CN117251134A - Data processing methods, devices, storage media and electronic equipment for neural networks - Google Patents

Abstract

The disclosure provides a data processing method and device of a neural network, a storage medium and electronic equipment, and relates to the technical field of computers. The method comprises the following steps: acquiring a plurality of groups of data to be processed, which are input to a current operation unit in a neural network, wherein the data to be processed comprise an order and a mantissa; extracting common order factors of the multiple groups of data to be processed, and multiplying mantissas in each group of data to be processed by the residual orders after the common order factors are extracted to obtain data to be accumulated corresponding to each group of data to be processed; and accumulating the data to be accumulated corresponding to each group of data to be processed, and multiplying the accumulated result by the common order factor to obtain the output data of the current operation unit. The method and the device are beneficial to improving the data processing efficiency of the neural network and reducing the hardware power consumption.

Description

Data processing methods, devices, storage media and electronic equipment for neural networks

Technical field

The present disclosure relates to the field of computer technology, and in particular, to a neural network data processing method, a neural network data processing device, a computer-readable storage medium, and an electronic device.

Background technique

In algorithms based on neural networks, data can be divided into floating-point data and fixed-point data, and the core computing unit is the multiply-accumulate operation. However, in related technologies, floating-point data or fixed-point data need to be processed based on different neural networks, which reduces the versatility of the data processing method of the neural network; when performing an accumulation operation, the data to be accumulated with a large bit width is directly accumulated. , making the data processing method of the neural network slow in operation, increasing the hardware power consumption, thereby reducing the data processing efficiency of the neural network.

It should be noted that the information disclosed in the above background section is only used to enhance understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art.

Contents of the invention

The present disclosure provides a data processing method of a neural network, a data processing device of a neural network, a computer-readable storage medium and an electronic device, thereby improving the problem of low data processing efficiency of the neural network at least to a certain extent.

Additional features and advantages of the disclosure will be apparent from the following detailed description, or, in part, may be learned by practice of the disclosure.

According to a first aspect of the present disclosure, a data processing method for a neural network is provided, including: acquiring multiple sets of data to be processed input to the current operation unit in the neural network, where the data to be processed includes an order and a mantissa; extracting the Describe the common order factors of multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor to obtain the accumulated order corresponding to each group of data to be processed. Data; accumulate the data to be accumulated corresponding to each group of data to be processed, and multiply the accumulation result by the common order factor to obtain the output data of the current operation unit.

According to a second aspect of the present disclosure, a data processing device for a neural network is provided, including: a data acquisition module to be processed, configured to acquire multiple sets of data to be processed input to a current computing unit in the neural network, the data to be processed being The data includes an order and a mantissa; a common order factor extraction module is configured to extract the common order factor of the multiple groups of data to be processed, and extract the common order factor from the mantissa in each group of data to be processed. Multiply the remaining orders to obtain the data to be accumulated corresponding to each group of data to be processed; the output data acquisition module is configured to accumulate the data to be accumulated corresponding to each group of data to be processed, and compare the accumulated results with the above The common order factors are multiplied together to obtain the output data of the current operation unit.

According to a third aspect of the present disclosure, there is provided a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the data processing method of the neural network of the above-mentioned first aspect and its possible implementation are implemented. Way.

According to a fourth aspect of the present disclosure, an electronic device is provided, including: a processor; and a memory for storing executable instructions of the processor. Wherein, the processor is configured to execute the data processing method of the neural network of the first aspect and its possible implementation by executing the executable instructions.

The technical solution of the present disclosure has the following beneficial effects:

In the data processing process of the neural network, multiple sets of input data to be processed are obtained, the common order factors of the data to be processed are extracted, and the mantissa in each set of data to be processed is compared with the remaining order after extracting the common order factor. Multiply to obtain the data to be accumulated, then accumulate the data to be accumulated, and multiply the accumulation result by the common order factor to obtain the output data of the current operation unit. On the one hand, based on existing fixed-point multipliers and adders, common order factors are extracted for each set of data to be processed, so that the present disclosure can not only handle floating-point operations, but also fixed-point operations, improving the data processing of neural networks. The method is versatile; on the other hand, the mantissa in each set of data to be processed is multiplied by the remaining order after extracting the common order factor to obtain the data to be accumulated, and then the data to be accumulated is accumulated, which reduces the time required The bit width of the accumulated data and the accumulated operation improves the operation speed, further improves the data processing efficiency of the neural network and reduces the hardware power consumption.

It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and do not limit the present disclosure.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description, serve to explain the principles of the disclosure. Obviously, the drawings in the following description are only some embodiments of the present disclosure. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 shows the system architecture of the operating environment of this exemplary embodiment;

Figure 2 shows a flow chart of a data processing method of a neural network in this exemplary embodiment;

Figure 3 shows a schematic diagram of three groups of data to be processed performing multiplication and accumulation operations in this exemplary embodiment;

Figure 4 shows a schematic diagram of the process of obtaining the order code of the common order factor in this exemplary embodiment;

Figure 5 shows a schematic diagram of the process of accumulating the data to be accumulated after extracting the common order factors in this exemplary embodiment;

Figure 6 shows a process flow chart of the accumulation operation in a data processing method of a neural network in this exemplary embodiment;

Figure 7 shows a flow chart of another data processing method of a neural network in this exemplary implementation;

Figure 8 shows a schematic diagram of the data processing process of a convolutional neural network when the data to be processed is image data in this exemplary implementation;

Figure 9 shows a schematic diagram of the operation process of the accumulation operation in the data processing of a convolutional neural network when the data to be processed is image data in this exemplary implementation;

Figure 10 shows a schematic structural diagram of a data processing device of a neural network in this exemplary embodiment;

FIG. 11 shows a schematic diagram of an electronic device in this exemplary embodiment.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concepts of the example embodiments. To those skilled in the art. The described features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to provide a thorough understanding of embodiments of the disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of the specific details being omitted, or other methods, components, devices, steps, etc. may be adopted. In other instances, well-known technical solutions have not been shown or described in detail to avoid obscuring aspects of the disclosure.

Furthermore, the drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the figures are functional entities and do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and/or processor devices and/or microcontroller devices.

In computer systems, data can be expressed in two ways: fixed-point and floating-point; if each data is expressed in fixed-point, the decimal point position of the data is fixed and the value range of the data is limited, and it can be used to represent a decimal or an integer; such as Can be used to represent monetary data in fixed-point mode. For example, 88.00 or 00.88 can be used to represent a monetary value with four digits of precision and two decimal places.

Since the fixed decimal point position determines the integer part and the decimal part of the fixed number of digits, using fixed-point method to represent data results in the form of data being too rigid, which is not conducive to expressing particularly large numbers or small numbers at the same time. If fixed-point method is used to represent two Data with very different sizes requires a long machine word length, resulting in low utilization of the data storage unit. Therefore, in practical applications, most modern computers use floating point to express data. Floating point uses scientific notation to represent real numbers. Data expressed in floating point can include order and mantissa, base and sign bits, such as The decimal fixed-point data 123.45 can be expressed in floating point mode as (-1)°×1.2345×10 ² , where (-1)° can be the sign bit, 1.2345 can be the mantissa, 10 can be the base, and 2 can be the exponent code. 10 ² can be an order, and floating point numbers can achieve the effect of floating decimal points through exponents, thereby controlling the numerical range of the data. Floating point data can represent a larger range of data and has higher calculation accuracy, but at the same time, the cost of hardware implementation is high.

In related technologies, the typical hardware structure of an artificial intelligence computing engine applied to neural networks may include a convolution computing engine, a vector operation engine, a storage unit, an overall scheduling unit and a partial and accumulation unit. Among them, the core operation in the convolution calculation engine and vector operation engine is multiplication and accumulation.

Taking the multiplication and accumulation operation process of floating point data A, B, C, D, E, F and other floating point numbers as an example, any floating point data can be expressed in the following form:

A＝2 ^ea ×ma (1)

Among them, ea represents the exponent code, 2 represents the base, and ma is the mantissa. Therefore, the operation process of the multiply-accumulate operation unit is as follows:

Before the data is calculated, all input operands in the multiply-accumulate operation are first converted into floating point representation format. Then during the data transmission process, only the order and mantissa of each data need to be transmitted. When performing an accumulation operation, accumulating each multiplication result is equivalent to first calculating A×B+C×D, and then adding the result of E×F.

It can be seen from formula (2) that each multiplication operation can include the order part and the mantissa part. Taking A×B as an example, the multiplication operation can include the order part 2 ^ea+eb and the mantissa part (ma×mb), Among them, ma×mb is equivalent to a fixed-point multiplication operation, which can be completed by using a multiplier in hardware implementation; 2 ^ea+eb × (ma×mb) is equivalent to a shift operation on the result of ma×mb. In hardware implementation Implementation can be accomplished using a shifter; however, this may result in a very large bit width of A×B=2 ^ea+eb ×(ma×mb). For example, in the floating point number A=2 ^ea ×ma, The bit width of ea and ma is 8 bits in total, of which the bit width of ea is 4 bits, and the data range that can be represented is 0 to 15. When performing the multiplication operation A×B, the data range of ea+eb obtained is 0 to 30. , at this time, the bit width to be used becomes 5bit.

In the above process, the bit width of the two data orders increases by 1 bit after one addition operation. In formula (2), in addition to the addition of orders, the data also needs to be accumulated. Not only does the data accumulate The bit width becomes larger during operation, and as the number of accumulations increases, the bit width requirements for the input data of the adder are also getting larger and larger. The increase in the amount of operations slows down the operation speed, thereby increasing the hardware power consumption and reducing the cost. Data processing efficiency.

In view of the above problems, exemplary embodiments of the present disclosure provide a data processing method of a neural network. The system architecture of the running environment of this exemplary embodiment will be described below with reference to Figure 1 .

Referring to FIG. 1 , the system architecture 100 may include a terminal device 110 and a server 120 . The terminal device 110 may be an electronic device such as a smartphone, a tablet computer, or a laptop computer. The server 120 generally refers to a backend system that provides services related to data processing of the neural network in this exemplary embodiment. For example, it may be a server that implements data processing of the neural network. The server 120 may be one server or a cluster formed by multiple servers, which is not limited in this disclosure. The terminal device 110 and the server 120 may be connected through a wired or wireless communication link for data exchange.

In one embodiment, the neural network can be deployed on the terminal device 110, and the terminal device 110 executes the data processing method of the neural network in this exemplary embodiment.

In one implementation, the neural network can also be deployed on the server 120. After the terminal device 110 sends the data to be processed to the server 120, the server 120 executes the data processing method of the neural network in this exemplary implementation.

The data processing method of the neural network is explained below with reference to Figure 2. Figure 2 shows an exemplary flow of the data processing method of the neural network, including the following steps S210 to S230:

Step S210: Obtain multiple sets of data to be processed that are input to the current computing unit in the neural network. The data to be processed includes orders and mantissas;

Step S220: Extract the common order factors of multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor to obtain the order factors corresponding to each group of data to be processed. Accumulate data;

Step S230: Accumulate the data to be accumulated corresponding to each group of data to be processed, and multiply the accumulation result by the common order factor to obtain the output data of the current operation unit.

Based on the above method, on the one hand, based on existing fixed-point multipliers and adders, common order factors are extracted for each set of data to be processed, so that the present disclosure can not only handle floating-point operations, but also fixed-point operations, improving neural The versatility of the network's data processing method; on the other hand, multiply the mantissa in each set of data to be processed by the remaining order after extracting the common order factor to obtain the data to be accumulated, and then accumulate the data to be accumulated, The bit width of the data to be accumulated and the accumulation operation is reduced, the operation speed is improved, the data processing efficiency of the neural network is further improved, and the hardware power consumption is reduced.

Each step in Figure 2 is described in detail below.

Referring to FIG. 2 , in step S210 , multiple sets of data to be processed that are input to the current computing unit in the neural network are obtained, and the data to be processed include orders and mantissas.

Among them, the neural network can be an algorithmic mathematical model that imitates the behavioral characteristics of biological neural networks to perform distributed parallel information processing. The neural network can rely on the complexity of the system by adjusting the interconnection relationships between a large number of internal nodes, thereby achieving For the purpose of processing information, biological neural networks usually mainly refer to the neural networks of the human brain. For example, the neural network may include a convolutional neural network, which may be a type of feedforward neural network that includes convolutional calculations and has a deep structure that can perform feature learning.

The data to be processed can be the current computing unit input to the neural network, the data that needs to be calculated, it can be any type of data, such as image data, text data or audio data, or the above-mentioned data after being processed by a part of the intermediate layer of the neural network. data. The data to be processed can include the order and the mantissa; at the hardware level, the data to be processed can be represented by shifting the number of bits represented by the order of the mantissa. For example, the data A to be processed can be expressed as A=2 ^ea ×ma , where 2 ^ea can be the order of the data A to be processed, ma can be the mantissa of the data A to be processed, then at the hardware level, the data A to be processed can be represented by shifting ma to the left by ea bits. This disclosure does not impose any special restrictions on the specific content of the data to be processed.

Multiple sets of data to be processed can be input into the current operation unit for operation. The current operation unit can perform multiplication and accumulation operations. The multiplication and accumulation operation unit can contain multiple accumulators and multipliers, among which the multiplier completes the multiplication operation of data; the accumulator Responsible for adding the products obtained by multiplication operations. When performing an accumulation operation, in addition to the result of each multiplier in the current module, the input also needs to add the output data of the previous module of the current module.

In one implementation, each group of data to be processed in the multiple groups of data to be processed in the present disclosure can be used to perform multiplication operations to obtain each group of data to be processed, and data to be processed between different groups can be used to Perform an accumulation operation; for example, as shown in Figure 3, where A, B, C, D, E, and F are six data to be processed. There are three sets of data in Figure 3: A and B are one set, and C and D are One group, E and F are one group; each group of data to be processed can be used to perform multiplication operations, that is, A×B, C×D, E×F, to obtain each group of data to be processed; different groups of data to be processed Accumulation operations can be performed, that is, A×B, C×D, and E×F can be added to obtain A×B+C×D+E×F, plus the output data of the upper-level operation unit to obtain The output data of the current computing unit.

In one implementation, the above-mentioned data to be processed may include output data of an upper-level computing unit of the current computing unit.

Among them, the upper-level operation unit can be an operation unit that completes the multiplication-accumulation operation, and can output the operation result of the multiplication-accumulation operation, that is, the output data of the upper-level operation unit. The output data of the upper-level operation unit can be expressed as the order and Regarding the form of mantissa multiplication, this disclosure does not place any special restrictions on the operation mode of the upper-level operation unit and the specific form of the output data of the upper-level operation unit.

In step S220, extract the common order factors of multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor to obtain the corresponding order of each group of data to be processed. The data to be accumulated;

Wherein, each group of data to be processed can include two data to be processed, each group of data to be processed can be used to perform multiplication operations, and data to be processed between different groups can be used to perform accumulation operations; in one embodiment , each group of data to be processed may also include the output data of the upper-level computing unit of the current computing unit. This disclosure does not specifically limit the specific content of each group of data to be processed.

In this exemplary embodiment, each set of data to be processed can be regarded as a whole, and each set of data to be processed can be expressed in the form of the total order multiplied by the total mantissa. For example, a set of data to be processed is A×B , where A=2 ^ea ×ma, B=2 ^eb ×mb, then A×B can be expressed as A×B=2 ^ea ×ma×2 ^eb ×mb=2 ^ea+eb ×(ma×mb), in In this set of data to be processed, the total order is 2 ^ea+eb and the total mantissa is (ma×mb).

Among them, the total order can be the product of different orders in each group of data to be processed, and the total mantissa can be the product of different mantissas in each group of data to be processed. It should be noted that if a certain set of data to be processed only includes one piece of data, the order of the data may be the total order of the set of data to be processed, and the mantissa of the data may be the total mantissa of the set of data to be processed.

The common order factor can be the common factor between the total orders of each set of data to be processed. For example, there are two sets of data to be processed, A×B and C×D, where A×B=2 ^ea+eb ×(ma ×mb), C×D=2 ^ec+ed ×(mc×md), then the common order factor can be the total order of a set of data to be processed 2 ^ea+eb or 2 ^ec+ed ; in one implementation , the common order factor can also be a preset order value, and the preset order value can be an order value less than the minimum value of the total order; the public order factor can also be the output data of the upper-level computing unit The order; this disclosure does not place special restrictions on the method of obtaining the public order factor and the total order.

In one implementation, the above-mentioned extraction of common order factors of multiple sets of data to be processed may include:

Determine the total order of each set of data to be processed, and compare the total orders to obtain the common order factor.

For example, there are two groups of data to be processed: the first group of data to be processed A and B, and the second group of data to be processed C and D. Then it can be determined that the total order of the first group of data to be processed is 2 ^{ea + eb} , and the order of the first group of data to be processed is determined to be 2 ea + eb . The total order of the two sets of data to be processed is 2 ^ec+ed . Compare the total order 2 ^ea+eb and 2 ^ec+ed . When the base of the total order is the same, you can compare the order code ea of the total order. +eb and ec+ed to compare the total order to obtain the common order factor.

In one implementation, the above-mentioned comparison of total orders to obtain a common order factor may include the following steps:

Determine the common order factor based on the minimum value of the total order.

In one implementation, if the current computing unit does not have an upper-level computing unit, the data to be processed only includes data processed by the current computing unit, and the total order may include the total order of each group of data to be processed; if the current When the arithmetic unit has an upper-level arithmetic unit, the data to be processed may include the output data of the upper-level arithmetic unit of the current arithmetic unit. When determining the common order factor, the total order may include the upper-level arithmetic operation of the current arithmetic unit. The order 2 ^emin ' of the output data of the unit (that is, the common order factor of the upper-level operation unit of the current operation unit) and the total order of each other group of data to be processed can be determined based on the minimum value of the total order. Order factor 2 ^emin . In addition, when there is only one set of data to be processed, the total order of the set of data to be processed can be directly used as the common order factor.

For example, determine the common order factor in two groups of data to be processed: the first group of data to be processed A and B, the second group of data to be processed C and D, the total order of the first group of data to be processed is 2 ^{ea + eb} , the total order of the second group of data to be processed is 2 ^ec+ed . Compare 2 ^ea+eb and 2 ^ec+ed . Since the base of the total order is 2, you can compare the order code ea of the total order. +eb and ec+ed are compared, and the minimum value emin among the order codes ea+eb and ec+ed is selected as the order code of the common order factor, and the common order factor 2 ^emin can be obtained.

Determining the common order factor based on the minimum value of the total order can reduce the bit width of the data to be accumulated to the greatest extent; since the minimum value of the total order is the greatest common factor of each group of data to be processed, the total order The minimum value of the number is extracted as the common order factor, and the remaining order of the minimum bit width can be obtained.

In one implementation, the remaining order may include a quotient obtained by dividing the total order of each set of data to be processed by a common order factor.

For example, if the current computing unit is the first-level computing unit, the output results of the previous-level computing unit are not included in the data to be processed at this time. A and B, C and D, E and F, etc. can be calculated according to formula (3). Multiple groups of data to be processed are used to extract the common order factor 2 ^emin . In formula (3), the common order factor and the data to be accumulated can be included. The data to be accumulated can include the mantissa and remainder corresponding to each group of data to be processed. The operation result obtained by accumulating the products of orders.

For example, 2 ^(ea+eb)-emin ×K1, K1=(ma×mb), where the remaining order 2 ^(ea+eb)-emin can include the total order of the data to be processed 2 ^ea+eb divided by the common order Factor 2 ^emin to get the quotient.

In one implementation, the current operation unit may include a multiplier and an adder. The multiplier may be used to multiply the mantissa in each set of data to be processed by the remaining order after extracting the common order factor. For example, the multiplier may be used to The device multiplies the remaining order and mantissa of each group of data to be processed in formula (3).

Among them, emin is the minimum value among ea+eb, ec+ed, ee+ef,..., 2 ^emin can be used as a common order factor to extract the product of each set of data to be processed, K1=(ma×mb), K2=(mc×md), K3=(me×mf)…

In one implementation, the current operation unit may include an order processing unit and a mantissa processing unit, and the order and mantissa of each data to be processed may be sent to the order processing unit and the mantissa processing unit respectively for processing.

For example, if there are multiple data to be processed A, B, C, D, E, F and the output data Result _prev of the upper-level computing unit of the current computing unit, the data to be processed can be calculated according to formula (4); since each The data to be processed can include the order and the mantissa. Therefore, the order and mantissa of each data to be processed can be sent to the order processing unit and the mantissa processing unit respectively for processing.

Among them, emin can be the minimum value among ea+eb, ec+ed, ee+ef, and emin'. Emin can be used as a common order factor to extract the product of each set of data to be processed, K1=(ma×mb) , K2=(mc×md), K3=(me×mf), R' can be the mantissa of the output data of the upper-level operation unit of the current operation unit.

In one implementation, in the order processing unit, as shown in Figure 4, the order processing unit needs to receive the order code emin' of the common order factor of the upper-level operation unit of the current operation unit, and then convert each The order code of the data to be processed is sent to the adder to obtain the total order code of other groups of data to be processed. At this time, the order code of the total order of each group of data to be processed can include ea+eb, ec+ed, ee+ef and emin', input all the total order code into the MIN module to obtain the minimum value of the total order code, so that the current operation can be obtained based on the minimum value of the total order code The order emin of the common order factor of the unit.

In one implementation, if a certain set of data to be processed is the output data of the upper-level operation unit, the total order of the set of data to be processed may be the common order factor of the upper-level operation unit.

In one implementation, if the data to be processed includes the output data of the upper-level operation unit, then the total order of each group of data to be processed can be compared with the common order factor of the upper-level operation unit. Get the common order factor.

In one implementation, in the mantissa processing unit, as shown in Figure 5, the mantissa of the data to be processed can be input to the multiplier to obtain the mantissas K1, K2 and K3 corresponding to each group of data to be processed, and receive The mantissa R' of the upper-level computing unit of the current computing unit; then input ea+eb, ec+ed, ee+ef, emin' and emin into the subtractor to calculate ea+eb-emin, ec+ed-emin, ee +ef-emin, the value of emin'-emin; then, K1, K2, K3 and R' can be performed according to the value of ea+eb-emin, ec+ed-emin, ee+ef-emin, emin'-emin Shift operation, and add the shifted results to obtain the mantissa R of the output result of the current operation unit. Finally, the order codes emin and R of the common order factor in the output result of the current operation unit can be output to The next-level computing unit of the current computing unit continues to process the data.

In one embodiment, the present disclosure extracts common order factors for each set of data to be processed based on fixed-point multipliers and fixed-point adders, so that this solution can handle both floating-point operations and fixed-point operations, improving neural processing. The versatility of the network's data processing method; in addition, the remaining order after extracting the common order factor of each group of data to be processed is multiplied by the mantissa to obtain the data to be accumulated, which reduces the bit width of the addition operation. For example, in the data to be processed A and B with a bit width of 16 bits, the order codes ea and eb can use 5 bits, and the mantissas ma and mb can use 11 bits; then the data range that ea and eb can represent is 0 to 31. , if ea and eb are added directly, the data range of ea+eb obtained can be 0 to 62. At this time, the bit width to be used becomes 6 bits, and as the number of additions increases, the bit width to be used will be getting bigger and bigger; and this disclosure extracts the common order factor when calculating A×B, and when performing the order addition operation, by subtracting the order code emin of the common order factor, ea+eb-emin The data range is still 0 to 31, and the bit width that can still be used for the operation result after the addition operation is 5 bits.

Continuing to refer to Figure 2, in step S230, the data to be accumulated corresponding to each group of data to be processed is accumulated, and the accumulation result is multiplied by the common order factor to obtain the output data of the current operation unit.

The data to be accumulated may include the mantissa corresponding to each group of data to be processed and the remaining order after extracting the common order factor. For example, in this disclosure, for a group of data to be processed A=2 ^ea ×ma and B=2 ^eb ×mb is processed, and the common order factor of the group of data to be processed is 2 ^emin , then the data to be accumulated corresponding to the group of data to be processed can be expressed as 2 ^{(ea +eb)-emin} × (ma×mb), this disclosure There are no special restrictions on the specific content of the accumulated data.

In one implementation, accumulating data to be accumulated corresponding to each group of data to be processed, and multiplying the accumulation result by a common order factor to obtain the output data of the current operation unit may include:

The adder is used to accumulate the data to be accumulated corresponding to each group of data to be processed, and the multiplier is used to multiply the accumulated result by the common order factor to obtain the output data of the current operation unit.

In one implementation, the output data of the current operation unit may be the product of the common order factor and the accumulation result. The output data of the current operation unit may be expressed as the multiplication of the order and the mantissa. This disclosure is useful for the current The specific form of the output data of the arithmetic unit is not particularly limited.

In one implementation, the output data of the current computing unit can be input to the next-level computing unit to continue data processing.

In one implementation, the current operation unit may also include a carry-preserving adder, and the adder may be used to accumulate data to be accumulated corresponding to each group of data to be processed, which may include the following steps:

When the number of data items to be accumulated is less than or equal to the preset number of items, the data to be accumulated is added through the adder to obtain the accumulation result;

When the number of items of data to be accumulated is greater than the preset number of items, a carry-preserving adder and an adder are combined to add the data to be accumulated to obtain the accumulation result.

A carry-preserving adder may be an adder used to sum a large number of operands. Input three source operands and output two operation results. Through CSA, the addition of three numbers can be compressed into the addition of two numbers. When adding multiple data, the carry can be retained and only one carry transfer is performed. .

The number of preset items may be two items, three items, or other items. This disclosure does not place a special limit on the specific number of the preset items.

In the calculation process of each CSA, three operands a ₁ , a ₂ , a ₃ can be input, and two output data b ₁ and b ₂ can be obtained according to the following formula (5) and formula (6):

b ₁ ＝a ₁ ∧a ₂ ∧a ₃ (5)

b ₂ =((a ₁ &a ₂ )|(a ₂ &a ₃ )|(a ₁ &a ₃ ))<<1 (6)

In one implementation, if there are only two items of data to be accumulated, the data to be processed can be added directly through an adder; if there are more than two items of data to be accumulated, a Carry Save Adder (CSA) can be used ) compresses the accumulation operation. When the data to be accumulated is compressed to two items, the adder is then used to add the two items to obtain the accumulation result.

For example, as shown in Figure 6, a carry-preserving adder and an adder can be combined to sum the accumulated data a ₀ to a ₅ . Since CSA can input three data and output two data, you can select three data from the above six data to be accumulated and input them into the first CSA, and then input the remaining three data into the second CSA; then input the third data into the second CSA. The two outputs of one CSA and one output of the second CSA are used as input data into the third CSA to obtain two output data; the two output data of the third CSA and the remaining data of the second CSA are An output data of is used as input data and is input into the fourth CSA to obtain two output data b ₀ and b ₁ . At this time, the six data to be accumulated a ₀ ~ a ₅ are compressed into two through multiple CSAs. Output data b ₀ and b ₁ , and then input b ₀ and b ₁ to the adder for addition operation to obtain the accumulation result of the data to be accumulated.

It can be seen from formulas (5) and (6) that the CSA operation logic is relatively simple. Therefore, using CSA to compress the accumulated data before adding it effectively improves the operation speed and reduces the power consumption caused by the accumulation operation.

In one implementation, FIG. 7 shows an exemplary flow of the neural network data processing method of the present disclosure. Referring to FIG. 7 , data processing can be performed based on the neural network according to steps S701 to S709.

Step S701, input multiple sets of data to be processed into the neural network, where the data to be processed includes order and mantissa;

Step S702, obtain the total order of each group of data to be processed;

Step S703, use the minimum value of the total order of each group of data to be processed as the common order factor;

Step S704, extract common order factors from each group of data to be processed;

Step S705, use the mantissa of each group of data to be processed and the remaining order after extracting the common order factor as the data to be accumulated;

Step S706: Determine whether the current number of data items to be accumulated is two items. If so, jump to step S707; otherwise, jump to step S708;

Step S707, directly add the two items of data to be accumulated through an adder to obtain the accumulation result;

Step S708, combine the carry-preserving adder and the adder to perform an addition operation on the data to be accumulated, and the accumulation result has been obtained;

Step S709: Multiply the accumulated result by the common order factor to obtain the output data of the current operation unit.

In one embodiment, the neural network may include a convolutional neural network, and the data to be processed may include image block data and convolution kernels covered by the convolution kernel of the convolutional neural network in the image data.

For example, the present disclosure can be combined with a convolutional neural network for image data processing. As shown in Figure 8, the image data to be processed can be first input into the convolutional neural network. The convolutional neural network includes a convolution kernel, and can be processed according to the image data. The image block covered by the medium convolution kernel obtains the target image block; the data to be processed can include the pixel data A, B, E, F of the image block and the convolution kernel W, X, Y, Z, which can be expressed as the order and mantissa. In the form of multiplication; the pixel data of the target image block can be multiplied by the operator corresponding to the convolution kernel, and then the products can be added to perform the convolution operation. Specifically, the data to be processed can be convolved according to the following formula Product operation:

Among them, 2 ^emin is the minimum value of the total order of each group of data to be processed; each group of data to be processed can be multiplied, and different groups of data to be processed can be accumulated; K1=(ma×mw), K2 =(mb×mx), K3=(me×my), K4=(mf×mz).

During the convolution operation between the pixel data of the target image block and the convolution kernel, there are four groups of data to be processed, namely A and W, B and X, E and Y, and F and Z. Each can be first The data to be processed is expressed in the form of multiplying the order and the mantissa to perform a convolution operation; and then based on the order code of the total order of each group of data to be processed, ea+ew, eb+ex, ee+ey, ef+ez The minimum value of the order code emin of the common order factor is obtained, and then the common order factor 2 ^emin is obtained; the common factor can be extracted for each set of data to be processed according to the common order factor to obtain the data to be accumulated; the data to be accumulated It can include the mantissas K1, K2, K3, K4 corresponding to each group of data to be processed and the remaining order after extracting the common order factor; since the number of items to be accumulated is four, as shown in Figure 9, it can be combined with CSA After compressing the data to be accumulated into two items, the compressed two items are then input to the adder for addition operation to obtain the accumulation result; finally, the current target image block and volume can be obtained based on the product of the common order factor and the accumulation result. The convolution result of the convolution kernel is then output to the convolution operation of the next target image block and the convolution kernel to obtain the convolution result of the next target image block and the convolution kernel.

In addition, exemplary embodiments of the present disclosure also provide a data processing apparatus for a neural network. Referring to Figure 10, the data processing device 1000 of the neural network may include:

The data to be processed acquisition module 1010 is configured to acquire multiple sets of data to be processed input to the current computing unit in the neural network, where the data to be processed includes orders and mantissas;

The common order factor extraction module 1020 is configured to extract the common order factors of multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor to obtain each A set of data to be accumulated corresponding to the data to be processed;

The output data acquisition module 1030 is configured to accumulate the data to be accumulated corresponding to each group of data to be processed, and multiply the accumulation result by the common order factor to obtain the output data of the current operation unit.

In one implementation, the above-mentioned neural network may include a convolutional neural network, and the data to be processed may include image block data and convolution kernels covered by the convolution kernel of the convolutional neural network in the image data.

In one implementation, the above-mentioned data to be processed may also include output data of an upper-level computing unit of the current computing unit.

In one implementation, the above-mentioned extraction of common order factors of multiple sets of data to be processed includes:

Determine the total order of each set of data to be processed, and compare the total orders to obtain the common order factor.

In one implementation, the above method compares the total orders to obtain the common order factors, including:

The common order factor is determined based on the minimum value of the total order.

In one implementation, the above-mentioned remaining order may include a quotient obtained by dividing the total order of each group of data to be processed by a common order factor.

In one implementation, the above-mentioned current operation unit may include a multiplier and an adder. The above-mentioned multiplication of the mantissa in each group of data to be processed by the remaining order after extracting the common order factor includes:

Use a multiplier to multiply the mantissa in each set of data to be processed by the remaining order after extracting the common order factor;

The above-mentioned accumulation of data to be accumulated corresponding to each group of data to be processed is performed, and the accumulated result is multiplied by the common order factor to obtain the output data of the current operation unit, including:

The data to be accumulated corresponding to each group of data to be processed is accumulated through the adder, and the accumulated result is multiplied by the common order factor through the multiplier to obtain the output data of the current operation unit.

In one implementation, the above-mentioned current operation unit may also include a carry-preserving adder, and the adder is used to accumulate data to be accumulated corresponding to each group of data to be processed, including:

When the number of data items to be accumulated is less than or equal to the preset number of items, the data to be accumulated is added through the adder to obtain the accumulation result;

When the number of items of data to be accumulated is greater than the preset number of items, a carry-preserving adder and an adder are combined to add the data to be accumulated to obtain the accumulation result.

The specific details of each part of the above-mentioned device have been described in detail in the method implementation, and will not be described again.

Exemplary embodiments of the present disclosure also provide a computer-readable storage medium, which can be implemented in the form of a program product, which includes program code. When the program product is run on an electronic device, the program code is used to cause the electronic device to The steps described in the "Exemplary Methods" section of this specification above according to various exemplary embodiments of the present disclosure are performed. In an alternative embodiment, the program product may be implemented as a portable compact disk read-only memory (CD-ROM) and include the program code, and may be run on an electronic device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

The Program Product may take the form of one or more readable media in any combination. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: electrical connection with one or more conductors, portable disk, hard disk, random access memory (RAM), read only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave carrying readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A readable signal medium may also be any readable medium other than a readable storage medium that can send, propagate, or transport the program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any suitable medium, including but not limited to wireless, wireline, optical cable, RF, etc., or any suitable combination of the foregoing.

Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., as well as conventional procedural programming. Language—such as "C" or a similar programming language. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server execute on. In situations involving remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device, such as provided by an Internet service. (business comes via Internet connection).

Exemplary embodiments of the present disclosure also provide an electronic device. The electronic device may include a processor and memory. The memory stores executable instructions of the processor, such as program codes. The processor performs the method in this exemplary embodiment by executing the executable instructions.

Referring to FIG. 11 , an electronic device will be exemplarily described in the form of a general-purpose computer. It should be understood that the electronic device 1100 shown in FIG. 11 is only an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

As shown in FIG. 11 , the electronic device 1100 may include: a processor 1110 , a memory 1120 , a bus 1130 , an I/O (input/output) interface 1140 , and a network adapter 1150 .

The processor 1110 may include one or more processing units. For example, the processor 1110 may include a central processing unit (CPU), an AP (Application Processor, an application processor), a modem processor, a display processor ( Display Process Unit (DPU), GPU (Graphics Processing Unit, graphics processor), ISP (Image Signal Processor, image signal processor), controller, encoder, decoder, DSP (Digital Signal Processor, digital signal processor) , baseband processor and/or NPU (Neural-NetworkProcessing Unit, neural network processor), etc. The data processing method of the neural network in this exemplary embodiment can be executed by a GPU or an NPU. In one embodiment, after obtaining the data to be processed, the neural network can be deployed in the GPU, and the data to be processed can be processed by the GPU. Extract the common order factor, and then use the remaining data after extracting the common order factor from the data to be processed as the data to be accumulated. Add the accumulated data through the CSA and adder to obtain the accumulation result. Add the common order factor and the accumulated data. The product of the results is used as the output data of the current operation unit

The memory 1120 may include volatile memory, such as RAM 1121, cache unit 1122, and may also include non-volatile memory, such as ROM 1123. Memory 1120 may also include one or more program modules 1124, such program modules 1124 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples. This may include the implementation of a network environment. For example, the program module 1124 may include each module in the device 1000 described above.

The bus 1130 is used to realize connections between different components of the electronic device 1100 and may include a data bus, an address bus and a control bus.

Electronic device 1100 may communicate with one or more external devices 1100 (eg, terminal device, keyboard, mouse, external controller, etc.) through I/O interface 1140.

The electronic device 1100 can communicate with one or more networks through the network adapter 1150. For example, the network adapter 1150 can provide mobile communication solutions such as 3G/4G/5G, or provide wireless communication solutions such as wireless LAN, Bluetooth, near field communication, etc. . Network adapter 1150 may communicate with other modules of electronic device 1100 via bus 1130 .

Although not shown in Figure 11, other hardware and/or software modules may also be provided in the electronic device 1100, including but not limited to: a display, microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, Tape drives and data backup storage systems, etc.

It should be noted that although several modules or units of equipment for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to exemplary embodiments of the present disclosure, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above may be further divided into being embodied by multiple modules or units.

Those skilled in the art will understand that various aspects of the present disclosure may be implemented as systems, methods, or program products. Therefore, various aspects of the present disclosure may be embodied in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which may be collectively referred to herein as "Circuits", "modules" or "systems". Other embodiments of the present disclosure will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the disclosure that follow the general principles of the disclosure and include common common sense or customary technical means in the technical field that are not disclosed in the disclosure. . It is intended that the specification and embodiments be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise structures described above and illustrated in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the disclosure is limited only by the appended claims.

Claims (10)

1. A neural network data processing method, characterized by including: Obtain multiple sets of data to be processed that are input to the current computing unit in the neural network, where the data to be processed includes orders and mantissas; Extract the common order factor of the multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor to obtain the corresponding order of each group of data to be processed. Data to be accumulated; The data to be accumulated corresponding to each group of data to be processed are accumulated, and the accumulation result is multiplied by the common order factor to obtain the output data of the current operation unit. 2. The method according to claim 1, wherein the neural network includes a convolutional neural network, and the data to be processed includes image block data covered by the convolution kernel of the convolutional neural network in the image data. and the convolution kernel. 3. The method according to claim 1, wherein the data to be processed includes output data of an upper-level computing unit of the current computing unit. 4. The method according to claim 1, characterized in that: extracting the common order factor of the multiple groups of data to be processed, and extracting the common order factor from the mantissa in each group of data to be processed. The remaining orders after are multiplied, including: Determine the total order of each set of data to be processed, and obtain the common order factor by comparing the total order; The remaining order includes the quotient obtained by dividing the total order of each group of data to be processed by the common order factor. 5. The method of claim 4, wherein the step of comparing the total orders to obtain the common order factor includes: The common order factor is determined based on the minimum value of the total order. 6. The method according to claim 1, characterized in that the current operation unit includes a multiplier and an adder; the mantissa in each group of data to be processed and the remainder after extracting the common order factor Multiplication of orders, including: Use the multiplier to multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor; The accumulated data corresponding to each group of data to be processed is accumulated, and the accumulated result is multiplied by the common order factor to obtain the output data of the current operation unit, including: The adder is used to accumulate the data to be accumulated corresponding to each group of data to be processed, and the multiplier is used to multiply the accumulation result by the common order factor to obtain the output data of the current operation unit. 7. The method according to claim 6, wherein the current operation unit further includes a carry-preserving adder, and the adder is used to accumulate data to be accumulated corresponding to each group of data to be processed, including: When the number of items of the data to be accumulated is less than or equal to the preset number of items, the data to be accumulated are added by an adder to obtain the accumulation result; When the number of items of the data to be accumulated is greater than the preset number of items, the data to be accumulated is added using a carry-preserving adder and an adder to obtain the accumulation result. 8. A neural network data processing device, characterized in that it includes: The data to be processed acquisition module is configured to acquire multiple sets of data to be processed input to the current computing unit in the neural network, where the data to be processed includes an order and a mantissa; A common order factor extraction module configured to extract the common order factors of the multiple groups of data to be processed, and multiply the mantissa in each group of data to be processed by the remaining order after extracting the common order factor. , obtain the data to be accumulated corresponding to each group of data to be processed; The output data acquisition module is configured to accumulate data to be accumulated corresponding to each group of data to be processed, and multiply the accumulation result by the common order factor to obtain the output data of the current operation unit. 9. A computer-readable storage medium with a computer program stored thereon, characterized in that when the computer program is executed by a processor, the method of any one of claims 1 to 7 is implemented. 10. An electronic device, characterized in that it includes: processor; memory for storing executable instructions for the processor; wherein the processor is configured to perform the method of any one of claims 1 to 7 via execution of the executable instructions.