, that is, not to insert between the Relu1 operator and the Conv2 operator Format conversion operator.

For another example, as shown in Figure 8, taking the first computing unit as the Conv2 operator and the second computing unit as the Relu2 operator as an example, if the mode type of the Relu operator configured in the operator information base is a format-independent class (such as , op.pattern=formatAgnostic), the processor determines that the Relu2 operator supports any data format, that is, the Relu2 operator supports the output data format of the Conv2 operator, then the processor determines not to perform format conversion between the Conv2 operator and the Relu2 operator Processing, that is, do not insert a format conversion operator between the Conv2 operator and the Relu2 operator.

According to the solution of the embodiment of this application, for two consecutive operators, if the latter operator supports the format of the output data of the previous operator, then the processor determines not to perform format conversion between the two operators, if If the latter operator does not support the format of the output data of the former operator, then the processor determines to perform format conversion processing between these two operators (that is, insert a format conversion operator between these two operators), so that Convert the output data format of the previous operator to the data format supported by the latter operator.

It can be understood that in the data format processing method provided by the embodiment of the present application, when the first computing unit does not support the format of the input data of the first computing unit, the first format conversion unit is added before the first computing unit, and the second When the computing unit supports the format of the output data of the first computing unit, format conversion processing is not performed between the first computing unit and the second computing unit, thus reducing the number of format conversion units in the network and reducing the format processing in the network , which improves network performance.

It should be noted that, in the embodiment of the present application, when performing format processing, when the second operator does not support the format of the output data of the first operator, insert between the first operator and the second operator Data conversion unit, when the third operator supports the output data format of the second operator, no data conversion unit will be inserted between the second operator and the third operator, when the fourth operator When a child supports the format of the output data of the third operator, no data conversion unit will be inserted between the third operator and the fourth operator until the Nth operator does not support the N-1th operator The data conversion unit will be inserted between the N-1th operator and the Nth operator only when the format of the output data of the operator. Therefore, this application considers the continuity of the data format supported by multiple operators, and Compared with the prior art, format conversion processing can be greatly reduced and network performance can be improved.

(Optional) S704. If it is determined that the format of the output data of the last computing unit of the computing network is different from the format of the original output data of the computing network, add a second format converting unit after the last computing unit of the computing network.

It can be understood that in the above step S704, the processor may determine whether the format of the output data of the last computing unit of the computing network is the same as the format of the original output data, and determine the format of the output data of the last computing unit of the computing network by the processor When the format of the original output data is different, a second format conversion unit is added after the last calculation unit of the calculation network.

Optionally, the original output data of the above computing network is the original output data processed by the network. The format of the original output data of the computing network is a data format not adapted to the hardware. For example, in order to adapt to the hardware, the Conv operator needs to convert the format of the input data to the NC1HWC0 format. The NC1HWC0 format is a data format that adapts to the hardware, while the original output data of the computing network is a data format that is not adapted to the hardware. For example, The original output data of the computing network can be in NCHW format or NHWC format.

Optionally, the format of the original output data of the computing network may be the same as that of the original input data of the computing network, or may be different from the format of the original input data of the computing network, which is not limited in this application.

The above-mentioned second format conversion unit is used to convert the format of the output data of the last computing unit into the format of the original output data of the computing network.

Exemplarily, when the format of the output data of the last operator of the computing network is different from the format of the original output data of the computing network, in order to ensure that the format of the output data of the computing network is consistent with the format of the original output data, you can A format conversion operator is inserted after a calculation unit to convert the format of the output data of the last operator to the format of the original output data.

As shown in Figure 9, taking the original output data format of the computing network as NHWC and the Relu2 operator as the last operator of the computing network as an example, since the output data format of the Relu2 operator is NC1HWC0, it is different from the original output data format different, therefore, a format conversion operator (the second format conversion unit) is added after the Relu2 operator to convert the format of the output data of the Relu2 operator into the NHWC format to ensure that the format of the output data of the computing network is consistent with that of the original output data The format is the same.

It can be understood that in the data format processing method provided by the embodiment of the present application, when the first computing unit does not support the format of the input data of the first computing unit, the first format conversion unit is added before the first computing unit, and the second When the computing unit supports the format of the output data of the first computing unit, format conversion processing is not performed between the first computing unit and the second computing unit, thus reducing the number of format conversion units in the network and reducing the format processing in the network , which improves network performance. Moreover, when the format of the output data of the last computing unit of the computing network is different from the format of the original output data, the processor converts the format of the output data of the last computing unit to keep the format of the output data of the computing network consistent with the original The format of the output data is consistent.

Next, the intermediate logic of the above solution in the embodiment of the present application is explained.

Exemplarily, taking the computing network shown in FIG. 1 as an example, the sequence of operators in the computing network sequentially includes a Conv1 operator, a Relu1 operator, a Conv2 operator, and a Relu2 operator.

As shown in Figure 1, for the Conv1 operator, since the format of the original input data is NHWC, the data format supported by the Conv1 operator is NC1HWC0, so the Conv1 operator does not support the format of the original input data, so before the Conv1 operator and Conv1 Insert the format conversion operator after the operator respectively. As shown in (a) in Figure 10, the format conversion operator before the Conv1 operator is used to convert the format NHWC of the original input data into the data format NC1HWC0 supported by the Conv1 operator, and the format conversion operator after the Conv1 operator uses It is used to convert the format NC1HWC0 of the output data of the Conv1 operator to NHWC.

As shown in (a) in Figure 10, for the Relu1 operator, the mode configured in the operator information base is a format-independent class, that is, the Relu1 operator supports any data format, and can choose ) are processed in the same data format, so format conversion operators are inserted before and after the Relu1 operator respectively. As shown in (b) in Figure 10, the format conversion operator before the Relu1 operator is used to convert the original input data format NHWC to NC1HWC0, and the format conversion operator after the Relu1 operator is used to convert the output of the Relu1 operator The format of the data is NC1HWC0 converted to NHWC.

As shown in (b) in Figure 10, since the format conversion operator after the Conv1 operator is logically inverse to the format conversion operator before the Relu1 operator, it can be eliminated, and (c) in Figure 10 can be obtained after elimination Computational network shown. It can be understood that, by exposing the hardware format to the FE level, this application finds an opportunity to eliminate the format conversion at the FE level, reduces the format conversion process, and improves the network performance.

As shown in (c) in Figure 10, for the Conv2 operator, since the input data format of the Conv2 operator is NHWC, and the data format supported by the Conv2 operator is NC1HWC0, the Conv2 operator does not support the input data of the Conv2 operator format, so the format conversion operator is inserted before the Conv2 operator and after the Conv2 operator. As shown in (d) in Figure 10, the format conversion operator before the Conv2 operator is used to convert the format NHWC of the original input data into the data format NC1HWC0 supported by the Conv2 operator, and the format conversion operator after the Conv2 operator is used It is used to convert the format NC1HWC0 of the output data of the Conv2 operator to NHWC. It can be understood that the input data of the Conv2 operator in (c) in Figure 10 is the output data of the Relu1 operator after the format conversion operator converts the data, so the format of the input data of the Conv2 operator is the Relu1 operator The output data of the child is converted into the data format after the format conversion operator.

As shown in (d) in Figure 10, since the logic of the format conversion operator after the Relu1 operator and the format conversion operator before the Conv2 operator are reciprocal, they can be eliminated, and (e) in Figure 10 can be obtained after elimination Computational network shown.

As shown in (e) in Figure 10, for the Relu2 operator, since the mode configured by the Relu operator in the operator information base is a format-independent class, that is, the Relu2 operator supports any data format, and it can be selected to be the same as the previous operator ( That is, the Conv2 operator) is processed in the same data format, so the format conversion operator is inserted before the Relu2 operator and after the Relu2 operator. As shown in (f) in Figure 10, the format conversion operator before the Relu2 operator is used to convert the format NHWC of the original input data to NC1HWC0, and the format conversion operator after the Relu2 operator is used to convert the output of the Relu2 operator The format of the data is NC1HWC0 converted to NHWC.

As shown in (f) in Figure 10, since the format conversion operator after the Conv2 operator is logically inverse to the format conversion operator before the Relu2 operator, it can be eliminated, and (g) in Figure 10 can be obtained after elimination Computational network shown.

It can be understood that the computing network shown in (f) in FIG. 10 is the computing network shown in FIG. 9 obtained by adopting the solution of the embodiment of the present application. Compared with the computing network shown in Figure 2, this computing network does not perform format conversion inside the operator, but performs format conversion at the FE layer, so it does not need to be like the solution shown in Figure 2, each Conv operator internal Transformation logic is required. Compared with the computing network shown in Figure 5, the computing network in this application embodiment does not need to insert format conversion operators before and after each Conv operator, only when the latter operator does not support the previous operator. Only when the format of the output data is specified, a format conversion operator is inserted for format conversion processing. Therefore, the solution of the embodiment of the present application reduces the number of format conversion processing in the network and improves network performance.

It should be noted that this application exposes the hardware format to the FE level through the configuration of the operator information base, finds the opportunity to eliminate the format conversion at the FE level, reduces the format conversion process, and improves the network performance.

The data format processing method provided by the embodiment of the present application considers the continuity of the data formats supported by multiple operators when determining whether to insert a format conversion operator. If the latter operator supports the output data of the previous operator format, then make sure not to insert a format conversion operator between these two operators, if the latter operator does not support the format of the output data of the previous operator, then insert a format conversion operator between these two operators , thereby reducing the amount of data conversion processing and improving network performance.

The foregoing mainly introduces the solutions provided by the embodiments of the present invention from the perspective of method steps. It can be understood that, in order to realize the above functions, the computer includes hardware structures and/or software modules corresponding to each function. Those skilled in the art should easily realize that the present application can be realized in a combined form of hardware and computer software in combination with the units and algorithm steps of each example described in the embodiments disclosed herein. Skilled artisans may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present invention.

The embodiment of the present application may divide the computer into functional modules according to the above method example, for example, each functional module may be divided corresponding to each function, or two or more functions may be integrated into one processing module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. It should be noted that the division of modules in the embodiment of the present invention is schematic, and is only a logical function division, and there may be another division manner in actual implementation.

Fig. 11 shows a schematic composition diagram of a data format processing apparatus 1100, and the data format processing apparatus 1100 can be used to implement the methods and functions involved in any of the above-mentioned embodiments.

The data format processing apparatus 1100 includes: a processing unit 1101 and an obtaining unit 1102 . Exemplarily, the processing unit 1101 is configured to control and manage the actions of the above-mentioned data format processing device, and is used to execute the processing performed by the processor in the above-mentioned embodiment. Optionally, if the data format processing device 1100 includes a storage unit, then The processing unit 1101 may also execute programs or instructions stored in the memory, so that the data format processing apparatus 1100 implements the methods and functions involved in any of the foregoing embodiments.

Exemplarily, the above-mentioned processing unit 1101 may be used to execute steps S702-S704 in FIG. 7, and/or other processes for the technologies described herein. The obtaining unit 1102 may be configured to perform, for example, obtaining data formats supported by the first computing unit and the second computing unit, and/or other processes for the techniques described herein. Wherein, all relevant content of each step involved in the above method embodiment can be referred to the function description of the corresponding function module, and will not be repeated here.

Exemplarily, in terms of hardware implementation, a processor may execute the function of the processing unit 1101, and a communication interface may execute the function of the acquisition unit 1102, for example, the interface of the operator information library loading module may acquire the first computing unit and The data format supported by the second computing unit. Wherein, the processing unit 1101 can be embedded in or independent of the processor of the data format processing device 1100 in the form of hardware, and can also be stored in the memory of the data format processing device 1100 in the form of software, so that the processor can call and execute the above functions The operation corresponding to the unit.

The embodiment of the present application also provides a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a computer, the computer can execute the data format processing method shown in FIG. 7 .

The embodiment of the present application also provides a computer program product, which causes the computer to execute the data format processing method shown in FIG. 7 when the computer program product is run on the computer.

The embodiment of the present application also provides an electronic device, the electronic device includes a processor and a memory, the memory is coupled to the processor; the memory is used to store computer program code; the computer program code includes computer instructions, When the processor executes the above computer instructions, the electronic device is made to execute the data format processing method shown in FIG. 7 .

The embodiment of the present application also provides an electronic device, the electronic device includes a central processing unit (Central Process Unit, CPU) and a neural network processor (Neural-network Processing Unit, NPU), and the CPU is used to execute the A data format processing method, the NPU is used to execute the computing network obtained by the CPU based on the data format processing method.

The steps of the methods or algorithms described in connection with the disclosure of this application can be implemented in the form of hardware, or can be implemented in the form of a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (Random Access Memory, RAM), flash memory, erasable programmable read-only memory (Erasable Programmable ROM, EPROM), electrically erasable Programmable read-only memory (Electrically EPROM, EEPROM), registers, hard disk, removable hard disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be a component of the processor. The processor and storage medium can be located in the ASIC. In addition, the ASIC may be located in the core network interface device. Certainly, the processor and the storage medium may also exist in the core network interface device as discrete components.

Those skilled in the art should be aware that, in the above one or more examples, the functions described in the present invention may be implemented by hardware, software, firmware or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer-readable storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a general purpose or special purpose computer.

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention and are not intended to limit the scope of the present invention. Protection scope, any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the protection scope of the present invention.

Claims (15)

1. A data format processing method, applied to a computing network, where the computing network sequentially includes a first computing unit and a second computing unit, the method comprising:

   Acquiring data formats supported by the first computing unit and the second computing unit;

   if the first computing unit is determined not to support the format of the input data of the first computing unit, adding a first format conversion unit before the first computing unit, wherein the first format conversion unit is used for converting the format of the input data of the first computing unit into a data format supported by the first computing unit;

   if the second computing unit is determined to support the format of the output data of the first computing unit, determining that format conversion processing is not performed between the first computing unit and the second computing unit.
2. The method of claim 1, wherein, in the case where the first computing unit is a first computing unit in the computing network, the format of the input data of the first computing unit is the format of the original input data of the computing network.
3. The method according to claim 1, wherein in case the first computing unit is a computing unit located at an intermediate position in the computing network, the format of the input data of the first computing unit is the format of the output data of a computing unit preceding the first computing unit.
4. A method according to any of claims 1-3, wherein, in case the second computing unit supports a plurality of data formats, determining the format in which the second computing unit supports the output data of the first computing unit comprises:

   and if the plurality of data formats supported by the second computing unit comprise the format of the output data of the first computing unit, determining that the second computing unit supports the format of the output data of the first computing unit.
5. The method of any of claims 1-4, wherein the obtaining the data formats supported by the first computing unit and the second computing unit comprises:

   acquiring preset mode types of the first computing unit and the second computing unit; the preset mode types comprise a format independent class, a reduced reduction class and a Broadcast class;

   and determining the data formats supported by the first computing unit and the second computing unit according to the preset mode types of the first computing unit and the second computing unit.
6. The method according to any one of claims 1-5, further comprising:

   if the format of the output data of the last computing unit of the computing network is different from the format of the original output data of the computing network, adding a second format conversion unit after the last computing unit of the computing network, wherein the second format conversion unit is used for converting the format of the output data of the last computing unit into the format of the original output data of the computing network.
7. A data format processing apparatus, applied to a computing network, the computing network comprising, in order, a first computing unit and a second computing unit, the apparatus comprising: a processing unit and an acquisition unit;

   the acquisition unit is used for acquiring the data formats supported by the first calculation unit and the second calculation unit;

   the processing unit is configured to add a first format conversion unit before the first computing unit if it is determined that the first computing unit does not support the format of the input data of the first computing unit, where the first format conversion unit is configured to convert the format of the input data of the first computing unit into a data format supported by the first computing unit;

   the processing unit is further configured to determine that format conversion processing is not performed between the first computing unit and the second computing unit if it is determined that the second computing unit supports the format of the output data of the first computing unit.
8. The apparatus of claim 7, wherein, in the case where the first computing unit is a first computing unit in the computing network, the format of the input data of the first computing unit is the format of the original input data of the computing network.
9. The apparatus of claim 7, wherein, in the case where the first computing unit is a computing unit located at an intermediate position in the computing network, the format of input data of the first computing unit is the format of output data of a computing unit preceding the first computing unit.
10. The apparatus according to any of claims 7-9, wherein the processing unit is configured to determine that the second computing unit supports the format of the output data of the first computing unit if it is determined that the plurality of data formats supported by the second computing unit includes the format of the output data of the first computing unit.
11. The device according to any one of claims 7 to 10, wherein,

    the acquisition unit is further used for acquiring preset mode types of the first calculation unit and the second calculation unit; the preset mode types comprise a format independent class, a reduced reduction class and a Broadcast class;

    the processing unit is further configured to determine a data format supported by the first computing unit and the second computing unit according to the preset mode types of the first computing unit and the second computing unit acquired by the acquiring unit.
12. The apparatus according to any one of claims 7-11, wherein the processing unit is further configured to:

    if the format of the output data of the last computing unit of the computing network is different from the format of the original output data of the computing network, adding a second format conversion unit after the last computing unit of the computing network, wherein the second format conversion unit is used for converting the format of the output data of the last computing unit into the format of the original output data of the computing network.
13. An electronic device comprising a processor and a memory, the memory coupled to the processor; the memory is used for storing computer program codes; the computer program code comprising computer instructions which, when executed by the processor, cause the electronic device to perform the method of any of claims 1-6.
14. A computer readable storage medium comprising computer instructions which, when run on an electronic device, cause the electronic device to perform the method of any of claims 1-6.
15. A computer program product, characterized in that the computer program product, when run on a computer, causes the computer to perform the method according to any of claims 1-6.

Images