(19)
(11)EP 3 654 209 A1

(12)EUROPEAN PATENT APPLICATION

(43)Date of publication:
20.05.2020 Bulletin 2020/21

(21)Application number: 19212365.1

(22)Date of filing:  31.08.2017
(51)Int. Cl.: 
G06F 17/16  (2006.01)
G06N 3/04  (2006.01)
G06N 3/063  (2006.01)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
MA MD

(62)Application number of the earlier application in accordance with Art. 76 EPC:
17923228.5 / 3605402

(71)Applicant: Cambricon Technologies Corporation Limited
Beijing 100190 (CN)

(72)Inventors:
  • LIU, Shaoli
    Beijing, 100190 (CN)
  • CHEN, Tianshi
    Beijing, 100190 (CN)
  • WANG, Bingrui
    Beijing, 100190 (CN)
  • ZHANG, Yao
    Beijing, 100190 (CN)

(74)Representative: Huang, Liwei 
Cäcilienstraße 12
40597 Düsseldorf
40597 Düsseldorf (DE)

 
Remarks:
This application was filed on 28-11-2019 as a divisional application to the application mentioned under INID code 62.
 


(54)CHIP DEVICE AND RELATED PRODUCTS


(57) A chip device and related products are provided. The chip device includes a main unit and a plurality of basic units in communication with the main unit. The main unit is configured to obtain a data block to-be-calculated and an operation instruction, divide the data block to-be-calculated into a distribution data block and a broadcast data block according to the operation instruction, obtain a plurality of basic data blocks by splitting the distribution data block, and distribute the plurality of basic data blocks to the basic units and broadcast the broadcast data block to the basic units. The basic units are configured to obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block, send the operation results to the main unit. The main unit is further configured to obtain an instruction result corresponding to the data block to-be-calculated and the operation instruction by processing the operation results. The technical solutions provided by the present disclosure have the advantages of short operation time and low energy consumption.




Description

TECHNICAL FIELD



[0001] The present disclosure relates to the field of communication and chip technology, and particularly to related to a chip device and related products.

BACKGROUND



[0002] Artificial neural networks (ANN) have been a research hotspot prompted since the 1980s. ANN abstracts human brain neural networks from the perspective of information processing to establish a simple model and form different networks according to different connection modes. It is also often referred to as neural networks or neural-like networks in engineering and academia. Neural networks are an operation model consisting of a larger number of interconnected nodes (or neurons). Operations of the existing neural networks are performed based on a central processing unit (CPU) or a graphic processing unit (GPU), and the operations have high power consumption and long operation time.

SUMMARY



[0003] The embodiments of the present disclosure provide a method for operations in a neural network and related products, which can reduce operation time and decrease power consumption of the module.

[0004] According to a first aspect of the present disclosure, there is provided a method for operations in a neural network, and the method is applicable to a chip device. The chip device includes a main unit and a plurality of basic units. The method includes the following. A data block to-be-calculated and an operation instruction is obtained by the main unit, and a distribution data block and a broadcast data block is obtained by dividing the data block to-be-calculated according to the operation instruction. The main unit obtains a plurality of basic data blocks by splitting the distribution data block, distributes the plurality of basic data blocks to a plurality of basic units, and broadcasts the broadcast data block to the plurality of basic units. The plurality of basic units obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block, and send the operation results to the main unit. The main unit obtains an instruction result corresponding to the data block to-be-calculated and the operation instruction by processing the operation results.

[0005] In one implementation, the main unit may broadcast the broadcast data block to the plurality of basic units as follows. The main unit may broadcast the broadcast data block to the plurality of basic units by one broadcast.

[0006] In one implementation, the plurality of basic units may obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block and send the operation results to the main unit as follows. The basic units may obtain inner-product processing results by performing inner-product operations for each of the basic data blocks and the broadcast data block, obtain the operation results by performing accumulation operation for each of the inner-product processing results respectively, and then send the operation results to the main unit.

[0007] In one implementation, when the operation results are the inner-product processing results, the main unit may obtain an instruction result corresponding to the data block to-be-calculated and the operation instruction by processing the operation results as follows. The main unit may obtain accumulated results by performing accumulation operation for each of the operation results respectively, and obtain the instruction result corresponding to the data blocks to-be-calculated and the operation instruction by arranging the accumulated results.

[0008] In one implementation, the main unit may broadcast the broadcast data block to the plurality of basic units as follows. The main unit may obtain a plurality of broadcast data sub-blocks by splitting the broadcast data block; and broadcast the plurality of broadcast data sub-blocks to the plurality of basic units by multiple broadcasts.

[0009] In one implementation, the plurality of basic units may obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block and send the operation results to the main unit as follows. The basic units may obtain inner-product processing results by performing one inner-product operation for each of the broadcast data sub-blocks and each of the basic data blocks, obtain operation sub-results by performing accumulation operation for each of the inner-product processing results, and send the operation sub-results to the main unit.

[0010] In one implementation, the basic units may obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block and send the operation results to the main unit as follows. The basic unit may obtain n processing sub- results, by multiplexing each of the broadcast data sub-blocks n times, to perform inner-product operation for the broadcast data sub-blocks and n basic data blocks, obtain n operation sub- results by performing accumulation operation for each of the n processing sub-results respectively, and send the n operation sub-results to the main unit, the n is an integer greater than or equal to 2.

[0011] According to a second aspect of the present disclosure, there is provided a chip device. The chip device includes a main unit and a plurality of basic units. The main unit is configured to obtain a data block to-be-calculated and an operation instruction, to divide the data block to-be-calculated into a distribution data block and a broadcast data block according to the operation instruction, to obtain a plurality of basic data blocks by splitting the distribution data block, to distribute the plurality of basic data blocks to a plurality of basic units, and to broadcast the broadcast data block to the plurality of basic units; the basic units are configured to obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block, and to send the operation results to the main unit; and the main unit is configured to obtain an instruction result corresponding to the data block to-be-calculated and the operation instruction by processing the operation results.

[0012] In one implementation, the chip device may further include a branch unit disposed between the main unit and at least one basic unit, and the branch unit may be configured to forward data.

[0013] In one implementation, the main unit is configured to broadcast the broadcast data block to the plurality of basic units by one broadcast.

[0014] In one implementation, the basic units are configured to obtain inner-product processing results by performing inner-product operations for each of the basic data blocks and the broadcast data block, to obtain the operation results by performing accumulation operation for each of the inner-product processing results respectively, and to send the operation results to the main unit.

[0015] In one implementation, when the operation results are the inner-product processing results, the main unit is configured to obtain accumulated results by performing accumulation operation for each of the operation results respectively, and to obtain the instruction result corresponding to the data blocks to-be-calculated and the operation instruction by arranging the accumulated results.

[0016] In one implementation, the main unit is configured to obtain a plurality of broadcast data sub-blocks by splitting the broadcast data block, and to broadcast the plurality of broadcast data sub-blocks to the plurality of basic units by multiple broadcasts.

[0017] In one implementation, the basic units are configured to obtain inner-product processing results by performing one inner-product operation for each of the broadcast data sub-blocks and each of the basic data blocks, to obtain operation sub-results by performing accumulation operation for each of the inner-product processing results; and to send the operation sub-results to the main unit.

[0018] In one implementation, the basic units are configured to obtain n processing sub-results by multiplexing each of the broadcast data sub-blocks n times to perform inner-product operation for the broadcast data sub-blocks and n basic data blocks; to obtain n operation sub-results by performing accumulation operation for each of the n processing sub-results respectively, and to send the n operation sub-results to the main unit, and the n is an integer greater than or equal to 2.

[0019] In one implementation, the main unit may include one or any combination of a main register and a main on-chip cache circuit; and a basic unit may include one or any combination of a basic register and a basic on-chip cache circuit.

[0020] In one implementation, the main unit may include one or any combination of a vector arithmetic unit circuit, an arithmetic logic unit circuit, an accumulator circuit, a matrix transposition circuit, a direct memory access circuit and a data rearrangement circuit.

[0021] In one implementation, the basic unit may include one or any combination of an inner-product arithmetic unit circuit and an accumulator circuit.

[0022] In one implementation, the main unit is coupled with a plurality of branch units respectively, and each of the branch units is coupled with at least one basic unit.

[0023] In one implementation, the main unit is coupled with a plurality of branch units mutually connected in series, and each of the branch units is respectively coupled with at least one basic unit.

[0024] In one implementation, the plurality of branch units are configured to forward data transmitted between the main unit and the at least one basic unit.

[0025] In one implementation, the branch units are configured to forward data transmitted between the main unit and the at least one basic unit or other branch units.

[0026] In one implementation, the data may include one or any combination of a vector, a matrix, a three-dimensional data block, a four-dimensional data block, and an n-dimensional data block.

[0027] In one implementation, the broadcast data block is configured as a multiplier data block and the distribution data block is configured as a multiplicand data block, when the operation instruction is a multiply instruction; and the broadcast data block is configured as an input data block and the distribution data block is configured as a convolution kernel, when the operation instruction is a convolution instruction.

[0028] According to a third aspect of the present disclosure, there is provided a method for applying the chip device provided by the second aspect, the chip device is configured to perform one or any combination of a matrix-multiplying-matrix operation, a matrix-multiplying-vector operation, a convolution operation, or a full connected operation.

[0029] According to a fourth aspect of the present disclosure, a chip is provided that integrates the chip device provided by the second aspect.

[0030] According to a fifth aspect of the present disclosure, an intelligent device is provided, the intelligent device may include the chip provided by the fourth aspect.

[0031] The embodiments of the present disclosure have the following beneficial effects. In the embodiments of the present disclosure, when receiving the data block to-be-calculated and the operation instruction, the data block to-be-calculated is divided into the distribution data block and the broadcast data block, the distribution data block is split into the plurality of basic data blocks, and then the plurality of basic data blocks are distributed to the plurality of basic units to perform inner-product operations. In this way, a largest amount of calculation in the inner product operations is distributed to the plurality of basic units for simultaneous execution, thereby reducing calculation time and saving power consumption.

BRIEF DESCRIPTION OF THE DRAWINGS



[0032] To describe technical solutions of the present disclosure more clearly, the following briefly introduces the accompanying drawings required for describing the embodiments. Apparently, the accompanying drawings in the following description only illustrate some embodiments of the present disclosure. Those of ordinary skill in the art may also obtain other drawings based on these accompanying drawings without creative efforts.

FIG. 1a is a schematic structural diagram of a chip device according to one embodiment of the present disclosure.

FIG. 1b is a schematic structural diagram of another chip device according to one embodiment of the present disclosure.

FIG. 1c is a schematic diagram illustrating data distribution of a chip device according to one embodiment of the present disclosure.

FIG. 1d is a schematic diagram illustrating data return of a chip device according to one embodiment of the present disclosure.

FIG. 2 is a flow chart illustrating a method for operations in a neural network according to one embodiment of the present disclosure.

FIG. 2a is a schematic diagram illustrating a multiplication between a matrix A and a matrix B according to one embodiment of the present disclosure.

FIG. 3 is a flow chart illustrating a method for operations in a neural network according to one embodiment of the present disclosure.

FIG. 3a is a schematic diagram illustrating data of a simple sample of a fully connection 1.

FIG. 3b is a schematic diagram illustrating data of multiple samples of a fully connection 2.

FIG. 3c is a schematic diagram illustrating data of M convolution kernels of a convolution 1.

FIG. 3d is a schematic diagram illustrating input data of a convolution 2.

FIG. 3e is a schematic diagram of an operation window of a three-dimension data block of input data.

FIG. 3f is a schematic diagram of another operation window of a three-dimension data block of input data.

FIG. 3g is schematic diagram of yet another operation window of a three-dimension data block of input data.


DETAILED DESCRIPTION OF ILLUSTRATED EMBODIMENTS



[0033] Technical solutions in the embodiments of the present disclosure will be described clearly and completely hereinafter with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are merely some rather than all embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0034] The terms such as "first", "second", "third", and "fourth" used in the specification, the claims, and the accompanying drawings of the present disclosure are used for distinguishing between different objects rather than describing a particular order. The terms "include" and "comprise" as well as variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, device, or apparatus including a series of steps or units is not limited to the listed steps or units, it can optionally include other steps or units that are not listed; alternatively, other steps or units inherent to the process, method, product, or device can be included either.

[0035] The term "embodiment" or "implementation" referred to herein means that a particular feature, structure, or feature described in conjunction with the embodiment may be contained in at least one embodiment of the present disclosure. The phrase appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to an independent or alternative embodiment that is mutually exclusive with other embodiments. It is expressly and implicitly understood by those skilled in the art that an embodiment described herein may be combined with other embodiments.

[0036] The following describes a CPU as an example to illustrate a method for operation in a neural network. Matrix multiplication has been widely used in the neural network, and in the embodiments of the present disclosure, a multiplication of a matrix A and a matrix B is described as an example to illustrate an AND operation in the CPU. Assume that a multiplication result of the matrix A and the matrix B is a matrix C, i.e., C=A*B, which is illustrated below.



[0037] For the CPU, the matrix C can be calculated as follows. Multiplication calculation may be sequentially performed for the first line, the second line, and the third line, in other words, after the multiplication calculation for a line is completed, the multiplication calculation for another line can be performed. Taking the above formula as an example, the multiplication calculation for the first line is first performed and completed, by the CPU, to obtain results of a11 * b11+a12 * b21+a13 * b31, a11 * b12+a12 * b22+a13 * b32, and a11 * b13+a12 * b23+a13 * b33, and then the multiplication calculation for the second line is performed and completed to obtain results of a21 * b11+a22 * b11+a22 * b21+a23 * b31, a21 * b12+a22 * b22+a23 * b32, and a21 * b13+a22 * b23+a23 * b33, and the multiplication calculation for the third line is finally performed and completed to obtain results of a31 * b11+a32 * b21+a33 * b31, a31 * b12+a32 * b22+a33 * b32, and a31 * b13+a32 * b31+a33 * b33.

[0038] Therefore, for a CPU or a GPU, it needs to perform calculations line by line, that is, after a calculation operation for the first line is completed, a calculation operation for the second line is performed, and then a calculation operation for the third line is performed until calculation operations for all lines has been completed. A neural network may include thousands of rows to be calculated, such that a calculation operation for the neural network may cost a long period of time. During the calculation operation for the neural network, the CPU is constantly in operation state with high energy consumption.

[0039] FIG. 1b is a schematic structural diagram of a chip device. As illustrated in FIG. 1b, the chip device includes a main unit, basic units, and branch units. The main unit may include a register and/or an on-chip cache circuit, and may further include one or any combination of a vector arithmetic unit circuit, an arithmetic logic unit (ALU) circuit, an accumulator circuit, a matrix transposition circuit, a data rearrangement circuit, and/or a direct memory access (DMA) circuit. Each of the basic units may include a basic register and/or a basic on-chip cache circuit, and may further include one or any combination of an inner-product arithmetic unit circuit, a vector arithmetic unit circuit, an accumulator circuit, and/or the like. Each of the above-mentioned circuits can be an integrated circuit. In an implementation, the chip device is provided with branch units, the main unit is coupled with the branch units, and each of the branch units is connected to a corresponding basic unit. The basic units are configured to perform an inner-product operation between data blocks. The main unit is configured to receive and transmit external data, and to distribute the external data to the branch units. The branch unit is configured to receive and transmit data from the main unit or the basic unit. The structure of the chip device illustrated in FIG. 1b is applicable for calculations of complex data. Since the number of units connected to the main unit is limited, it is necessary to add branch units between the main unit and basic units to provide accesses for more basic units, so as to achieve calculations of complex data blocks.

[0040] The branch unit and a basic unit may be connected by any type of connection structure, which is not limited to an H-type structure illustrated in FIG. 1b. In one implementation, data from the main unit may be broadcasted or distributed to a basic unit and data from a basic unit may be gathered to the main unit. That is, data may be transmitted in a broadcast manner, a distribution manner, and a gather manner, which are described in detail as follows.

[0041] A data transfer manner of the main unit to the basic unit may include the follows.

[0042] In one implementation, the main unit is connected with multiple branch units respectively, and each of the branch units is connected with multiple basic units respectively.

[0043] In another implementation, the main unit is connected with a branch unit, and the branch unit is connected with another branch unit, by such analogy, multiple branch units are connected in series, and then, each of the multiple branch units is connected with multiple basic units respectively.

[0044] In yet another implementation, the main unit is connected with each of multiple branch units respectively, and each branch unit is connected in series with multiple basic units.

[0045] In still another implementation, the main unit is connected with a branch unit, and the branch unit is connected with another branch unit, by such analogy, multiple branch units are connected in series, and then, each of the multiple branch units is connected with multiple basic units in series.

[0046] When distributing data, the main unit transmits data to some or all of the basic units, and data received by each of the basic units may be different.

[0047] When broadcasting data, the main unit transmits data to some or all of the basic units, and data received by each of the basic units is the same.

[0048] When gathering data, some or all of the basic units transmit data to the main unit. It should be noted that the chip device illustrated in FIG. 1a or FIG. 1b may be a single physical chip. In practical applications, the chip device may also be integrated in another chip (such as a CPU, a GPU). The specific embodiments in the present disclosure do not limit the physical representation of the above-mentioned chip device.

[0049] FIG. 1c is a schematic diagram illustrating data distribution of a chip device. As illustrated by arrows in FIG. 1c, the arrows indicate a distribution direction of the data. As illustrated in FIG. 1c, after receiving external data, the external data is split and distributed, by the main unit, to the multiple branch units, and then the multiple branch units transmit the data split to the basic units.

[0050] FIG. 1d is a schematic diagram illustrating data return of a chip device. As illustrated by arrows in FIG. 1d, which indicate a return direction of the data. As illustrated in FIG. 1d, the basic units return data (for example, an inner-product operation result) to the branch units, and then the branch units return the data to the main unit.

[0051] FIG. 1a is a schematic structural diagram illustrating another chip device. The chip device includes a main unit and basic units, and the main unit is connected with the basic units respectively. Referring to a structure illustrated in FIG. 1a, since each of the basic units is directly and physically connected with the main unit, the number of basic units connected to the structure is limited, which is suitable for simple data calculation only.

[0052] As illustrated in FIG. 2, a method for operations in a neural network by using the above-mentioned chip device is provided. The method is implemented by using a chip device illustrated in FIG. 1a or FIG. 1b. As illustrated in FIG. 2, the method includes the following.

[0053] At S201, the main unit of the chip device may acquire a data block to-be- calculated and an operation instruction.

[0054] The data block to-be-calculated at S201 may be a matrix, a vector, a three-dimensional data, a four-dimensional data, a multi-dimensional data, and the like. The form of the data block to-be-calculated is not limited in the embodiments of the present disclosure. The operation instruction may be a multiplication instruction, a convolution instruction, an addition instruction, a subtraction instruction, a Basic Linear Algebra Subprograms (BLAS) function, an activation function, or the like.

[0055] At S202, the main unit may divide the data block to-be-calculated into a distribution data block and a broadcast data block according to the operation instruction.

[0056] The S202 may be implemented as follows.

[0057] When the operation instruction is a multiplication instruction, a multiplier data block is determined to be the broadcast data block, and a multiplicand data block is determined to be the distribution data block.

[0058] When the operation instruction is a convolution instruction, an input data block is determined to be the broadcast data block, and a convolution kernel is determined to be the distribution data block.

[0059] Operations at S203 may include operations at S2031 and S2032.

[0060] At S2031, the main unit may split the distribution data block into multiple basic data blocks, and distribute the basic data blocks to the multiple basic units.

[0061] At S2032, the main unit may broadcast the broadcast data block to the multiple basic units.

[0062] In one implementation, the operations at S2031 and S2032 may also be performed cyclically. In a case that there is large amounts of data-to-be-calculated, the main unit may split the distribution data block into multiple basic data blocks and split each of the multiple basic data blocks into m basic data sub-blocks, and split the broadcast data block into m broadcast data sub-blocks. The main unit may distribute one basic data sub-block and broadcast one broadcast data sub-block each time. The basic data sub-block and the broadcast data sub-block are data blocks capable of performing parallel neural network computations. Take a multiplication of a matrix B of size 1000*1000 and a matrix A of size 1000*1000 as an example. A basic data block may be zth row data of the matrix A, and a basic data sub-block may be the first 20 columns of data in the zth row data of the matrix A, the broadcast data sub-block may be the first 20 rows of data in the zth column data of matrix B.

[0063] The basic data block at S203 may be a minimum data block capable of performing an inner-product operation. In a matrix multiplication, for example, the basic data block may be a row of data of a matrix. For example, in a convolution operation, the basic data block may be the weight of a convolution kernel.

[0064] A manner for distributing data blocks at S203 may refer to the description of the following embodiments, and details are not described herein again. The reference of a manner for broadcasting the broadcast data block may be made in the description of the following embodiments, and details are not described herein again.

[0065] Operations at S204 may include operations at S2041 and S2042.

[0066] At S2041, the basic unit of the chip device may perform an inner-product operation on the basic data block and the broadcast data block to obtain an operation result (alternatively, may obtain an intermediate result).

[0067] At S2042, when the operation result is not an intermediate result, the operation result may be transmitted to the main unit in a return manner.

[0068] The return manner at S204 may refer to the description of the following embodiments, and details are not described herein again.

[0069] At S205, the main unit may obtain the instruction result corresponding to the data blocks to-be-calculated and the operation instruction by processing the operation result.

[0070] A processing manner at S205 may be an accumulation operation, a sort algorithm, or the like. The present disclosure is not limited to a specific processing manner. The specific processing manner needs to be configured according to different operation instructions, for example, may also include a nonlinear transformation or the like.

[0071] In technical solutions provided by the present disclosure, when performing an operation, the main unit may perform the following operations. External data including a data block to-be-calculated and an operation instruction may be received, and the data block to-be- calculated and the operation instruction may then be acquired. A distribution data block and a broadcast data block of the data block to-be-calculated may be determined according to the operation instruction. The distribution data block may be split into multiple basic data blocks. The broadcast data block may broadcast to the multiple basic units. The multiple basic data blocks may be distributed to the multiple basic units. The multiple basic units may respectively perform inner-product operations for the multiple basic data blocks and the broadcast data block to obtain operation results, and return the operation results to the main unit. The main unit obtains the instruction result of the operation instruction according to the operation results returned. The technical point of this technical solution can be illustrated as follows. For a neural network, inner-product operations between data blocks need large amounts of calculations, resulting in a larger computational overhead and a longer period of calculation time. Therefore, the embodiments of the present disclosure first distinguish a distribution data block and a broadcast data block in data blocks to-be-calculated according to an operation instruction and the data blocks to-be-calculated. The broadcast data block is a data block that must be used when implementing an inner-product operation, and the distribution data block can be split in the inner-product operation. Taking a matrix multiplication as an example, the data block to-be-calculated may include a matrix A and a matrix B. An operation instruction can be a multiplication instruction (A*B). According to the rule of matrix multiplication, the matrix A is determined as the distribution data block subject to split, and the matrix B is determined as the broadcast data block. For the matrix multiplication, the multiplicand matrix A can be split into multiple basic data blocks, and the multiplier matrix B can be a broadcast data block. According to the definition of the matrix multiplication, data of each line of the multiplicand matrix A is subject to inner-product operations with the multiplier matrix B. Therefore, in technical solutions of the present application, the matrix A may be divided into M basic data blocks, and each of the M basic data blocks may be a row of data of the matrix A. Thus, for matrix multiplication, a time-consuming operation is performed by multiple basic units individually, such that in an inner-product operation, the multiple basic units can quickly calculate the operation results in parallel to reduce calculation time. In addition, less calculation time can also reduce the operating time duration of the chip device, thereby reducing power consumption.

[0072] The effects of the technical solutions provided by the present disclosure will be described below by practical examples. FIG. 2a is a schematic diagram illustrating a multiplication operation for a matrix A and a vector B. As illustrated in FIG. 2a, the matrix A has M rows and L columns, the vector B has L rows. Assume that a time duration required for the inner-product operation of one row of the matrix A and the vector B is t1. When a CPU or a GPU is employed to perform operations, only after a calculation operation for a line is completed, a calculation operation can be performed for another. Therefore, a time duration T0 of a calculation operation for the above-mentioned matrix A and the vector B by a GPU or a CPU will be M*t1. With technical solutions provided by the embodiments of the present disclosure, assume that the number of basic units is M, the matrix A will be split into M basic data blocks, and each basic data block can be a row of data of the matrix A. The M basic units can simultaneously perform inner-product operations in parallel, and then the calculation time duration is t1. In the technical solutions provided by the embodiments of the present disclosure, the time duration required is T1 = t1+t2+t3, where t2 is a time duration required for the main unit to split data, and t3 is a time duration required for processing the operation results of the inner- product operations to obtain an instruction result. The calculation amount required for splitting data and processing operation results are very small, and therefore, the time consumption is very small, that is, T0>>T1, and the total calculation time duration can be significant reduced in the technical solutions of the embodiments of the present disclosure. Moreover, as for the power consumption generated during processing the data to-be-calculated, the operating time duration of chip device provided by the present disclosure can be particularly short due to T0>>T1. Experiments have illustrated that when the operating time duration of the chip device is very short, the energy consumption will be much lower than the energy consumption of long operating time duration, thereby reducing energy consumption.

[0073] In the foregoing operations at S203, the main unit may broadcast the broadcast data block to the multiple basic units in multiple implementation manners, which may specifically be described as follows.

[0074] In a method A, the broadcast data block is broadcast to multiple basic units by one time. The term "broadcast" refers to performing "one-to-many" data transmission, that is, the main unit simultaneously sends the same data block to multiple (all or part of) basic units. For example, in a matrix multiplication between a matrix A and a matrix B, where the matrix B is implemented as a broadcast data block, the matrix B will be broadcast to the multiple basic units. For another example, in a convolution operation, the input data block is implemented as a broadcast data block, and the input data block will be broadcast to multiple basic units at one time. The advantage of the method A is that data transmission amount between the main unit and the multiple basic units can be saved by transmitting all the broadcast data to multiple basic units via only one broadcast.

[0075] In a method B, the broadcast data block is divided into multiple broadcast data sub-blocks, and the multiple broadcast data sub-blocks will be broadcast to multiple basic units by multiple times. For example, the matrix B may be broadcast to the multiple basic units by multiple times. Specifically, each N columns of data of the matrix B may be broadcast each time. The advantage of the method B is that configuration requirements of the multiple basic units can be decreased. Normally, a storage space of a register of a basic unit is relatively small, when the matrix B with a large amount of data is sent to the basic unit at one time, then a larger register space of the basic unit for storing the data of the matrix B will be needed. Since there are a large number of the basic units in the chip device, an increase of the register space of each of the basic units will inevitably increase the cost of the chip device. Therefore, in this case, the broadcast data block is broadcasted to the basic units in multiple times, such that the basic unit only needs to store a part of the data of the broadcast data block each time, thereby reducing the cost of the chip device.

[0076] It should be noted that, the method for distributing multiple basic data blocks to multiple basic units at S203, may also adopt the method A or the method B. The only difference is that the transmission method is unicast and the transmitted data is the multiple basic data blocks.

[0077] Operations at S204 may be implemented as follows.

[0078] When the method A is adopted to broadcast the broadcast data block and to distribute basic data block (as illustrated in FIG. 3a), the basic unit may perform inner-product operations to the basic data block and the broadcast data block to obtain an inner-product processing result. That is to perform an inner-product operation for one line at one time, and to transmit the inner-product processing result (one kind of the operation results) to the main unit, and the main unit accumulates the inner-product processing result. In practical applications, the basic unit may accumulate the inner-product processing result and transmit an accumulated result (another kind of the operation results) to the main unit. The above-mentioned method can reduce the amount of data transmission between the main unit and the basic unit, thereby improving calculation speed.

[0079] When the method B is adopted to broadcast the broadcast data block, the basic unit may perform an inner-product operation for a basic data block and a broadcast data sub-block to obtain an operation sub-result when receiving the broadcast data sub-block. The operation sub-result may be sent to the main unit by the basic unit, and the main unit performs accumulation operation for the processing result. In another implementation, when the basic unit receives n basic data blocks, n operation sub-results may be obtained by multiplexing the broadcast data block to perform inner-product operation with the n basic data blocks. The basic unit may send the n operation sub-results to the main unit, and the main unit may perform accumulation operation for the n processing results. The above accumulation operation can also be performed by the basic unit.

[0080] In the above case, the amount of data of the broadcast data block is generally very large, as well as the distribution data block. The chip device is a kind of a hardware configuration, the basic units of the chip device may be innumerable in theory. However, in practice, the number of the basic units may be generally several tens, which is limited and may change (such as increase) constantly with technology development. Therefore, as for the matrix multiplication operations in the neural network, the matrix A may have thousands of rows, and the matrix B may also have thousands of columns, such that it is impossible to transmit the matrix B to the basic unit at one time. Accordingly, an implementation method may be performed as follows. A part of the data of the matrix B (for example, the first five columns of the matrix B), may be broadcast to the basic units at each time. A similar implementation method may be adopted for the matrix A. The basic unit may perform an inner-product sub-operation each time, and then, a result of the inner-product sub-operation can be stored in the register, and after all the inner-product operations of the row are executed, an operation result may be obtained by performing an accumulation operation for all results of the inner-product operations of the row. The operation result obtained may be sent to the main unit, thereby improving calculation speed.

[0081] Referring to FIG. 3, a method for operations in a neural network is provided. The embodiments of the present disclosure describe a matrix multiplication of a matrix A and a matrix B as an example. The matrix multiplication of the matrix A and the matrix B is illustrated in FIG. 2a. For convenience of illustration, the calculation method of the neural network illustrated in FIG. 3 is performed in the chip device illustrated in FIG. 1b. As illustrated in FIG. 1b, the chip device is provided with 16 basic units. For convenience of description and distribution, the value of M as illustrated in FIG. 3a is 32, the value of N is 15, and the value of L is 20. It should be understood that the calculating device may have any number of basic units. The method illustrated in FIG 3 may include the following.

[0082] At S301, the main unit may receive the matrix A, the matrix B, and a multiplication operation instruction for the matrix A and the matric B.

[0083] At S302, the main unit may determine that the matrix B is implemented as a broadcast data block and the matrix B is implemented as a distribution data block according to the multiplication operation instruction. The matrix A is split into 32 basic data blocks, and each of the 32 basic data blocks is a row of data of the matrix A.

[0084] At S303, the main unit may equally distribute the 32 basic data blocks to the 16 basic units, that is, each basic unit receives two basic data blocks. It should be noted that, the basic data blocks may be distributed non-repetitively in an arbitrary order.

[0085] The basic data blocks at S303 may be distributed in other manners. For example, when the basic data blocks cannot be distributed equally to each basic unit, then the basic data blocks may be distributed to each basic unit unequally. A data block that cannot be equally distributed to the basic units, may be first split into several parts, and then the several parts may be equally distributed to the basic units. The manner in which the basic data blocks are distributed to multiple basic units is not limited in the embodiments of the present disclosure.

[0086] At S304, the main unit may select the first few columns of the matrix B (such as the first five columns), and broadcast data of the first five columns of the matrix B to the 16 basic units.

[0087] At S305, the 16 basic units may multiplex the data of the first five columns twice to perform inner-product operations with the two basic data blocks, and perform an accumulation operation to obtain 32*5 first-processing results, and then send the 32*5 first-processing results to the main unit.

[0088] At S306, the main unit may select data of the middle five columns of the matrix B, and broadcast the data of the middle five columns of the matrix B to 16 basic units.

[0089] At S307, the 16 basic units may multiplex the data of the middle five columns twice to perform inner-product operations with the two basic data blocks, and perform an accumulation operation to obtain 32*5 middle-processing results, and then send the 32*5 middle-processing results to the main unit.

[0090] At S308, the main unit may select data of the last five columns of the matrix B, and broadcast the data of the last five columns of the matrix B to the 16 basic units.

[0091] At S309, the 16 basic units may multiplex the data of the last 5 columns twice to perform inner-product operations with the two basic data blocks, and perform an accumulation operation to obtain 32*5 last-processing results, and then send 32*5 last-processing results to the main unit.

[0092] At S310, the main unit may combine the 32*5 first-processing results, the 32*5 middle-processing results, and the 32*5 last-processing results in an order from the first to the last to obtain a 32*15 matrix C. The matrix C is an instruction result of the multiplication operation instruction A*B.

[0093] In the technical solutions illustrated in FIG. 3, the matrix A is split into 32 basic data blocks, and then the matrix B may be broadcast in batches, such that the basic unit can obtain instruction results in batches. Since an inner-product operation is split and calculated by 16 basics units, the calculation time duration can be greatly reduced, which has advantages of short calculation time and low energy consumption.

[0094] FIG. 1a is a chip device provided by the disclosure. As shown in FIG. 1a, the chip device includes a main unit and a plurality of basic units. The main unit is a hardware chip unit, and the plurality of basic units are also hardware chip units.

[0095] The main unit is configured to perform each continuous operation in a neural network and to transmit data with the basic units.

[0096] The basic units are configured to perform parallel acceleration operations in the neural network according to the data transmitted by the main unit, and to transmit operation results to the main unit.

[0097] The above-mentioned parallel acceleration operations may include, but are not limited to, large-scale and parallelizable operations such as multiplication operations between data blocks, convolution operations, and the like.

[0098] The above-mentioned continuous operation includes, but is not limited to, accumulation operation, continuous operations such as matrix transposition operation, data sorting operation, and the like.

[0099] In one implementation, the main unit is configured to obtain a data block to-be-calculated and an operation instruction, to divide the data block to-be-calculated into a distribution data block and a broadcast data block according to the operation instruction, to obtain a plurality of basic data blocks by splitting the distribution data block, and to distribute the plurality of basic data blocks to the basic units and broadcast the broadcast data block to the basic units. The basic units are configured to obtain operation results by performing inner-product operations for the basic data blocks and the broadcast data block and to send the operation results to the main unit. The main unit is further configured to obtain an instruction result corresponding to the data block to-be-calculated and the operation instruction by processing the operation results.

[0100] In one implementation, the chip device further includes a branch unit disposed between the main unit and at least one basic unit. The branch unit is configured to forward data

[0101] In one implementation, the main unit is further configured to broadcast the broadcast data block to multiple basic units by one broadcast.

[0102] In one implementation, the basic units are further configured to obtain inner-product processing results by performing inner-product operations for each of the basic data blocks and the broadcast data block, to obtain the operation results by performing accumulation operation for each of the inner-product processing results respectively, and to send the operation results to the main unit.

[0103] In one implementation, the main unit is further configured to obtain accumulated results by performing accumulation operation for each of the operation results respectively when the operation results are inner-product processing results, and to obtain the instruction result by arranging the accumulated results, where the instruction result is corresponding to the data blocks to-be-calculated and the operation instruction.

[0104] In one implementation, the main unit is further configured to obtain a plurality of broadcast data sub-blocks by splitting the broadcast data block and to broadcast the plurality of broadcast data sub-blocks to the plurality of basic units by multiple broadcasts.

[0105] In one implementation, the basic units are further configured to obtain inner-product processing results by performing one inner-product operation for each of the broadcast data sub-blocks and each of the basic data blocks, to obtain operation sub-results by performing accumulation operation for each of the inner-product processing results, and to send the operation sub-results to the main unit.

[0106] In one implementation, the basic units are configured to obtain n processing sub-results by multiplexing each of the broadcast data sub-blocks n times to perform inner-product operation with n basic units, to obtain n operation sub-results by performing accumulation operation for each of the n processing sub-results respectively, and to send the n operation sub-results to the main unit, where the n is an integer greater than or equal to two.

[0107] According to one implementation of the present disclosure, an application method for the chip device illustrated in FIG. 1a is provided. The application method is applicable to perform one or any combination of a matrix-multiplying-matrix operation, a matrix-multiplying- vector operation, a convolution operation, or a fully connected operation.

[0108] In one implementation, the main unit may further perform neural network operations such as a pooling operation, a normalization operation (for example, a batch normalization and a local response normalization (LRN)), and the like.

[0109] According to one implementation of the present disclosure, a chip is provided, which includes the chip device illustrated in FIG. 1a or FIG. 1b.

[0110] According to one implementation of the present disclosure, an intelligent device is provided, which includes the above-mentioned chip that integrates the chip device illustrated in FIG. 1a or FIG. 1b. The intelligent device includes, but is not limited to, a smart phone, a tablet computer, a personal digital assistant, a smart watch, a smart camera, a smart television, a smart refrigerator and other smart devices. The above-mentioned devices are for illustrative purposes only, and the specific embodiments of the present application are not limited to the specific forms of the above-mentioned devices.

[0111] For the matrix multiplication operation, reference can be made to the description of the embodiments illustrated in FIG. 3, and details are not described herein again.

[0112] In one implementation, the chip device perform a fully connected operation as follows.

[0113] Assume input data of a fully connected layer is a vector of length L (such as a vector B in FIG. 3a illustrated a single sample of a fully connection 1, that is, the input data of the neural network is a single sample), output data of the fully connected layer is a vector of length M, and a weight of the fully connected layer is an M*L matrix (such as a matrix A in FIG. 3b illustrated a single sample of the fully connection 1). The weight matrix of the fully connected layer is implemented as the matrix A (i.e., distribution data block), and the input data of the fully connected layer is implemented as the vector B (i.e., broadcast data block). The fully connected operation can be performed according to the method illustrated in FIG. 2.

[0114] In another implementation, assume that the input data of the fully connected layer is a matrix (that is, input data of the neural network is multiple samples which are operated together as a batch). The input data of the fully connected layer represents N input samples and each sample is a vector of length L, and then the input data is represented by a L*N matrix such as a matrix B in FIG. 3b illustrated multiple samples of the fully connection 1, and each output of each sample in the fully connected layer is a vector of length M. The output data of the fully connected layer is an M*N matrix, such as a result matrix in FIG. 3a illustrated multiple samples of the fully connection 1, the weight of the fully connected layer is an M*L matrix (i.e., matrix A in FIG. 3a illustrated multiple samples of the fully connection 1). In this case, the weight matrix of the fully connected layer is implemented as the matrix A (i.e., distribution data block), and the input data matrix of the fully connected layer is implemented as the matrix B (i.e., broadcast data block). Alternatively, the weight matrix of the fully connected layer is implemented as the matrix B (i.e., broadcast data block), and the input data matrix of the fully connected layer is implemented as the matrix A (i.e., distribution data block). The specific operations can be performed according to the method illustrated in FIG. 2.

[0115] When the artificial neural network operation is performed by the chip device, input data of a convolutional layer, a pooling layer, a regularization layer, and the like in the neural network are illustrated in FIG. 3d illustrated input data of a convolution 2, where the regularization layer is also called a normalization layer, such as batch normalization (BN) or local response normalization(LRN). For the sake of clarity, the three-dimensional data block representing each sample is described with C=5, H=10, and W=12 as an example. In practice, the sizes of N, C, H, and W are not limited to those values illustrated in FIG. 3d. Each of the three-dimensional data blocks in FIG. 3d represents a sample corresponding to the input data of a layer. The three dimensions of each three-dimensional data block are represented by C, H and W, respectively, and the total number of the three-dimensional data blocks is N.

[0116] For the calculations of the above-mentioned neural network layers, when receiving the input data, the main unit arranges each sample of the input data in a certain order through the data rearrangement circuit of the main unit, where the order may be arbitrary.

[0117] For example, the input data may be arranged in an order that coordinates of data in the dimension C vary, such as in an order of NHWC, NWHC, and the like, where C is the dimension of the innermost layer of a data block, N is the dimension of the outermost layer of the data block, and H and W are the dimensions of the middle layer of the data block. In this way, the data in dimension C is arranged adjacently and consecutively, thereby facilitating an increase of operation parallelism and parallel operations for multiple feature maps.

[0118] The following implementations are described for the understanding of dimensions C, H, and W according to different neural network operations. For a convolution operation and a pooling operation, H and W are dimensions in which relevant operation windows slide when performing convolution and pooling operations. The operation window sliding in dimension W is illustrated in FIG. 3e illustrating a sliding a of a convolution 3 and illustrated in FIG. 3f illustrating a sliding b of the convolution 3, and the operation window sliding in dimension H is illustrated in FIG. 3g.). The size of the operation window is the same as the size of a convolution kernel in M convolution kernels. For example, as illustrated in FIG. 3c, each of the M convolution kernels is a 5*3*3 three-dimensional data block, and then the operation window is also a 5*3*3 three-dimensional data block. For the M convolution kernels illustrated in FIG. 3c, KH represents the dimension H of input data, and KW represents the dimension W of the input data. Gray squares in FIG. 3e, 3f, and 3g represent data to be calculated each time the operation window is slid, and the operation window may be operated to first slide in the dimension H and then to slide in dimension W; alternatively, the operation window may be operated to first slide in dimension W and then to slide in dimension H. Specifically, for the convolution operation, operations in each position of the operation window include performing inner-product operations for the data blocks indicated by the gray squares with the M convolution kernel data blocks illustrated in FIG. 3c respectively, such that one value may be output by each of the M convolution kernels each time the operation window is slid, that is, there are M values will be output at each position of the operation window (that is, each time the operation window is slid). For the pooling operation, operations in each position of the operation window include obtaining a maximum value, an average value, or the like, among the data blocks indicated by the gray squares in dimension H and dimension W (that is, 9 data blocks on the same plane of the gray data block in dimension H and dimension W), that is, there are C values will be output at each position of the operation window. C is another dimension in the three-dimensional data block of a single sample other than H and W, and N represents a total of N samples simultaneously performing the operation in a layer. For the LRN in the regularization operation, the dimension C is defined as a dimension along which the LRN operation selects a continuous data block (i.e., a data block of Y*1*1) each time, where Y in the data block of Y*1*1 is a value of the number of data selected in the dimension C, the value of Y is less than or equal to the maximum value of the number of data in the dimension C, the first "1" represents the number of data selected in the dimension H, and the second "1" represents the number of data selected in the dimension W. That is, when the regularization operation LRN is performed for each of the three-dimensional data blocks of one sample, each regularization operation LRN is performed for a part of the continuous data with the same W coordinate and H coordinate but different C coordinate. For the regularization operation BN, an average and a variance (or a standard deviation) of all values that have the same coordinate on dimension C in the three-dimensional data block of the N samples are calculated.

[0119] Each of squares in FIG. 3c to FIG. 3g represents a value, which may also be called a weight. The numbers used in the diagram are for illustrative purposes only. In practice, a dimension value may be any value (for example, a certain dimensional value is one, in this case the four-dimensional data block automatically becomes a three-dimensional data block. For another example, when the number of samples simultaneously calculated is one, the input data is a three-dimensional data block. For yet another example, when the number of convolution kernel is one, the convolution kernel data is a three-dimensional data block). A convolution operation between input data B and convolution kernel A is performed by the chip device.

[0120] A weight (all convolution kernels) of a convolution layer is illustrated in FIG. 3c. The number of convolution kernels is M, and each convolution kernel consists of C matrices of KH rows and KW columns, and the weight of the convolutional layer may be expressed as a four-dimensional data block with four dimensions of M, C, KH, and KW. The input data of the convolutional layer is a four-dimensional data block, which consists of N three-dimensional data blocks, and each of the three-dimensional data blocks consists of C characteristic matrices of H rows and W columns (i.e., a data block with four dimensions of N, C, H, W), as shown in FIG. 3d. A weight of each of the M convolution kernels is distributed from the main unit to one of K basic unit, and stored in an on-chip buffer and/or register of the basic unit (in this case, the M convolution kernels are implemented as distribution data blocks, and each convolution kernel can be implemented as a basic data block. Of course, in practical applications, the basic data block can also be changed into a smaller dimension, such as a matrix of a plane in a kernel). The specific method of distribution can be implemented as follows. When the number of the convolution kernels is M and M<=K, each of the M basic units is distributed with one weight of convolution kernels; when the number of the convolution kernels is M and M>K, each of the M basic units is distributed with one or more weights of convolution kernels. The convolution kernel weight set distributed to the ith basic unit is denoted by Ai, and there are a total of Mi convolution kernels. For each basic unit, such as the ith basic unit, the received convolution kernel weight Ai distributed by the main unit is stored in register and/or on-chip buffer of the basic unit. The main unit transmits each part of input data (such as data in operation windows illustrated in FIG. 3e, FIG. 3f and FIG. 3g) to each basic unit in broadcast manner (such as the over-mentioned method A or method B).When broadcasting, the weights of the operation window can be broadcast to all the basic units by multiple broadcasts. Specifically, a part of the weights of the operation window may be broadcasted each time. For example, a matrix of data in a plane is broadcasted each time, and taking FIG. 3e as an example, a KH*KW matrix on plane C can be broadcasted each time. In practical applications, first front n rows or first front n columns of a KH*HW matrix on plane C can also be broadcast at one time. The manner in which the above partial data is transmitted and the arrangement of partial data is not limited by the present disclosure .The input data may be arranged in an arbitrary dimension order, and then each part of the input data is sequentially broadcast to the basic unit. In one implementation, the foregoing distribution data (that is, the convolution kernels), may also be sent in a manner similar to that of the operation window of the input data, and details are not described herein again. In one implementation, the input data can be arranged in a cycle where the dimension C is in the innermost layer. In this way, the data in dimension C is arranged adjacently and consecutively, thereby facilitating an increase of operation parallelism and parallel operations for multiple feature maps. In one implementation, the input data can be arranged in a dimension order of NHWC or NWHC. Each basic unit, such as the ith basic unit, calculates the inner product of the convolution kernel in the weight Ai and a corresponding part of the received broadcast data (i.e., a part of data in the operation window), where the data of the corresponding part of the weight Ai can be read directly from the on-chip buffer, or can be read into the register for multiplexing. In one implementation, an accumulation operation may be performed for each of the inner-product processing results of basic units to obtain operation results, and then the operation results may be transmitted back to the main unit. Alternatively, each of the inner- product processing results of basic units may be transmitted back to the main unit, and then the main unit may perform an accumulation operation for each of the inner-product processing results. In one implementation, a sub-sum obtained by performing the inner-product operation via each basic unit may be stored and accumulated in the on-chip buffer and/or the register of the basic unit, and then transmitted back to the main unit. In one implementation, in some cases, a sub-sum obtained by performing the inner-product operation via each basic unit may be stored and accumulated in the on-chip buffer and/or the register of the basic unit, and then transmitted back to the main unit, in other cases, a sub-sum obtained by performing the inner-product operation via each basic unit may be directly transmitted to the basic unit and then accumulated by the main unit.

[0121] A method for implementing a basic linear algebra subprograms (BLAS) function by the chip device is provided as follows.

[0122] GEMM calculation refers to the matrix multiplication operation in the BLAS library, which is normally expressed as C = alpha*op(A)*op(B) + beta*C, where A and B are two input matrices, C is a output matrix, alpha and beta are scalars, and op represents an operation for matrix A or B. In addition, there will be some supplementary integers as parameters to explain the width and height of the matrices A and B.

[0123] Operations for implementing a GEMM calculation by the chip device are described as follows.

[0124] The main unit of the chip device performs corresponding op operations on matrix A and matrix B respectively. The op operation may be a matrix transposition or other operations, such as nonlinear function operation, pooling operation, and so on. The matrix op operation is implemented by the vector operation function of the main unit. The op of a matrix may be null, it indicates that the main unit performs no operation on the matrix.

[0125] A matrix multiplication of op(A) and op(B) is completed by employing the method illustrated in FIG. 2.

[0126] Each value of the results of op(A)*op(B) is multiplied by alpha, by employing the vector operation function of the main unit.

[0127] Data at corresponding positions in op(A)*op(B) and beta*C are added, by employing the vector operation function of the main unit.

[0128] GEMV calculation refers to the matrix-multiplying-vector operation in the BLAS library, which is normally expressed as C = alpha*op(A)*B+ beta*C, where A is an input matrix, B is an input vector, C is a output matrix, alpha and beta are scalars, and op represents an operation for matrix A.

[0129] Operations for implementing a GEMV calculation by the chip device are described as follows.

[0130] The main unit of the chip device performs a corresponding op operation on matrix A. The chip device completes the matrix-multiplying-vector operation of the matrix op(A) and the vector B. Each value of the results of op(A)*B is multiplied by alpha, by employing the vector operation function of the main unit. Data at corresponding positions in op(A)* B and beta*C are added, by employing the vector operation function of the main unit.

[0131] A method for implementing an activation function operation is provided as follows.

[0132] An activation function operation generally refers to performing a nonlinear operation for each data of a data block (such as a vector or a multi-dimensional matrix). For example, the activation function may be expressed as y=max(m, x), where x is an input value, y is an output value, and m is a constant. For example, the activation function may be expressed as y=tanh(x), where x is an input value and y is an output value. For another example, the activation function can be expressed as y=sigmoid(x), where x is an input value and y is an output value. For yet another example, the activation function may be a piecewise linear function. For still yet another example, the activation function can be any function that outputs a value when a value is inputted.

[0133] In a process for implementing an activation function operation, the chip device may input a vector to calculate an activation vector corresponding to the vector, by employing the vector operation function of the main unit. The main unit performs an activation function operation on each value of an input vector to obtain an output value, when the input of the activation function is a value, the output is also a value, and output the output value into the corresponding position of an output vector.

[0134] The above-mentioned input vector may be obtained from, but not limited to, external data of the chip device, and calculation result data of the basic unit forwarded by the branch unit of the chip device.

[0135] The above-mentioned calculation result data may specifically be an operation result of a matrix-multiplying-vector operation. The calculation result data may further be an operation result of a matrix-multiplying-matrix operation. The calculation result data may be an operation result of an add-offset operation implemented by the main unit.

[0136] A method for implementing an add-offset operation by the main unit is provided as follows.

[0137] Two vectors or two matrices can be added by the main unit. A vector can be added with each row or each columns of a matrix by the main unit

[0138] For example, the above-mentioned matrix may be a result of a matrix-multiplying-matrix operation performed by the chip device. The matrix may be a result of a matrix-multiplying-vector operation performed by the chip device. The matrix may be determined from external data received by the chip device.

[0139] The above-mentioned input data and the calculation result data are merely illustrative. In practical applications, data of other types or sources may also be applicable. The specific embodiments of the present disclosure do not limit the source and expression manner of the above-mentioned data.

[0140] It is to be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of action combinations, however, it will be appreciated by those skilled in the art that the present disclosure is not limited by the sequence of actions described. According to the present disclosure, certain steps or operations may be performed in other order or simultaneously. Besides, it will be appreciated by those skilled in the art that the embodiments described in the specification are exemplary embodiments and the actions and modules involved are not necessarily essential to the present disclosure.

[0141] In the foregoing embodiments, the description of each embodiment has its own emphasis. For the parts not described in detail in one embodiment, reference may be made to related descriptions in other embodiments.

[0142] In the embodiments of the disclosure, the apparatus disclosed in embodiments provided herein may be implemented in other manners. For example, the device/apparatus embodiments described above are merely illustrative; for instance, the division of the unit is only a logical function division and there can be other manners of division during actual implementations, for example, multiple units or components may be combined or may be integrated into another system, or some features may be ignored, omitted, or not performed. In addition, coupling or communication connection between each illustrated or discussed component may be direct coupling or communication connection, or may be indirect coupling or communication among devices or units via some interfaces, and may be electrical connection, mechanical connection, or other forms of connection.

[0143] In addition, the functional units in various embodiments of the present disclosure may be integrated into one processing unit, or each unit may be physically present, or two or more units may be integrated into one unit. The above-mentioned integrated unit/module can be implemented in the form of hardware. For example, the hardware can be a circuit, including a digital circuit, an analog circuit, and the like. Physical implementations of hardware structures include, but are not limited to, physical elements. The physical elements include, but not limited to, transistors, memristors, and the like. The calculating modules in a calculating device can be any suitable hardware processor, such as a CPU, a GPU, an FPGA, a DSP, an ASIC, and the like. The storage unit may be any suitable magnetic storage medium or magneto-optical storage medium such as an RRAM, a DRAM, a SRAM, an EDRAM, an HBM, an HMC, and the like.

[0144] The units described may or may not be physically separated, that is, they may be in the same place or may be distributed to multiple network elements. All or part of the units may be selected according to actual needs to achieve the purpose of the technical solutions of the embodiments.

[0145] The embodiments of the present disclosure have been described in detail above, and the principles and implementations of the present disclosure are described in the specific examples. The description of the above embodiments is merely used to facilitate understanding the method and core ideas of the present disclosure. For a person of ordinary skill in the art, in light of the present disclosure, there may be changes in the specific implementation and application scope. The contents of the present specification should not be construed as limiting the disclosure.


Claims

1. A matrix-multiplying-matrix operation method used in a chip device, wherein the chip device comprises a main unit and a plurality of basic units, and the method comprises:

receiving, by the main unit, a matrix A, a matrix B, and a multiplication instruction A*B; dividing, by the main unit, the matrix A into M basic data blocks; distributing, by the main unit, the M basic data blocks to the plurality of basic units; broadcasting, by the main unit, the matrix B to the plurality of basic units;

performing, by the plurality of basic units, inner-product operations in parallel on the basic data blocks and the matrix B to obtain a plurality of processing results; sending, by the plurality of basic units, the plurality of processing results to the main unit; and

combining, by the main unit, the plurality of processing results to obtain a computation result of the multiplication instruction.


 
2. The method of claim 1, wherein distributing the M basic data blocks to the plurality of basic units includes:
distributing the M basic data blocks to the plurality of basic units non-repetitively in an arbitrary order.
 
3. The method of claim 1, wherein broadcasting, by the main unit, the matrix B to the plurality of basic units; performing, by the plurality of basic units, the inner-product operations in parallel on the basic data blocks and the matrix B to obtain the plurality of processing results; sending, by the plurality of basic units, the plurality of processing results to the main unit; and combining, by the main unit, the plurality of processing results to obtain the computation result of the multiplication instruction include:
broadcasting, by the main unit, first α columns of the matrix B to the plurality of basic units; performing, by the plurality of basic units, the inner-product operations in parallel on the basic data blocks and the first α columns of the matrix B to obtain a plurality of first-processing results; sending, by the plurality of basic units, the plurality of first-processing results to the main unit; broadcasting, by the main unit, middle β columns of the matrix B to the plurality of basic units; performing, by the plurality of basic units, inner-products operations in parallel on the basic data blocks and the middle β columns of the matrix B to obtain a plurality of middle-processing results; sending, by the plurality of basic units, the plurality of middle-processing results to the main unit; broadcasting, by the main unit, last γ columns of the matrix B to the plurality of basic units; performing, by the plurality of basic units, inner-products operations in parallel on the basic data blocks and the last γ columns of the matrix B to obtain a plurality of last-processing results; sending, by the plurality of basic units, the plurality of last-processing results to the main unit; combining, by the main unit, the plurality of first-processing results, the middle-processing results and the last-processing results in a first-middle-last order to obtain a matrix C, wherein the matrix C is the computation result of the multiplication instruction.
 
4. The method of claim 1, wherein if the plurality of basic units are k basic units, and k is an integer greater than or equal to 2, distributing, by the main unit, the M basic data blocks to the plurality of basic units includes:

if M>k, distributing, by the main unit, one or more of the M basic data blocks to one of the k basic units; and

if Mk, distributing, by the main unit, one of the M basic data blocks to one of the k basic units.


 
5. The method of any one of claims 1 to 4, wherein the chip device further includes branch units configured to connect the main unit to a plurality of basic units, and the method further includes:
transmitting, by the branch units, data among the main unit and the plurality of basic units.
 
6. The method of any one of claims 1 to 5, wherein the main unit includes one or any combination of a vector arithmetic unit circuit, an arithmetic logic unit circuit, an accumulator circuit, a matrix transposition circuit, a direct memory access circuit, and a data rearrangement circuit.
 
7. The method of any one of claims 1 to 6, wherein each of the basic units includes one or any combination of an inner-product arithmetic unit circuit, an accumulator circuit, and the like.
 
8. A chip device comprising a main unit and a plurality of basic units, wherein the chip device is adapted to carry out the method of any one of claims 1 to 7.
 
9. A chip device comprising a main unit and a plurality of basic units, wherein the chip device is adapted to carry out the method of any one of claims 1 to 3, wherein the chip device further includes branch units configured to connect the main unit to a plurality of basic units, and the method further includes transmitting, by the branch units, data among the main unit and the plurality of basic units, and wherein the branch units include a plurality of branch units, and wherein each of the branch units is connected to the main unit and at least one of the basic units.
 
10. The chip device of claim 9, wherein the main unit includes one or any combination of a vector arithmetic unit circuit, an arithmetic logic unit circuit, an accumulator circuit, a matrix transposition circuit, a direct memory access circuit, and a data rearrangement circuit.
 
11. A computation device comprising the chip device of any one of claims 8 to 10.
 
12. A computer readable storage medium configured to store a computer program for exchanging electronic data, wherein the computer program enables the computer to execute the method of any one of claims 1 to 7.
 




Drawing