Background technology
Data processing is the step of most of algorithm needs to pass through or stage, after computer introduces data processing field,
More and more data processings realize have computing device carrying out the calculating of matrix data in existing algorithm by computer
Shi Sudu is slow, and efficiency is low.
Apply for content
The embodiment of the present application provides a kind of computational methods and Related product, can lift the processing speed of computing device, carry
High efficiency.
First aspect, there is provided a kind of computational methods, applied in computing device, the computing device include storage medium,
Register cell and matrix operation unit, the described method includes:
The computing device controls the matrix operation unit to obtain the first operational order, and first operational order is used for
Realize the computing between vector and matrix, the vector that first operational order includes performing needed for described instruction reads instruction,
The required vector is at least one vector, and at least one vector is the vector that length is identical or length is different;
The computing device controls the matrix operation unit to read instruction according to the vector and is sent out to the storage medium
Send reading order;
The computing device controls the matrix operation unit to be read using batch reading manner from the storage medium
The vector reads the corresponding vector of instruction, and performs first operational order to the vector.
It is described that the vector execution first operational order is included in some possible embodiments:
The computing device controls the matrix operation unit to be held using the calculation of multistage pipelining-stage to the vector
Row first operational order.
In some possible embodiments, include pre-set fixation in each pipelining-stage in the multistage pipelining-stage
Arithmetic unit, the fixation arithmetic unit in each pipelining-stage differ;
The computing device controls the matrix operation unit to be held using the calculation of multistage pipelining-stage to the vector
Row first operational order includes:
The computing device controls the matrix operation unit to be opened up according to the corresponding calculating network of first operational order
Flutter, utilize K1Selecting operation device in level pipelining-stage to the vector be calculated first as a result, again by described first
As a result it is input to K2Selecting operation device in level pipelining-stage perform be calculated second as a result, and so on, until by i-th -1
A result is input to KjSelecting operation device in level pipelining-stage, which performs, is calculated i-th of result;I-th of result is defeated
Enter to the storage medium and stored;
Wherein, KjBelong to any pipelining-stage in i pipelining-stage, j is less than or equal to i, and j and i are positive integer, described more
Quantity i, the selected execution sequence K of the multistage pipelining-stage of level pipelining-stagejAnd the KjSelection fortune in level pipelining-stage
It is to determine that the Selecting operation device is the fixed computing according to the calculating topological structure of first operational order to calculate device
Arithmetic unit in device.
In some possible embodiments, it is described multistage pipelining-stage in each pipelining-stage included by fixation arithmetic unit with
And the quantity of the fixed arithmetic unit is by user side or the self-defined setting in computing device side.
In some possible embodiments, the arithmetic unit in the multistage pipelining-stage in each pipelining-stage include it is following in
Any one or multinomial combination:Addition of matrices arithmetic unit, matrix multiplication operation device, matrix scalar multiplication arithmetic unit, non-linear fortune
Calculate device and matrix comparison operation device.
In some possible embodiments, first operational order includes any one of following:Vectorial derivation instruction
VDIER, vector generation diagonal matrix instruction VDIAG, vector multiply transposed matrix instruction VMULT.
In some possible embodiments, the instruction format of first operational order includes command code and at least one behaviour
Make domain, command code is used for the function of indicating the operational order, and arithmetic element is by identifying that the command code can carry out different vectors
Computing, operation domain are used to indicate the data message of the operational order, wherein, data message can be immediate or register number,
For example, when obtaining a vector, vector start address and vector can be obtained in corresponding register according to register number
Length, the vector of appropriate address storage is obtained further according to vector start address and vector length in storage medium.Alternatively, may be used
Any one of following middle information or multinomial combination are obtained in corresponding registers:Vectorial line number, row needed for described instruction
Number, data type, mark, storage address (first address) and dimension length, the dimension length refer to vector line length and/
Or the length of vector row.
In some possible embodiments, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage is included in advance
The matrix multiplication operation device first set, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix and compare fortune
Device is calculated, third level pipelining-stage includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First fortune
Calculate instruction and instruct VDIER for vectorial derivation,
The computing device controls the matrix operation unit to be held using the calculation of multistage pipelining-stage to the vector
Row first operational order includes:
The computing device controls the matrix operation unit to input the vector to non-thread in third level pipelining-stage
Property arithmetic unit to the matrix carry out matrix element local derviation the first result is calculated;Second result is inputted to described and is deposited
Storage media is stored.
In some possible embodiments, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage is included in advance
The matrix multiplication operation device first set, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix and compare fortune
Device is calculated, third level pipelining-stage includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First fortune
Calculate instruction and instruct VDIAG for vector generation diagonal matrix,
The computing device controls the matrix operation unit to be held using the calculation of multistage pipelining-stage to the vector
Row first operational order includes:
The computing device controls the matrix operation unit by the vector input to the matrix in the pipelining-stage of the second level
Comparison operation device carries out matrix element address and compares and obtain the first result at interval of the corresponding vector element of n+1 positions write-in;
First result is inputted to the storage medium and is stored.
In some possible embodiments, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage is included in advance
The matrix multiplication operation device first set, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix and compare fortune
Device is calculated, third level pipelining-stage includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First fortune
Calculate instruction and multiply transposed matrix instruction VMULT for vector,
The computing device controls the matrix operation unit to be held using the calculation of multistage pipelining-stage to the vector
Row first operational order includes:
The computing device controls the matrix operation unit by the vector input to the matrix in first order pipelining-stage
Multiplicative operator multiplies into row vector is calculated the first result;First result is inputted to the storage medium and is deposited
Storage.
In some possible embodiments, the vector, which reads instruction, to be included:Vectorial storage needed for described instruction
Vectorial mark needed for location or described instruction.
In some possible embodiments, when the vector, which is read, is designated as mark vectorial needed for described instruction,
The computing device controls the matrix operation unit to read instruction according to the vector and is sent out to the storage medium
Reading order is sent to include:
It is single that the computing device controls the matrix operation unit to be used according to the mark from the register cell
Position reading manner reads the corresponding storage address of the mark;
The computing device controls the matrix operation unit to be sent to the storage medium and reads the storage address
Reading order simultaneously obtains the vector using batch reading manner.
In some possible embodiments, the computing device further includes:Buffer unit, the method further include:
Pending operational order is cached in the buffer unit by the computing device.
In some possible embodiments, control the matrix operation unit to obtain the first computing in the computing device and refer to
Before order, the method further includes:
The computing device determines first operational order and the second operational order before first operational order
With the presence or absence of incidence relation, if first operational order and second operational order there are incidence relation, will described in
First operational order is cached in the buffer unit, after second operational order is finished, from the buffer unit
Extract first operational order and be transmitted to the arithmetic element;
Described definite first operational order whether there is with the second operational order before the first operational order to be associated
System includes:
Required the first vectorial storage address section in first operational order is extracted according to first operational order,
Required the second vectorial storage address section in second operational order is extracted according to second operational order, if described
First storage address section has overlapping region with the second storage address section, it is determined that first operational order with
Second operational order has incidence relation, if the first storage address section and the second storage address section are not
With overlapping region, it is determined that first operational order does not have incidence relation with second operational order.
Second aspect, there is provided a kind of computing device, the computing device include being used for the method for performing above-mentioned first aspect
Functional unit.
The third aspect, there is provided a kind of computer-readable recording medium, it stores the computer journey for electronic data interchange
Sequence, wherein, the computer program causes computer to perform the method that first aspect provides.
Fourth aspect, there is provided a kind of computer program product, the computer program product include storing computer journey
The non-transient computer-readable recording medium of sequence, the computer program are operable to make computer perform first aspect offer
Method.
5th aspect, there is provided a kind of chip, the chip include the computing device that as above second aspect provides.
6th aspect, there is provided a kind of chip-packaging structure, the chip-packaging structure include as above the 5th aspect and provide
Chip.
7th aspect, there is provided a kind of board, the board include the chip-packaging structure that as above the 6th aspect provides.
Eighth aspect, there is provided a kind of electronic equipment, the electronic equipment include the board that as above the 7th aspect provides.
In certain embodiments, the electronic equipment includes data processing equipment, robot, computer, printer, scanning
Instrument, tablet computer, intelligent terminal, mobile phone, automobile data recorder, navigator, sensor, camera, server, cloud server,
Camera, video camera, projecting apparatus, wrist-watch, earphone, mobile storage, wearable device, the vehicles, household electrical appliance, and/or medical treatment
Equipment.
In certain embodiments, the vehicles include aircraft, steamer and/or vehicle;The household electrical appliance include electricity
Depending on, air-conditioning, micro-wave oven, refrigerator, electric cooker, humidifier, washing machine, electric light, gas-cooker, kitchen ventilator;The Medical Devices include
Nuclear Magnetic Resonance, B ultrasound instrument and/or electrocardiograph.
Implement the embodiment of the present application, have the advantages that:
As can be seen that by the embodiment of the present application, computing device is provided with register cell and storage medium, is respectively used to
Store scalar data and vector data, and unit reading manner and batch are read the application for two kinds of memory distributions
Mode, the characteristics of by vector data distribution match the data reading mode of its feature, can be good at utilizing bandwidth, avoid
Because influence of the bottleneck of bandwidth to vectorial calculating speed, in addition, for register cell, since its storage is scalar
Data, there is provided the reading manner of scalar data, improve the utilization rate of bandwidth, so the technical solution that the application provides can
Utilize bandwidth well, avoid influence of the bandwidth to calculating speed, so it is fast with calculating speed, it is efficient the advantages of.
Embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, the technical solution in the embodiment of the present application is carried out clear, complete
Site preparation describes, it is clear that described embodiment is some embodiments of the present application, instead of all the embodiments.Based on this Shen
Please in embodiment, the every other implementation that those of ordinary skill in the art are obtained without creative efforts
Example, shall fall in the protection scope of this application.
Term " first ", " second ", " the 3rd " in the description and claims of this application and the attached drawing and "
Four " etc. be to be used to distinguish different objects, rather than for describing particular order.In addition, term " comprising " and " having " and it
Any deformation, it is intended that cover non-exclusive include.Such as contain the process of series of steps or unit, method, be
The step of system, product or equipment are not limited to list or unit, but alternatively further include the step of not listing or list
Member, or alternatively further include for the intrinsic other steps of these processes, method, product or equipment or unit.
Referenced herein " embodiment " is it is meant that a particular feature, structure, or characteristic described can wrap in conjunction with the embodiments
It is contained at least one embodiment of the application.Each position in the description occur the phrase might not each mean it is identical
Embodiment, nor the independent or alternative embodiment with other embodiments mutual exclusion.Those skilled in the art explicitly and
Implicitly understand, embodiment described herein can be combined with other embodiments.
It should be noted that this application involves matrix be specifically as follows m*n matrixes, wherein, m and N are more than or equal to 1
Integer, when m or n is 1, is represented by 1*n matrixes or m*1 matrixes, is referred to as vector;When m and n at the same time for 1 when, can be with
It is considered as the Special matrix of 1*1.Following matrixes all can be in above-mentioned three types matrix any one, do not repeating below.
The embodiment of the present application provides a kind of computational methods, which can be applied in computing device.It is this such as Fig. 1
A kind of structure diagram of possible computing device shown in inventive embodiments.Computing device as shown in Figure 1 includes:
Storage medium 201, for storage matrix (or vector).The preferable storage medium can be that scratchpad is deposited
Reservoir, it would be preferable to support the matrix data (or vector data) of different length;Necessary calculating data are temporarily stored in by the application
In scratchpad (Scratchpad Memory), make this arithmetic unit can be more during matrix operation is carried out
Flexibly effectively support the data of different length.Above-mentioned storage medium can also be the outer database of piece, database or other energy
Medium enough stored etc..
Register cell 202, for storing scalar data, wherein, which includes but not limited to:Matrix data or
Vector data (the application is also referred to as matrix/vector) storage medium 201 storage address and vector with matrix operation when
Scalar.In one embodiment, register cell can be scalar register heap, there is provided scalar is posted needed for calculating process
Storage, scalar register not only store matrix address, and also storage has scalar data.(i.e. matrix is deposited it is to be understood that matrix address
Store up address, such as first address) also it is scalar.When being related to the computing of matrix and vector, arithmetic element not only will be from register list
Matrix address is obtained in member, corresponding scalar, such as the line number of matrix, columns, matrix function are also obtained from register cell
According to type (alternatively referred to as data type), matrix dimension length (the concretely length of row matrix, length of rectangular array etc.).
Arithmetic element 203 (the application is also referred to as matrix operation unit 203), for obtaining and performing the first operational order.
As shown in Fig. 2, the arithmetic element includes multiple arithmetic units, which includes but not limited to:Addition of matrices arithmetic unit 2031, square
Battle array multiplicative operator 2032, size comparison operation device 2033 (or matrix comparison operation device), 2034 and of nonlinear operator
Matrix scalar multiplication arithmetic unit 2035.
This method is as shown in figure 3, include the following steps:
Step S301, arithmetic element 203 obtains the first operational order, and first operational order is used for realization vector and square
The computing of battle array, first operational order include:Perform the vector needed for the instruction and read instruction.
In step S301, needed for the above-mentioned execution instruction vector read instruction be specifically as follows it is a variety of, for example, this
Apply in an optional technical solution, it can be required vectorial storage that the vector needed for the above-mentioned execution instruction, which reads instruction,
Address.And for example, in the application in another optional technical solution, the vector needed for the above-mentioned execution instruction reads instruction can be with
For required vectorial mark, the form of expression of the mark can be a variety of, for example, the title of vector, and for example, vectorial identification
Number, register number or storage address of the vector in register cell for another example.
Illustrate that above-mentioned first operational order includes below by the example of a reality perform needed for the instruction to
Amount reads instruction, it is assumed here that and the vector operation formula is f (x)=A+B, wherein, A, B are vector.So in the first computing
In instruction in addition to carrying the vector operation formula, storage address vectorial needed for the vector operation formula can also be carried, is had
Body, such as the storage address of A is 0000-0FFF, the storage address of B is 1000-1FFF.And for example, it can carry A's and B
Mark, for example, A be identified as 0101, B be identified as 1010.
Step S302, arithmetic element 203 reads instruction according to the vector and sends reading order to the storage medium 201.
The implementation method of above-mentioned steps S302 is specifically as follows:
It can be required vectorial storage address that such as the vector, which reads instruction, and arithmetic element 203 is sent out to the storage medium 201
Give the reading order of the reading storage address and corresponding vector is obtained using batch reading manner.
And for example the vector reads instruction when can be required vectorial mark, and arithmetic element 203 is according to the mark from deposit
The corresponding storage address of the mark is read using unit reading manner at device unit, then arithmetic element 203 is to the storage medium
201 send the reading order of the reading storage address and obtain corresponding vector using batch reading manner.
Above-mentioned single reading manner is specifically as follows, and reads every time as the data of unit, i.e. 1bit data.Set at this time
The reason for unit reading manner i.e. 1 reading manner, is, for scalar data, the capacity that it takes is very small, if adopted
With batch data reading manner, then the data volume of reading is easily more than the capacity of required data, can so cause bandwidth
Waste, so using unit reading manner here for the data of scalar and reading to reduce the waste of bandwidth.
Step S303, arithmetic element 203 reads the corresponding vector of the instruction using batch reading manner, which is performed
First operational order.
Batch reading manner is specifically as follows in above-mentioned steps S303, and reading every time is the data of multidigit, such as every time
The data bits of reading is 16bit, 32bit or 64bit, i.e., no matter the data volume needed for it is how many, it reads equal every time
For the data of fixed long number, the data mode that this batch is read is very suitable for the reading of big data, for vector, due to
Capacity shared by it is big, if using single reading manner, its speed read can be very slow, so being read here using batch
Mode is taken to obtain the data of multidigit so as to quickly read vector data, is avoided because read vector data influences to gauge slowly excessively
The problem of calculating speed.
The computing device for the technical solution that the application provides is provided with register cell and storage medium, it stores mark respectively
Measure data and vector data, and the application for two kinds of memory distributions unit reading manner and batch reading manner,
The characteristics of by vector data distribution match its feature data reading mode, can be good at utilize bandwidth, avoid because of
Influence of the bottleneck of bandwidth to vectorial calculating speed, in addition, for register cell, since its storage is scalar number
According to there is provided the reading manner of scalar data, the utilization rate of bandwidth being improved, so the technical solution that the application provides can be very
Good utilization bandwidth, avoids influence of the bandwidth to calculating speed, so it is fast with calculating speed, it is efficient the advantages of.
Optionally, it is above-mentioned that vector execution first operational order is specifically as follows:
Arithmetic element 203 can be realized using the calculation of multistage pipelining-stage, wherein, which can be user
Side or the computing device side self-defined setting in advance, be that fixed design is good.Such as in herein described computing device
It is designed with i grades of pipelining-stages.It is embodiment below:
Arithmetic element can be according to the corresponding calculating network topology of first operational order, Selection utilization K1Level pipelining-stage
In Selecting operation device to it is described vector perform be calculated first as a result, then reselection utilizes K2Choosing in level pipelining-stage
Select arithmetic unit to first result perform be calculated second as a result, and so on, select KjSelection in level pipelining-stage
Arithmetic unit, which performs the i-th -1 result, to be calculated i-th as a result, until completing the computing of first operational order.Here
I-th of result is to export result (being specially output matrix).Further, arithmetic element 203 can store the output result
To storage medium 201.
Wherein, the quantity i of the multistage pipelining-stage, the execution sequence of the multistage pipelining-stage (select KjLevel pipelining-stage)
And the KjSelecting operation device in level pipelining-stage is all specifically the calculating topological structure according to first operational order
Definite, i is positive integer.In general, i=3.May be provided with arithmetic unit correspondingly in each pipelining-stage, the arithmetic unit include but
It is not limited to any one of following or multinomial combination:Addition of matrices arithmetic unit, matrix multiplication operation device, matrix scalar multiplication fortune
Calculate device, nonlinear operator, matrix comparison operation device and other matrix operation devices.It is, included in each pipelining-stage
The quantity of fixed arithmetic unit and fixed arithmetic unit can have user side or the self-defined setting in computing device side, not do
Limit.
It is to be understood that in the application in above-mentioned computing device, the K performed is selected every time1、K2…KjLevel pipelining-stage and
Selecting operation device in pipelining-stage can all be repeated selection, be the execution number for not limiting each pipelining-stage.Hereinafter will be with institute
Exemplified by the first operational order is stated as vectorial derivation, it is described in detail.
In the specific implementation, as Fig. 4 shows a kind of configuration diagram of pipelining-stage.Such as Fig. 4, may be present between i grades of pipelining-stages
The bypass design (bypass circuit illustrated) connected entirely, for according to the corresponding calculating network topology of the first operational order, choosing
Select some arithmetic unit (the Selecting operation device i.e. in the application) in the current desired pipelining-stage used and the pipelining-stage.It is optional
Ground, the data transfer being additionally operable between multiple pipelining-stages, such as the output result of third level pipelining-stage is forwarded to first order stream
As input, being originally inputted be able to can make water level as the input of any level-one of three-level pipelining-stage, the output of any level-one
Final output for arithmetic element etc..
With i=3, exemplified by three-level flowing water, arithmetic element by bypass circuit, can select respectively the execution sequence of pipelining-stage with
And the respective required arithmetic unit (alternatively referred to as arithmetic unit) used in pipelining-stage.As Fig. 5 shows a kind of operation stream of pipelining-stage
Journey schematic diagram.Correspondingly, arithmetic element the vector is performed the first pipelining-stage be calculated first as a result, (optional) by the
One result be input to the second pipelining-stage perform the second pipelining-stage be calculated second as a result, (optional) by the second result input
The 3rd pipelining-stage, which is performed, to the 3rd pipelining-stage is calculated the 3rd as a result, (optional) stores the 3rd result to storage medium
201。
Above-mentioned first pipelining-stage includes but not limited to:Matrix multiplication operation device etc..
Above-mentioned second pipelining-stage includes but not limited to:Addition of matrices arithmetic unit, size comparison operation device etc..
Above-mentioned 3rd pipelining-stage includes but not limited to:Nonlinear operator, matrix scalar multiplication arithmetic unit etc..
By three pipelining-stage computings of vector point primarily to improving the speed of computing, for the calculating for vector, example
Such as using general processor when calculating, the step of its computing, is specifically as follows, and processor carries out vector to be calculated first
As a result, then by the storage of the first result in memory, processor from memory reads the first result and performs is calculated for the second time
Two as a result, then by the storage of the second result in memory, processor is calculated for the third time from interior performed from the second result of reading
3rd as a result, then by the storage of the 3rd result in memory.It can be seen that from the step of above-mentioned calculating and carried out in general processor
When vector calculates, it does not shunt water level and is calculated, then be required to after calculating every time the data that will have been calculated into
Row preservation, needs to again read off when next time calculates, so this scheme needs repetition storage to read multiple data, for the application's
For technical solution, the first result that the first pipelining-stage calculates is directly entered the grading row calculating of the second flowing water, the second pipelining-stage meter
The second result calculated enters directly into the 3rd pipelining-stage and is calculated, the first result that the first pipelining-stage and the second pipelining-stage calculate
With the second result without storage, which reduce the occupied space of memory first, secondly, which obviate result multiple storage and
Read, improve the utilization rate of bandwidth, further increase computational efficiency.
In another embodiment of the application, each flowing water component can be freely combined or take level-one pipelining-stage.Such as will
Second pipelining-stage and the 3rd pipelining-stage merge, and either all merge first and second and the 3rd assembly line or each
Pipelining-stage is responsible for different computings can be with permutation and combination.For example, first order flowing water is responsible for comparison operation, and partial product computing,
Two level flowing water is responsible for the combination such as nonlinear operation and matrix scalar multiplication.It is that the i pipelining-stage designed in the application is supported to appoint
Multiple pipelining-stages of anticipating are in parallel, connect and merge, and to form different permutation and combination, the application does not limit.
It should be noted that the arithmetic unit in above-mentioned computing device in each pipelining-stage be in advance it is self-defined set,
Once it is determined that do not allow to change;That is i grades of pipelining-stage may be designed as the permutation and combination of any arithmetic unit, i grades of pipelining-stages once driving not
Change again, different operational orders can design different i level flowing water stage arrangements.Wherein, which can be according to specific instruction
Demand, the quantity of adaptability increase/less pipelining-stage.Finally, the flowing water stage arrangement designed for different instruction can be combined
Together, the computing device is formed.
Using above-mentioned computing device (i.e. arithmetic unit in every grade of pipelining-stage/arithmetic unit design is fixed), have below with
Lower beneficial effect:In addition to bandwidth is improved, no extra selection signal judges expense, without identical operational part between different pipelining-stages
Part is overlapping and redundancy, durability height, area are small.
Optionally, above-mentioned computing device can also include:Buffer unit 204, for caching the first operational order.Instruction exists
In implementation procedure, while also it is buffered in instruction cache unit, after an instruction has performed, if the instruction is at the same time
It is not to be submitted an instruction earliest in instruction in instruction cache unit, which will carry on the back and submits, once submit, this instruction
Change of the operation of progress to unit state will be unable to cancel.In one embodiment, instruction cache unit can be reset
Sequence caches.
Optionally, the above method can also include before step S301:
Determine that first operational order whether there is incidence relation with the second operational order before the first operational order, such as
First operational order there are incidence relation, is then performed with the second operational order before the first operational order in the second operational order
After finishing, first operational order is extracted from buffer unit and is transferred to arithmetic element 203.If the first operational order is with being somebody's turn to do
Instruction onrelevant relation before first operational order, then be directly transferred to arithmetic element by the first operational order.
Above-mentioned definite first operational order whether there is with the second operational order before the first operational order to be associated
The concrete methods of realizing of system can be:
Required the first vectorial storage address section, foundation in first operational order are extracted according to first operational order
Second operational order extracts required the second vectorial storage address section in second operational order, such as the first stored address area
Between with the second storage address section there is overlapping region, it is determined that the first operational order has with the second operational order to be associated
System.Such as the first storage address section and the non-overlapping region in the second storage address section, it is determined that the first operational order and second
Operational order does not have incidence relation.
There is overlapping region the first operational order of explanation occur in trivial of this storage and have accessed phase with the second operational order
Same vector, for vector, is as judgement since the space of its storage is bigger, such as using identical storage region
The no condition for incidence relation, in fact it could happen that situation be, the second operational order access storage region contain the first computing
The storage region accessed is instructed, is deposited for example, the second operational order accesses A vectors storage region, B vectors storage region and C vectors
Storage area domain, if A, B storage region are adjacent or A, C storage region are adjacent, the second operational order access storage region be, A,
B storage regions and C storage regions, or A, C storage region and B storage regions.In this case, if the first operational order
The storage region for A vectors and D vectors accessed, then the vectorial storage region that the first operational order accesses can not be with second
The vectorial storage region of operational order model essay is identical, if using identical Rule of judgment, it is determined that the first operational order with
Second operational order does not associate, but it was verified that the first operational order and the second operational order belong to incidence relation at this time, institute
With the application by whether have overlapping region to determine whether for incidence relation condition, the erroneous judgement of the above situation can be avoided.
Illustrate which kind of situation belongs to incidence relation below with the example of a reality, which kind of situation belongs to dereferenced pass
System.It is assumed here that the vector needed for the first operational order is A vector sums D vectors, the storage region of wherein A vectors is【0001,
0FFF】, the storage region of D vectors is【A000, AFFF】, it is A vectors, B vector sums C for the vector needed for the second operational order
Vector, its corresponding storage region are【0001,0FFF】、【1000,1FFF】、【B000, BFFF】, refer to for the first computing
For order, its corresponding storage region is:【0001,0FFF】、【A000, AFFF】, for the second operational order, it is corresponded to
Storage region be:【0001,1FFF】、【B000, BFFF】, so the storage region of the second operational order and the first operational order
Storage region there is overlapping region【0001,0FFF】, so the first operational order has incidence relation with the second operational order.
It is assumed here that the vector needed for the first operational order is E vector sums D vectors, the storage region of wherein A vectors is
【C000, CFFF】, the storage region of D vectors is【A000, AFFF】, it is A vectors, B for the vector needed for the second operational order
Vector sum C vectors, its corresponding storage region are【0001,0FFF】、【1000,1FFF】、【B000, BFFF】, for
For one operational order, its corresponding storage region is:【C000, CFFF】、【A000, AFFF】, come for the second operational order
Say, its corresponding storage region is:【0001,1FFF】、【B000, BFFF】, so the storage region of the second operational order and the
The storage region of one operational order does not have overlapping region, so the first operational order and the second operational order onrelevant relation.
In the application, if Fig. 6 A are a kind of instruction (concretely the first operational order, or operation that the application provides
Instruction) instruction set form schematic diagram, as shown in Figure 6A, operational order includes a command code and an at least operation domain, wherein,
Command code is used to indicate the function of the operational order, arithmetic element by identifying that the command code can carry out different vector operations,
Operation domain is used to indicate the data message of the operational order, wherein, data message can be immediate or register number, for example,
When obtaining a vector, vector start address and vector length can be obtained in corresponding register according to register number,
The vector of appropriate address storage is obtained in storage medium further according to vector start address and vector length.
I.e. the first operational order can include:Operation domain and at least one command code, by taking vector operation instruction as an example, such as
Shown in table 1, wherein, register 0, register 1, register file 2, register 3, register 4 can be operation domain.Wherein, each
Register 0, register 1, register 2, register 3, register 4 be used for marker register numbering, its can be one or
Multiple registers.It is to be understood that the quantity of register does not limit in command code, each register is used to storage computing and refers to
The related data information of order.
If Fig. 6 B are the fingers for another instruction (can be the first operational order, be alternatively referred to as operational order) that the application provides
Make collection form schematic diagram, as shown in Figure 6B, instruction include at least two command codes and at least an operation domain, wherein, it is described extremely
Few two command codes include the first command code and the second command code (diagram is respectively command code 1 and command code 2).The command code
1 is used for the type (i.e. certain major class instruction) of indicator, such as can concretely I/O instruction, logical order or operational order etc.
Deng, the command code 2 is used for the function (explanation of the specific instruction i.e. under major class instruction) of indicator, such as in operational order
Matrix operation command (such as Matrix Multiplication vector instruction MMUL, matrix inversion command M INV), vector operation instruction is (as vector is asked
Lead instruction VDIER etc.) etc., the application does not limit.
It is to be understood that the form of instruction can be user side or the self-defined setting in computing device side.The behaviour of instruction
Regular length, such as 8bit, 16bit etc. are may be designed as code.Instruction format as shown in Fig. 6 A has the advantage that feature:
Command code occupancy digit is few, decoding system design is simple.Instruction format as shown in Fig. 6 B has the advantage that feature:It is variable
Long, decoding average efficiency higher, in the case that certain major class instructs lower specific instruction less and calls frequency height, design its second
The length of command code (i.e. command code 2) is short and small, can improve decoding efficiency.Moreover it is possible to strengthen the readable and expansible of instruction
Property, optimizes the coding structure of instruction.
In the embodiment of the present application, instruction set includes the operational order of difference in functionality, concretely:
Vectorial derivation instructs (VDIER), according to the instruction, device from memory (preferable scratchpad or
Scalar register heap) specified address take out setting length matrix data, carried out in arithmetic element to vector (solve refined
Than matrix) computing, and result back into.Preferably, and by result of calculation being written back to memory, (preferable scratchpad is deposited
Reservoir or scalar register heap) specified address.
Vector generation diagonal matrix instruction (VDIAG), according to the instruction, from memory, (preferable scratchpad stores device
Device or scalar register heap) specified address take out setting length matrix data, carried out in arithmetic element to element vector
Element is mapped as the computing of matrix diagonals member generation diagonal matrix, and results back into.Preferably, and by result of calculation it is written back to storage
The specified address of device (preferable scratchpad or scalar register heap).
Vector multiplies itself transposition generator matrix instruction (VMULT, the application are also referred to as vector and multiply transposed matrix instruction), according to
The instruction, device take out setting length from the specified address of memory (preferable scratchpad or scalar register heap)
The matrix data of degree, multiplies the computing of itself transposition generator matrix in arithmetic element into row vector, and results back into.It is preferred that
, and result of calculation is written back to the specified address of memory (preferable scratchpad or scalar register heap).
It is to be understood that operation/operational order that the application proposes is mainly used for vector being extended for matrix form, easy to square
Computing between battle array under appropriate dimension.To realize above-mentioned function, the arithmetic unit designed in every grade of pipelining-stage is including but not limited to following
Any one of or multinomial combination:It is addition of matrices arithmetic unit, matrix multiplication operation device, matrix scalar multiplication arithmetic unit, non-thread
Property arithmetic unit, matrix comparison operation device.
Be exemplified below this application involves operational order (i.e. the first operational order) calculating.
Be exemplified by vectorial derivation instructs VDIER by first operational order give two vectors solve one of them to
The amount Jacobian matrix vectorial on another, that is, solve Jacobian matrix.During specific implementation, vector X and vector Y is given, is pressed
Jacobian matrixs of the Y on X is solved according to equation below.
Wherein, xnFor nth elements in vectorial X, ymFor m-th of element in vectorial Y, m and n are positive integer.
Correspondingly, the instruction format of vectorial derivation instruction VDIER is specially:
With reference to previous embodiment, arithmetic element can obtain vectorial derivation instruction VDIER, and after being decoded to it, pass through bypass
Nonlinear operator in the 3rd pipelining-stage of circuit Selection utilization carries out vector to ask local derviation to calculate (i.e. execution matrix element local derviation
Calculate) obtain the first result (exporting result).Alternatively, which is stored into storage medium.
By taking first operational order instructs VDIAG for vector generation diagonal matrix as an example, the diagonal side for giving vector is calculated
Battle array.During specific implementation, a vector X is given, inputs each element of vector as diagonal side in sequence according to equation below
The diagonal element of battle array, to obtain diagonal square matrix A correspondingly.
Correspondingly, the instruction format of vector generation diagonal matrix instruction VDIAG is specially:
With reference to previous embodiment, arithmetic element can obtain vector generation diagonal matrix instruction VDIAG, and after being decoded to it, lead to
The matrix comparison operation device crossed in the second pipelining-stage of bypass circuit Selection utilization carries out element address to vector and compares and every
Every the N+1 vectorial respective element of position write-in, to obtain the first result (i.e. output is as a result, diagonal square matrix).Alternatively, by this
One result is stored into storage medium.
Multiplied by first operational order for vector exemplified by transposed matrix instructs VMULT, calculate a given vector with it is another
One transposition for giving vector is multiplied.During specific implementation, two vector X and Y are given, is calculated according to equation below and obtains output matrix
A.X and Y can be given row vector or column vector, not limit.
A=x × yT
Correspondingly, vector multiply transposed matrix instruction VMULT instruction format be specially:
With reference to previous embodiment, arithmetic element can obtain vector and multiply transposed matrix instruction VMULT, and after being decoded to it, lead to
The matrix multiplication operation device crossed in the first pipelining-stage of bypass circuit Selection utilization, which multiplies vector into row vector, is calculated the first knot
Fruit (exports result).Alternatively, which is stored into storage medium.
It should be noted that the acquisition and decoding of above-mentioned various operational orders will be described in detail later.Ying Li
Solution, the calculating of each operational order (such as vectorial derivation instruction VDIER) is realized using the structure of above-mentioned computing device, can be obtained
Obtain following beneficial effect:The scalable of vector, it is possible to reduce instruction number, the use of reduction instruction;Difference can be handled to deposit
The matrix of form (row-major order and row main sequence) is stored up, avoids the expense converted to matrix;Support is stored according to certain intervals
Matrix format, avoid the executive overhead that is converted to matrix storage format and the space hold of storage intermediate result.
Setting length in above-mentioned operational order (i.e. vector operation instruction/first operational order) can voluntarily be set by user
Fixed, in an optional embodiment, which can be arranged to a value by user, certainly in practical applications,
The setting length can also be arranged to multiple values by user.The application embodiment does not limit the specific of the setting length
Value and number.For the purpose, technical scheme and advantage of the application are more clearly understood, below in conjunction with specific embodiment, and
Referring to the drawings, the application is further described.
Refering to Fig. 7, Fig. 7 is another computing device 50 that the application embodiment provides.Shown in Fig. 7, dress is calculated
Putting 50 includes:Storage medium 501, register cell 502 (preferably scalar data storage unit, scalar register unit),
Arithmetic element 503 (can also claim matrix operation unit 503) and control unit 504;
Storage medium 501, for storage matrix (or vector);
Scalar data storage unit 502, for storing scalar data, the scalar data includes at least:The vector exists
Storage address in the storage medium;
Control unit 504, for controlling the arithmetic element to obtain the first operational order, first operational order is used for
Realize the computing between vector and matrix, the vector that first operational order includes performing needed for described instruction reads instruction;
Arithmetic element 503, reading order is sent for reading instruction according to the vector to the storage medium;Foundation is adopted
The vector is read with batch reading manner and reads the corresponding vector of instruction, and first operational order is performed to the vector.
Optionally, above-mentioned vector reads instruction and includes:Vectorial storage address or described instruction institute needed for described instruction
Need the mark of vector.
Optionally during mark vectorial as needed for the vector reading is designated as described instruction,
Control unit 504, for controlling the arithmetic element to go out according to the mark from the register cell using single
Position reading manner reads the corresponding storage address of the mark, controls the arithmetic element to be sent to the storage medium and reads institute
State the reading order of storage address and the vector is obtained using batch reading manner.
Optionally, arithmetic element 503, specifically for the calculation using multistage pipelining-stage, institute is performed to the vector
State the first operational order.
Optionally, pre-set fixed arithmetic unit, Mei Geliu are included in each pipelining-stage in the multistage pipelining-stage
Fixation arithmetic unit in water level differs;
Arithmetic element 503, specifically for according to the corresponding calculating network topology of first operational order, utilizing K1Level
Selecting operation device in pipelining-stage to the vector be calculated first as a result, first result is input to K again2
Selecting operation device in level pipelining-stage perform be calculated second as a result, and so on, until the i-th -1 result is input to the
KjSelecting operation device in level pipelining-stage, which performs, is calculated i-th of result;I-th of result is inputted to the storage and is situated between
Matter is stored;
Wherein, KjBelong to any pipelining-stage in i pipelining-stage, j is less than or equal to i, and j and i are positive integer, described more
Quantity i, the selected execution sequence K of the multistage pipelining-stage of level pipelining-stagejAnd the KjSelection fortune in level pipelining-stage
It is to determine that the Selecting operation device is the fixed computing according to the calculating topological structure of first operational order to calculate device
Arithmetic unit in device.
Optionally, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage includes pre-set Matrix Multiplication
Method arithmetic unit, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix comparison operation device, third level stream
Water level includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First operational order is asked for vector
Instruction VDIER is led,
Arithmetic element 503, for inputting the vector to the nonlinear operator in third level pipelining-stage to the square
Battle array carries out matrix element local derviation and the first result is calculated;Second result is inputted to the storage medium and is stored.
Optionally, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage includes pre-set Matrix Multiplication
Method arithmetic unit, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix comparison operation device, third level stream
Water level includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First operational order is raw for vector
Diagonally battle array instruction VDIAG,
Arithmetic element 503, for the vector input to the matrix comparison operation device in the pipelining-stage of the second level to be carried out square
Battle array element address, which compares and writes corresponding vector element at interval of n+1 positions, obtains the first result;First result is defeated
Enter to the storage medium and stored.
Optionally, the multistage pipelining-stage is three-level pipelining-stage, and first order pipelining-stage includes pre-set Matrix Multiplication
Method arithmetic unit, second level pipelining-stage include pre-set addition of matrices arithmetic unit and matrix comparison operation device, third level stream
Water level includes pre-set nonlinear operator and matrix scalar multiplication arithmetic unit;First operational order multiplies for vector
Transposed matrix instructs VMULT,
Arithmetic element 503, for by it is described vector input to the matrix multiplication operation device in first order pipelining-stage carry out to
Amount, which multiplies, is calculated the first result;First result is inputted to the storage medium and is stored.
Optionally, the computing device further includes:
Buffer unit 505, for caching pending operational order;
Described control unit 504, for pending operational order to be cached in the buffer unit 504.
Optionally, control unit 504, for determining the before first operational order and first operational order
Two operational orders whether there is incidence relation, such as first operational order and second operational order there are incidence relation,
Then by first operational order caching with the buffer unit, after second operational order is finished, from described
Buffer unit extracts first operational order and is transmitted to the arithmetic element;
Described definite first operational order whether there is with the second operational order before the first operational order to be associated
System includes:
Required the first vectorial storage address section in first operational order is extracted according to first operational order,
Required the second vectorial storage address section in second operational order is extracted according to second operational order, such as described the
One storage address section has overlapping region with the second storage address section, it is determined that first operational order and institute
Stating the second operational order has incidence relation, and such as the first storage address section does not have with the second storage address section
Overlapping region, it is determined that first operational order does not have incidence relation with second operational order.
Optionally, above-mentioned control unit 503, can be used for obtaining operational order from instruction cache unit, and to the computing
After instruction is handled, there is provided to the arithmetic element.Wherein, control unit 503 can be divided into three modules, be respectively:
Fetching module 5031, decoding module 5032 and instruction queue module 5033,
Fetching module 5031, for obtaining operational order from instruction cache unit;
Decoding module 5032, for the operational order to acquisition into row decoding;
Instruction queue 5033, for after decoding operational order carry out sequential storage, it is contemplated that different instruction comprising
Register on there may exist dependence, for cache decode after instruction, launch after dependence is satisfied and refer to
Order.
Refering to Fig. 8, Fig. 8 is the flow chart that computing device provided by the embodiments of the present application performs operational order, such as Fig. 8 institutes
Show, the hardware configuration of the computing device is refering to the structure shown in Fig. 7, and storage medium as shown in Figure 7 is with scratchpad
Exemplified by, performing the process of vectorial derivation instruction VDIER includes:
Step S601, computing device control fetching module takes out vectorial derivation instruction, and vector derivation instruction is sent to
Decoding module.
Step S602, decoding module are sent to instruction queue to the vector derivation Instruction decoding, and by vector derivation instruction.
Step S603, in instruction queue, vector derivation instruction needs to obtain five in instruction from scalar register heap
The data in scalar register corresponding to a operation domain, the data include input vector X addresses, input vector X length, output
Matrix address, input vector Y address and input vector Y length.
Step S604, control unit determine whether are the vectorial derivation instruction and the operational order before vectorial derivation instruction
There are incidence relation, such as there are incidence relation, vectorial derivation instruction is deposited into buffer unit, associate management such as is not present, will
Vector derivation instruction is transmitted to arithmetic element.
Step S605, data in scalar register of the arithmetic element according to corresponding to five operation domains are from scratch pad memory
It is middle to take out the matrix data needed, vectorial derivative operation is then completed in arithmetic element.
Step S606, after the completion of arithmetic element computing, write the result into memory (preferable scratchpad or
Scalar register heap) specified address, reorder caching in the vector derivation instruct is submitted.
Optionally, in above-mentioned steps S605 when arithmetic element performs vectorial derivative operation, the computing device can use
Nonlinear operator performs matrix element local derviation and calculates to obtain corresponding Jacobian matrix.
In the specific implementation, after decoding module is to the vector derivation Instruction decoding, the control signal according to caused by decoding,
Vector input to the nonlinear operator in third level pipelining-stage acquired in S603 is performed into matrix using bypass circuit selection
Element local derviation is calculated first as a result, then would know that first result is to export result according to control signal.Accordingly
Ground, first result is write back as the output of arithmetic element, or is directly transferred to output terminal.
Operational order in above-mentioned Fig. 8 is by taking vectorial derivation instruction as an example, in practical applications, in embodiment as shown in Figure 8
Vectorial derivation instruction can multiply transposed matrix instruction VMULT equal matrix with vector generation diagonal matrix instruction VDIAG, vector and transport
Calculation/operational order is replaced, and is not repeated one by one here.
The embodiment of the present application also provides a kind of computer-readable storage medium, wherein, computer-readable storage medium storage is used for electricity
The computer program that subdata exchanges, it is any as described in above-mentioned embodiment of the method which make it that computer is performed
Implementation section or Overall Steps.
The embodiment of the present application also provides a kind of computer program product, and the computer program product includes storing calculating
The non-transient computer-readable recording medium of machine program, the computer program are operable to make computer perform such as above-mentioned side
Any implementation section or Overall Steps described in method embodiment.
The embodiment of the present application additionally provides a kind of accelerator, including:Memory:It is stored with executable instruction;Processor:
For performing the executable instruction in storage unit, in execute instruction according to the embodiment described in above method embodiment into
Row operation.
Wherein, processor can be single processing unit, but can also include two or more processing units.In addition,
Processor can also include general processor (CPU) or graphics processor (GPU);It is additionally may included in field programmable logic
Gate array (FPGA) or application-specific integrated circuit (ASIC), to be configured to neutral net and computing.Processor can also wrap
Include the on-chip memory for caching purposes (i.e. including the memory in processing unit).
In some embodiments, a kind of chip is also disclosed, is used to perform above method embodiment institute that includes above-mentioned
Corresponding neural network processor.
In some embodiments, a kind of chip-packaging structure is disclosed, that includes said chip.
In some embodiments, a kind of board is disclosed, that includes said chip encapsulating structure.
In some embodiments, a kind of electronic equipment is disclosed, that includes above-mentioned board.
Electronic equipment include data processing equipment, robot, computer, printer, scanner, tablet computer, intelligent terminal,
Mobile phone, automobile data recorder, navigator, sensor, camera, server, cloud server, camera, video camera, projecting apparatus, hand
Table, earphone, mobile storage, wearable device, the vehicles, household electrical appliance, and/or Medical Devices.
The vehicles include aircraft, steamer and/or vehicle;The household electrical appliance include TV, air-conditioning, micro-wave oven,
Refrigerator, electric cooker, humidifier, washing machine, electric light, gas-cooker, kitchen ventilator;The Medical Devices include Nuclear Magnetic Resonance, B ultrasound instrument
And/or electrocardiograph.
It should be noted that for foregoing each method embodiment, in order to be briefly described, therefore it is all expressed as a series of
Combination of actions, but those skilled in the art should know, the application and from the limitation of described sequence of movement because
According to the application, some steps can use other orders or be carried out at the same time.Secondly, those skilled in the art should also know
Know, embodiment described in this description belongs to alternative embodiment, involved action and module not necessarily the application
It is necessary.
In the above-described embodiments, the description to each embodiment all emphasizes particularly on different fields, and does not have the portion being described in detail in some embodiment
Point, it may refer to the associated description of other embodiment.
In several embodiments provided herein, it should be understood that disclosed device, can be by another way
Realize.For example, device embodiment described above is only schematical, such as the division of the unit, it is only one kind
Division of logic function, can there is an other dividing mode when actually realizing, such as multiple units or component can combine or can
To be integrated into another system, or some features can be ignored, or not perform.Another, shown or discussed is mutual
Coupling, direct-coupling or communication connection can be by some interfaces, the INDIRECT COUPLING or communication connection of device or unit,
Can be electrical or other forms.
The unit illustrated as separating component may or may not be physically separate, be shown as unit
The component shown may or may not be physical location, you can with positioned at a place, or can also be distributed to multiple
In network unit.Some or all of unit therein can be selected to realize the mesh of this embodiment scheme according to the actual needs
's.
In addition, each functional unit in each embodiment of the application can be integrated in a processing unit, can also
That unit is individually physically present, can also two or more units integrate in a unit.Above-mentioned integrated list
Member can both be realized in the form of hardware, can also be realized in the form of software program module.
If the integrated unit is realized in the form of software program module and is used as independent production marketing or use
When, it can be stored in a computer-readable access to memory.Based on such understanding, the technical solution of the application substantially or
Person say the part to contribute to the prior art or the technical solution all or part can in the form of software product body
Reveal and, which is stored in a memory, including some instructions are used so that a computer equipment
(can be personal computer, server or network equipment etc.) performs all or part of each embodiment the method for the application
Step.And foregoing memory includes:USB flash disk, read-only storage (ROM, Read-Only Memory), random access memory
(RAM, Random Access Memory), mobile hard disk, magnetic disc or CD etc. are various can be with the medium of store program codes.
One of ordinary skill in the art will appreciate that all or part of step in the various methods of above-described embodiment is can
To instruct relevant hardware to complete by program, which can be stored in a computer-readable memory, memory
It can include:Flash disk, read-only storage (English:Read-Only Memory, referred to as:ROM), random access device (English:
Random Access Memory, referred to as:RAM), disk or CD etc..
The embodiment of the present application is described in detail above, specific case used herein to the principle of the application and
Embodiment is set forth, and the explanation of above example is only intended to help to understand the present processes and its core concept;
Meanwhile for those of ordinary skill in the art, according to the thought of the application, can in specific embodiments and applications
There is change part, in conclusion this specification content should not be construed as the limitation to the application.