CN108614703A - Algorithm implant system based on embedded platform and its algorithm transplantation method - Google Patents
Algorithm implant system based on embedded platform and its algorithm transplantation method Download PDFInfo
- Publication number
- CN108614703A CN108614703A CN201611256319.6A CN201611256319A CN108614703A CN 108614703 A CN108614703 A CN 108614703A CN 201611256319 A CN201611256319 A CN 201611256319A CN 108614703 A CN108614703 A CN 108614703A
- Authority
- CN
- China
- Prior art keywords
- algorithm
- platform
- multinuclear
- implant system
- flow
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/70—Software maintenance or management
- G06F8/76—Adapting program code to run in a different environment; Porting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/40—Transformation of program code
- G06F8/53—Decompilation; Disassembly
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/60—Software deployment
- G06F8/61—Installation
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Prostheses (AREA)
Abstract
One algorithm implant system and its algorithm transplantation method based on embedded platform, wherein the algorithm implant system includes an acquisition assessment unit, for obtaining assessment algorithm；One algorithm flow adjustment unit is used for adjustment algorithm flow；One multinuclear allocation unit, for handling algorithm flow distribution multinuclear；One framework integration unit, treated for will pass through multinuclear, and the algorithm flow carries out framework integration；With a typing unit, it is used for the algorithm typing embedded platform, to which the algorithm that will be designed based on the ends PC is transplanted in the embedded platform.
Description
Technical field
The present invention relates to algorithm transplanting, further, are related to an algorithm implant system and its calculation based on embedded platform
Method transplantation method.
Background technology
Nowadays, the application of embedded mobile device is more and more extensive, therefore more and more PC (Personal
Computer, personal computer) algorithm also is intended to be transplanted and applied by the advantage of embedded platform.
CISC (Complex Instruction Set Computer, complicated order computer) and RISC (Reduced
Instruction Set Computer, Reduced Instruction Set Computer) be existing CPU two kinds of frameworks, due to its design reason
The difference with method is read, there is its respective advantage, pros and cons and application range.
The core of each microprocessors of CISC is the circuit of operating instruction.Instruction is made of multiple steps of completion task,
Numerical value transfer register or carry out sum operation.The instruction system of CISC is relatively abundanter, and there have special instruction to be specific to complete
Function, therefore, most PC are all to use CISC frameworks, for big multiple utility program, more suitable at cisc computer end
Complete preliminary design.
The format of all instructions of RISC is all consistent, and the period of all instructions is also identical, and uses flowing water
Line technology.This mentality of designing all simplifies number of instructions and addressing system, achieves and is more prone to, and parallel instructions are held
Stroke degree is good, becomes the more efficient of compiler.But RISC is to the function that is of little use, is usually completed by combined command,
Therefore when realizing specific function on RSIC, efficiency may be relatively low.Therefore for the Basic Design of big multiple utility program be
CISC is carried out, rather than is carried out in RSIC, but when programming is stablized, and practical application is carried out, is needed program portable extremely
Equipment end, and these equipment are typically embedded into formula equipment, using RSIC mode frameworks.
Due to the difference of the theory of the Basic Design thinking of RISC and CISC, the algorithm based on CISC designs wants graft application
To RISC frameworks embedded device when, need basic procedure and DSP (the Digital Signal in algorithm
Processing, digital processing unit) multinuclear unlatching on need to be adjusted so that the RSIC based on embedded platform
The equipment of framework may be employed.
Invention content
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the method make it migrate to the insertion of the RISC frameworks of DSP multinuclears to being adjusted based on the algorithm that CISC is designed
Formula equipment.
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the feature according to algorithm to be transplanted is designed so that algorithm to be transplanted maximally utilizes multi-core resource.
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the method carry out systematization integration to algorithm flow frame.
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the method may be used various ways and further optimized to algorithm after carrying out framework integration,
Operational efficiency of the algorithm that raising is transplanted in embedded end.
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the algorithm transplanting mode is suitable for the application of in the transplanting of the algorithm of three dimensional detection technology.
The algorithm implant system and its algorithm transplanting side that it is an object of the present invention to provide one based on embedded platform
Method, wherein the algorithm transplantation method is suitable for the application of in the transplanting of 3D structure light algorithms, by the 3D structure light algorithm flows at the ends PC
Migrate to the embedded platform of the RISC frameworks of DSP multinuclears.
In order to realize that the above at least goal of the invention, the present invention provide an algorithm implant system based on embedded platform,
It includes：One obtains assessment unit, for obtaining assessment algorithm；One algorithm flow adjustment unit is used for adjustment algorithm flow；One
Multinuclear allocation unit, for handling algorithm flow distribution multinuclear；One framework integration unit, for multinuclear will to be passed through
Treated, and the algorithm flow carries out framework integration；With a typing unit, for by the algorithm typing embedded platform, from
And the algorithm is transplanted in the embedded platform.
Assessment unit is obtained according to some embodiments, described in the algorithm implant system after obtaining the algorithm,
It is assessed in such a way that assembler language is into the maximized Cycle calculating of line frequency on DSP core.
According to some embodiments, the performance of algorithm described in algorithm evaluation unit evaluation described in the algorithm implant system
Whether reach expected standard, when reaching standard, can into be advanced into next step；When assessment is not up to expected standard, carry out
Theory of algorithm optimizes, and is assessed again.
According to some embodiments, algorithm flow adjustment unit described in the algorithm implant system judge whether to need into
The adjustment of every trade process flow, when judgement needs, into when the adjustment of every trade process flow, flow adjustment to be carried out to the algorithm；When
Judge to adjust the algorithm flow without flow into when the adjustment of every trade process flow, and directly opens multinuclear
Processing.
According to some embodiments, when the algorithm flow adjustment unit adjustment algorithm flow in the algorithm implant system
When the algorithm flow is divided into platform framework part and no platform frame part.
According to some embodiments, multinuclear allocation unit is to the platform framework part described in the algorithm implant system
Open multinuclear processing.
According to some embodiments, multinuclear allocation unit is to the no platform frame section described in the algorithm implant system
Divide and judge whether to need to open multinuclear, when judging that multinuclear need not be opened, distributes monokaryon；When judging to need to open multinuclear,
Distribute multinuclear.
According to some embodiments, multinuclear allocation unit is to the no platform frame section described in the algorithm implant system
The part for opening multinuclear is needed to judge adhesive size in the ranks in point, when in the ranks adhesive is big, using horizontal segmentation, when in the ranks
Adhesive hour, vertical segmentation.
According to some embodiments, the algorithm implant system includes a Memory adjustments unit, the Memory adjustments unit
For adjusting Memory Allocation after the multinuclear allocation processing in the algorithm, to improve operation efficiency.
According to some embodiments, Memory adjustments unit described in the algorithm implant system carries out the DSP memories
It overflows and judges, when judging to overflow, the memory of DSP core is divided into code segment and data segment, and it is low that code segment is linked to rate
Application heap, part data segment is linked to the high application heap of rate.
According to some embodiments, framework integration unit piece described in the algorithm implant system will pass through at DSP multinuclears
The algorithm after reason carries out framework integration on RISC control cores.
According to some embodiments, the algorithm implant system includes an algorithm optimization unit, the algorithm optimization unit
For optimizing the algorithm after the framework integration unit is integrated.
According to some embodiments, the optimization method of algorithm optimization unit described in the algorithm implant system is selected from following
Method：The calculating of correlation is placed on as possible under multinuclear and is handled；The time sequence spacing of thread is multiplexed；Utilize bottom language
Speech carries out the optimization of register level；It is pre-processed in the way of inline in the compilation phase；By the partial software filter in algorithm
Wave operator is substituted for hardware filtering.
According to some embodiments, the algorithm implant system includes an operation assessment unit, for assessing the algorithm
It is transplanted to the operational effect after the embedded platform, after assessment passes through, the typing unit will be described in the algorithm typing
Embedded platform.
According to some embodiments, typing unit described in the algorithm implant system carries out flash file using JTAG
It is burned onto the embedded platform.
According to some embodiments, algorithm described in the algorithm implant system is 3D structure light algorithms.
Another aspect of the present invention provides an algorithm transplantation method based on embedded platform comprising step：
(A) assessment algorithm；
(B) DSP is distributed；
(C) conformable frame；With
(D) to algorithm described in embedded platform typing.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein packet in the step (A)
Include step：After acquisition algorithm, carried out in such a way that assembler language is into the maximized Cycle calculating of line frequency on DSP core
Assessment.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein packet in the step (A)
Include step：Whether the performance for assessing the algorithm reaches expected standard, when reaching standard, can into be advanced into next step；When
When assessment is not up to expected standard, theory of algorithm optimization is carried out, is assessed again.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the step (B) includes
Step：Judge whether to need the adjustment into every trade process flow, when judging to need into when the adjustment of every trade process flow, to described
Algorithm carries out flow adjustment；When judge need not be into when the adjustment of every trade process flow, to the algorithm flow without flow
Adjustment, and directly open multinuclear processing.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the step (B) includes
Step：Judge whether to utilize the platform framework of the embedded platform will when judging that the platform framework can be utilized
The algorithm flow step is subdivided into platform framework part；When judging that the platform framework of the embedded platform cannot be utilized,
The algorithm flow step is subdivided into no platform frame part.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the step (B) is into one
Step includes step：Multinuclear processing is opened to the platform framework part.
According to some embodiments, the algorithm implant system based on embedded platform, wherein the step (B) is into one
Step includes step：The no platform frame part is judged whether to need to open multinuclear, when judging that multinuclear need not be opened, point
With monokaryon；When judging to need to open multinuclear, multinuclear is distributed.
According to some embodiments, the algorithm implant system based on embedded platform, wherein the algorithm is tied for 3D
Structure light algorithm, wherein normalization step distribute monokaryon processing, and multinuclear is opened in characteristic point filtering and the filtering of Node points.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the step (B) is into one
Step includes step：To needing the part for opening multinuclear to judge adhesive size in the ranks in the no platform frame part, when in the ranks
When adhesive is big, using horizontal segmentation, when in the ranks adhesive is small, vertical segmentation.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the algorithm is tied for 3D
The filtering of structure light algorithm, wherein characteristic point uses horizontal segmentation, the filtering of Node points to use vertical segmentation.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein the algorithm is tied for 3D
The processing of structure light algorithm, wherein picture signal, smothing filtering and convolutional filtering are subdivided into platform framework part；The normalization,
Characteristic point filters and the filtering of Node points is subdivided into no platform frame part.
According to some embodiments, described is the algorithm transplantation method of platform based on insertion, wherein the step (B) it
Include step (E) afterwards：Adjust DSP multi-core internal memories.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein right in the step (E)
The DSP memories carry out spilling judgement, and when judging to overflow, the memory of DSP core is divided into code segment and data segment, and by code
Section is linked to the low application heap of rate, and part data segment is linked to the high application heap of rate.
According to some embodiments, the algorithm transplantation method based on embedded platform, wherein in the step (C),
DSP multinuclears treated the algorithm will be passed through and carry out framework integration on RISC control cores.
According to some embodiments, the algorithm implant system based on embedded platform, wherein after the step (D)
Including step (F)：Optimize the algorithm.
According to some embodiments, described is the algorithm implant system based on embedded platform, wherein the step (F) is excellent
Change method is selected from following methods：The calculating of correlation is placed on as possible under multinuclear and is handled；The time sequence spacing of thread is carried out
Multiplexing；The optimization of register level is carried out using substrate；It is pre-processed in the way of inline in the compilation phase；By algorithm
In partial software filter operator be substituted for hardware filtering.
It is described based on Embedded algorithm transplantation method according to some embodiments, wherein including before the step (D)
Step：Transplantation effect is assessed, when assessment passes through, executes step (D).
It is described based on Embedded algorithm transplantation method according to some embodiments, wherein utilizing JTAG in the step (D)
It carries out flash file and is burned onto the embedded platform.
According to some embodiments, described based on Embedded algorithm transplantation method, wherein algorithm is 3D structure light algorithms.
Another aspect of the present invention provides a 3D structure light algorithm transplantation methods, wherein the 3D structure lights algorithm flow packet
Smothing filtering is included, is normalized, characteristic point filtering, Node points filtering, wherein the 3D structure lights algorithm passes through the aforementioned calculation
Method transplantation method migrates to embedded platform.
Description of the drawings
Fig. 1 is the algorithm implant system based on embedded platform of first preferred embodiment according to the present invention.
Fig. 2 is that the algorithm implant system based on embedded platform of first preferred embodiment according to the present invention executes stream
Journey schematic diagram.
Fig. 3 is the algorithm transplantation method block diagram based on embedded platform of second preferred embodiment according to the present invention.
Fig. 4 is the algorithm transplantation method flow based on embedded platform of second preferred embodiment according to the present invention
Figure.
Fig. 5 is a kind of 3D based on embedded platform of specific implementation mode of above preferred embodiment according to the present invention
The flow chart of 3D structure lights in structure light algorithm transplantation method at the ends PC.
Fig. 6 is the 3D structures based on embedded platform of the specific implementation mode of above preferred embodiment according to the present invention
Light algorithm transplantation method flow chart.
Specific implementation mode
It is described below for disclosing the present invention so that those skilled in the art can realize the present invention.It is excellent in being described below
Embodiment is selected to be only used as illustrating, it may occur to persons skilled in the art that other obvious modifications.It defines in the following description
The present invention basic principle can be applied to other embodiments, deformation scheme, improvement project, equivalent program and do not carry on the back
Other technologies scheme from the spirit and scope of the present invention.
It will be understood by those skilled in the art that the present invention exposure in, term " longitudinal direction ", " transverse direction ", "upper",
The orientation of the instructions such as "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom" "inner", "outside" or position are closed
System is to be based on the orientation or positional relationship shown in the drawings, and is merely for convenience of description of the present invention and simplification of the description, without referring to
Show or imply that signified device or element must have a particular orientation, with specific azimuth configuration and operation, therefore above-mentioned art
Language is not considered as limiting the invention.
Referring to Figures 1 and 2, the present invention provides an algorithm implant system 100 based on embedded platform, wherein the algorithm
The algorithm for designing or running at the ends PC can be migrated to an embedded platform by implant system 100, for example a mobile device is embedding
Enter formula platform, so that the algorithm can be run by the embedded platform, so as to be put down using embedded
The advantage of platform, or using the advantage of mobile device, extend the application range of the algorithm.Particularly, described based on insertion
The algorithm implant system 100 of formula platform be suitable for transplant predetermined frame under algorithm, citing ground but be not limited to, multi-core DSP, RISC.
The algorithm implant system 100 can be applied to the depth algorithm transplanting of the applications such as machine vision, intelligent robot, binocular sensing
In the process.
Specifically, according to this embodiment of the invention, the algorithm implant system 100 based on embedded platform includes
One obtains assessment unit 101, and the acquisition assessment unit 101 is used for acquisition algorithm, assesses algorithm performance.Namely
It says, it is described to obtain whether the algorithm that assessment unit 101 is got for entry evaluation apply the algorithm implant system
100。
Further, the assessment for obtaining assessment unit 101 and carrying out algorithm performance on DSP core, for example utilize compilation language
It says and is calculated into the maximized Cycle of line frequency (clock cycle).
Further, the evaluation criteria for obtaining assessment unit 101 after acquisition algorithm is：If the acquisition assessment is single
Member 101 assessment performances reach expected standard far away, then algorithm also need in theory optimize adjustment；If the acquisition
The performance that assessment unit 101 is assessed reaches expected standard, then can carry out in next step.
The embedded platform often presses what row carried out due to the limitation of hardware resource, when handling image, and PC algorithms
In some steps, often press what frame carried out, such as in traversal image when seeking extreme value to each pixel.Therefore, according to this hair
This bright embodiment, the algorithm implant system 100 based on embedded platform include an algorithm flow adjustment unit 102,
The algorithm that the algorithm flow adjustment unit 102 is used to need to be adjusted carries out the adjustment of flow, by the algorithm in PC
The flow of end design is adjusted to the flow suitable for the embedded platform.Such as the flow tune by frame operation that will be designed at the ends PC
The whole flow carried out by row for suitable for embedded platform.Certainly, the flow itself designed at the ends PC when the algorithm with it is embedded
When the process requirements of formula platform are consistent, then the adjustment of flow need not be carried out, can carried out in next step.
Certainly, before adjustment, the needs of algorithm flow adjustment unit 102 judge whether to need into every trade process flow
Adjustment, when judge obtain needing into when the adjustment of every trade process flow, then carrying out algorithm flow adjustment；When judgement is not required to
It then to be adjusted without algorithm flow into when the adjustment of every trade process flow, can directly utilize embedded platform frame, opened
The processing of DSP multinuclears.That is, when the algorithm flow of acquisition meets the frame system of embedded platform, then without algorithm
Flow adjusts.
More specifically, the algorithm flow adjustment unit 102 adjustment mode citing ground but be not limited to, according to the insertion
The algorithm is carried out the flow on platform framework and non-flat table frame and adjusted by the characteristics of formula platform.The platform framework is i.e. described
The packaging frame of the existing row processing module of embedded platform.These frames often have processing speed fast, modularization adjustment
Convenient feature, but disadvantage is also obvious, that is, the degree of freedom adjusted to data processing method to user is too low, because
This needs to build non-flat table frame.
In other words, the algorithm can be adjusted to flat by the algorithm flow adjustment unit 102 during adjustment
Table frame and non-flat table frame two parts.That is, the algorithm consistent with the embedded platform frame is put into the platform framework
Part, and with the embedded platform frame it is inconsistent be put into no platform frame part, it is described flat so as to be directed to respectively
Table frame part and the no platform frame part are handled.
Further, the algorithm implant system 100 based on embedded platform includes a multinuclear allocation unit 103, described
Multinuclear allocation unit 103 is used to distribute the multinuclear processing mode of the DSP.Citing ground but be not limited to, the multinuclear allocation unit
The 103 distribution each step executed in algorithm flow is to carry out monokaryon processing or carry out multinuclear cooperation processing.In some feelings
Under condition, the task of the internuclear distribution of DSP is that oneself is completed inside platform framework, thus the platform framework part be suitable for into
The processing of row multinuclear rather than platform framework part then need specifically to be judged.
Further, when the no platform frame part, which is assigned, opens multinuclear, the multinuclear allocation unit 103 needs
Judge the direction of the multinuclear segmentation of the DSP, such as horizontal segmentation, vertical segmentation.When the algorithm process image in the ranks
When adhesive is big, image is needed by horizontal segmentation, and when the adhesive in the ranks of image is small, then it needs by vertical segmentation.
The adhesive, refer to the pixel of image upper and lower several rows between pixel correlation for showing in algorithm calculating.
It citing ground in some embodiments, will be described after the algorithm flow unit is adjusted the algorithm
Each step in algorithm flow is adjusted to platform framework part and no platform frame part, to more described in the platform framework part
Core allocation unit 103 can carry out the unlatching of multinuclear, that is to say, that handle the platform framework part by the multinuclear of DSP.And
Can multinuclear be opened for no platform frame part judgement, if cannot, it distributes monokaryon and writes operator；If can open,
Need to judge the direction that multinuclear separates at this time, the big distribution horizontal segmentation of the adhesive of pixel, the adhesive of pixel is small
Distribute vertical segmentation.And by this method so that the DSP multi-core resources utilize maximization.
Further, the algorithm implant system 100 based on embedded platform includes a Memory adjustments unit 104, described
Memory adjustments unit 104 is for adjusting application heap.More specifically, the Memory adjustments unit 104 overflows to the DSP memories
Go out to judge, overflow if judging that memory exists, the memory of the DSP core can be used code by the Memory adjustments unit 104
The mode that section is detached with data segment adjusts, and code segment is linked in the low application heap of rate, and data segment is linked to rate
In high application heap；When judging memory there is no overflowing, can directly carry out in next step.At runtime due to algorithm, code
Section is typically to load once, and data segment is then constantly to be read out and refreshing, therefore the inside adjustment unit will be calculated
Method is distributed to different application heaps, so as to more efficiently utilize memory.
Further, the algorithm implant system 100 based on embedded platform includes a framework integration unit 105, described
Framework integration unit 105 is used to integrate the interface of the algorithm flow frame on RISC control cores.
In other words, the algorithm flow after algorithm flow unit adjustment, passes through the framework integration unit 105
It is integrated, is carried out after being handled by DSP respectively than platform framework part as will be described and the no platform frame part whole
It closes, RISC frames is integrated in the algorithm flow frame system to realize.
Further, the algorithm implant system 100 based on embedded platform includes an algorithm optimization unit 106, described
Algorithm optimization unit 106 is used to carry out the algorithm optimization of details, in order to further increase the algorithm described embedding
Enter the operational efficiency of formula platform framework.
Specifically, it the algorithm optimization unit 106 citing ground but is not limited to following methods may be used and optimizes：
1, the calculating of correlation is placed on as possible under multinuclear and is handled.
2, the time sequence spacing of thread is multiplexed.
3, the optimization of register level is carried out (as collected) using substrate.
4, it is pre-processed in the way of inline etc. in the compilation phase.
5, the partial software filter operator in algorithm is substituted for hardware filtering.
Further, the algorithm implant system 100 based on embedded platform includes an operation assessment unit 107, described
Operation assessment unit 107 is transplanted for running the assessment algorithm to the algorithm operational effect after the embedded platform.
In other words, whether the operation assessment unit 107 is used to test and assess the algorithm by successful implantation and transplanting effect
Fruit.
Further, the algorithm implant system 100 based on embedded platform includes a typing unit 108, the typing
Unit 108 is for will be transplanted embedded platform host computer described in the algorithm typing.For example carry out flash texts using JTAG
Part burning.
With reference to Fig. 3 and Fig. 4, above preferred embodiment according to the present invention, the present invention provides one based on embedded platform
Algorithm transplantation method 1000, the algorithm transplantation method 1000 are suitable for migrating to the algorithm that the ends PC are designed or run embedded flat
Platform, the algorithm transplantation method 1000 can be applied to the depth algorithm of the applications such as machine vision, intelligent robot, binocular sensing
In migration process.The algorithm transplantation method 1000 includes the following steps：
1001：Assessment algorithm；
1002：Adjustment algorithm flow；
1003：Distribute DSP；
1004：Adjust memory；
1005：Conformable frame；
1006：Optimization algorithm；
1007：Assess transplantation effect；With
1008：To embedded platform typing algorithm.
Wherein, in the step 1001, the assessment of performance is carried out on DSP core, for example utilize assembler language into line frequency
Rate maximizes Cycle (clock cycle) and calculates.
The standard assessed in the step 1001 is：When expected standard is much not achieved in the performance of assessment algorithm
When, adjustment is optimized to the algorithm；When the performance of algorithm reaches standard, step 1002 is executed.That is, when assessment
As a result when not up to expected, return is needed to optimize algorithm, then assessed, successively until the performance for being optimized to algorithm reaches
When to standard, step 1002 is executed.
Embedded platform is due to the limitation of hardware resource, often by row progress when handling image, and in PC algorithms
Some steps, generally require to carry out by frame, for example each pixel in traversal image exists not between the two when seeking extreme value
Unanimously, it is therefore desirable to which algorithm flow is adjusted.In the step 1002, the algorithm flow is adjusted so that
The algorithm at the ends PC is suitable for the execution flow of embedded platform.Before adjustment, it needs first to judge whether to need into every trade processing stream
The adjustment of journey, when judgement need not be opened into when the adjustment of every trade process flow, can directly use embedded platform frame
The processing of DSP multinuclears；When judgement needs, into when the adjustment of every trade process flow, to carry out the adjustment of algorithm flow.
Specifically, when being adjusted, the flow of the algorithm is divided by platform framework portion according to embedded platform feature
Divide and no platform frame part.The packaging frame of the existing row processing module of the platform framework, that is, embedded platform.
These frames often have processing speed fast, modularization feature easy to adjust, but disadvantage is also obvious, that is, give
The degree of freedom that user adjusts data processing method is too low, it is therefore desirable to build non-flat table frame.
In other words, during adjustment, the algorithm can be adjusted to platform framework and non-flat table frame two
Point.That is, the algorithm consistent with the embedded platform frame is put into the platform framework part, and with the embedded platform
Frame it is inconsistent be put into no platform frame part, so as to respectively be directed to the platform framework part and the no platform frame
Frame part is handled.
The step 1002 includes step 10021 as a result,：Judge whether to the adjustment of row process flow.
The step 10021 further comprises step：
100211：When judging to obtain needing carrying out the adjustment of process flow, algorithm flow adjustment is carried out；
100212：When judgement obtains that, using platform framework, it is more DSP need not be opened into when the adjustment of every trade process flow
Core.
Wherein further comprise step in the step 100211：
1002111：Judging whether can be using the platform framework of the embedded platform；
1002112：When judging that the platform framework can be utilized, the algorithm flow step is subdivided into platform framework
Part；
1002113：When judging that the platform framework of the embedded platform cannot be utilized, the algorithm flow step is drawn
It is divided into no platform frame part.
In the step 1003, the multinuclear processing mode of multi-core DSP is distributed, such as to step in the algorithm flow point
It is still handled using multinuclear with being handled using monokaryon, and what dividing mode is constantly used using multinuclear.
In some cases, the task of the internuclear distribution of DSP is oneself completion inside platform framework, therefore described flat
Table frame part is suitable for needing then further can specifically being judged by the processing of carry out multinuclear rather than platform framework part.
Further, when the no platform frame part, which is assigned, opens multinuclear, the multinuclear point for judging the DSP is needed
The direction cut, such as horizontal segmentation, vertical segmentation, when the adhesive in the ranks of the image of the algorithm process is big, image
It needs by horizontal segmentation, and when the adhesive in the ranks of image is small, then it needs by vertical segmentation.The adhesive refers to image
Pixel upper and lower several rows between pixel correlation for showing in algorithm calculating.
Citing ground, in some embodiments, after being adjusted to the algorithm, by each step in the algorithm flow
It is adjusted to platform framework part and no platform frame part, the unlatching of multinuclear can be carried out to the platform framework part, also
It is to say, the platform framework part is handled by the multinuclear cooperation of DSP.And can no platform frame part judgement be opened
Multinuclear is opened, if cannot, it distributes monokaryon and writes operator；If can open, need to judge the direction that multinuclear separates, pixel at this time
The big distribution horizontal segmentation of adhesive, the small distribution vertical segmentation of the adhesive of pixel.And by this method so that institute
That states DSP multi-core resources utilizes maximization.
The step 1003 further comprises step as a result,：
10031：To the platform framework part, the processing of DSP multinuclears is opened；
10032：Judge that the no platform frame part multinuclear can be opened；
The step 10032 further comprises step：
100321：When judging that multinuclear cannot be opened, distribution is handled using monokaryon, and writes operator；
100323：When judging that multinuclear can be opened, further judge that the adhesive in the ranks of the algorithm process image is
It is no big；When the adhesive in the ranks for handling image is big, using horizontal segmentation；When the adhesive in the ranks of the image of processing is little,
Using vertical segmentation.
In the step 1004, spilling judgement is carried out to the DSP memories, if judging, obtaining memory has spilling,
The memory of the DSP core is adjusted by the way of code segment and data segment separation, code segment is linked to the low memory of rate
In section, and data segment is linked in the high application heap of rate；When judging that memory does not overflow, can directly carry out in next step.
At runtime due to algorithm, code segment is typically to load once, and data segment is then constantly to be read out and refreshing, therefore
Algorithm is distributed to different application heaps, so as to more efficiently utilize memory.
The step 1004 further comprises step as a result,：
10041：Judge whether memory overflows；
10042：When judging that obtaining memory overflows, the code segment of the algorithm of the inside and data terminal are separately linked；
In the step 10042, code segment is linked in the low application heap of rate, it is high that data segment is linked to rate
In application heap.
In the step 1005, the interface of the algorithm flow frame is integrated on RISC control cores.For example, will
The platform framework part and the no platform frame part are integrated after being handled by DSP respectively, to realize pair
The algorithm flow frame system is integrated in RISC frames.
In the step 1006, the optimization of details is carried out to the algorithm, in order to further increase the algorithm in institute
State the operational efficiency of embedded platform frame.
Specifically, following methods may be used to optimize：
1, the calculating of correlation is placed on as possible under multinuclear and is handled.
2, the time sequence spacing of thread is multiplexed.
3, the optimization of register level is carried out (as collected) using substrate.
4, it is pre-processed in the way of inline etc. in the compilation phase.
5, the partial software filter operator in algorithm is substituted for hardware filtering.
In the step 1007, operation is assessed the algorithm and is transplanted to the algorithm operation effect after the embedded platform
Fruit.In other words, whether the algorithm of testing and assessing is by successful implantation and transplantation effect.
In the step 1008, embedded platform host computer described in the algorithm typing will be transplanted.For example it utilizes
JTAG (chip interior test compiler) carries out flash (flash memory) burning file.
Fig. 5 is a kind of 3D based on embedded platform of specific implementation mode of above preferred embodiment according to the present invention
The flow chart of 3D structure lights in structure light algorithm transplantation method at the ends PC.Fig. 6 is above preferred embodiment according to the present invention
Specific implementation mode the 3D structure light algorithm transplantation method flow charts based on embedded platform.In other words, pass through Fig. 6 institutes
The 3D structure light algorithms realized at the ends PC in Fig. 5 are transplanted to embedded platform by the method shown.
Specifically, referring to figure 5 and figure 6, the present invention provides a 3D structure light algorithms transplantation method 2000, the 3D structure lights
Algorithm transplantation method 2000 is suitable for the 3D structure light algorithms that the ends PC are designed or run migrating to embedded platform.The 3D structures
Algorithm flow step of the light at the ends PC include：Image inputs, smothing filtering, convolution and normalization, characteristic point detection, characteristic point filter
Wave, characteristic point network connection, Node points calculate, Node points connection, according to arrange table look-up and image output.The wherein described normalizing
Change the maximin for seeking frame image, the characteristic point filtering needs multinuclear to divide, and cannot carry out monokaryon processing, in Node
When point calculates, original image imports again, performance loss.The wherein described characteristic point detection, the characteristic point filtering and the spy
Sign spot net connection, which can integrate, is characterized filtering, and the Node points calculating is connected with the Node points and can be integrated as Node
Point filtering.The 3D structure lights algorithm transplantation method 2000 includes the following steps：
2001：Assessment algorithm；
2002：Adjustment algorithm flow；
2003：Distribute DSP；
2004：Adjust memory；
2005：Conformable frame；
2006：Optimization algorithm；
2007：Assess transplantation effect；With
2008：To embedded platform typing algorithm.
Wherein, in the step 2001, the assessment of performance is carried out on DSP core, for example utilize assembler language into line frequency
Rate maximizes Cycle (clock cycle) and calculates.
More specifically, in the step 2001, the 3D structure light algorithms in Fig. 5 are obtained, to the 3D structure lights algorithm
It is assessed.
The standard assessed in the step 2001 is：When the performance for assessing the 3D structure lights algorithm much reaches not
When to expected standard, adjustment is optimized to the 3D structure lights algorithm；When the performance of the 3D structure lights algorithm reaches standard
When, execute step 2002.That is, when assessment result is not up to expected, need to return to the 3D structure lights algorithm into
Row optimization, then assessed, successively when the performance for being optimized to the 3D structure lights algorithm reaches standard, execute step
2002。
Embedded platform is due to the limitation of hardware resource, often by row progress when handling image, and in PC algorithms
Some steps, generally require to carry out by frame, for example each pixel in traversal image exists not between the two when seeking extreme value
Unanimously, it is therefore desirable to which algorithm flow is adjusted.In the step 2002, the 3D structure lights algorithm flow is adjusted
It is whole so that the algorithm at the ends PC is suitable for the execution flow of embedded platform.Before adjustment, it needs first to judge whether to need into every trade
The adjustment of process flow, when judge can need not directly use embedded platform frame into when the adjustment of every trade process flow,
Open the processing of DSP multinuclears；When judgement needs, into when the adjustment of every trade process flow, to carry out the adjustment of algorithm flow.
Specifically, when being adjusted, the flow of the 3D structure lights algorithm is divided into according to embedded platform feature flat
Table frame part and no platform frame part.The envelope of the existing row processing module of the platform framework, that is, embedded platform
Frame up frame.These frames often have processing speed fast, modularization feature easy to adjust, but disadvantage is also apparent
, that is, the degree of freedom adjusted to data processing method to user is too low, it is therefore desirable to build non-flat table frame.
In other words, during adjustment, the 3D structure lights algorithm flow step can be adjusted to platform framework
With non-flat table frame two parts.That is, the algorithm consistent with the embedded platform frame is put into the platform framework part, and
With the embedded platform frame it is inconsistent be put into no platform frame part, so as to respectively be directed to the platform framework portion
Divide and the no platform frame part is handled.
Specifically, by Fig. 5 it can be seen that, the 3D structure lights algorithm in normalization step, need to full frame image into
Row processing, therefore need to be adjusted correspondingly when the embedded platform is handled.This implementation according to the present invention
Example method, the 3D structure lights algorithm flow after adjustment are divided into platform framework part, including：Picture signal processing, it is smooth to filter
Wave, convolutional filtering；With no platform frame part, including：Normalization, characteristic point filtering, the filtering of Node points.The characteristic point filtering
Including：Characteristic point detects, characteristic point filtering, characteristic point network connection, wherein characteristic point filtering can only carry out monokaryon processing.
Node is filtered：Node points calculate, the connection of Node points.
The step 2002 includes step 20021 as a result,：Judge whether to the adjustment of row process flow.
It further comprises step：
200211：When judging to obtain needing carrying out the adjustment of process flow, the adjustment of 3D structure light algorithm flows is carried out；
200212：When judgement obtains that, using platform framework, it is more DSP need not be opened into when the adjustment of every trade process flow
Core.
Wherein further comprise step in the step 200211：
2002111：Judging whether can be using the platform framework of the embedded platform；
2002112：When judging that the platform framework can be utilized, the algorithm flow is subdivided into platform framework portion
Point；
2002113：When judging that the platform framework of the embedded platform cannot be utilized, the algorithm flow is subdivided into
No platform frame part.
In the step 2002111, the foundation of judgement is to meet predetermined frame when the step of 3D structure lights algorithm
When structure, the embedded platform frame can be utilized, and when the step of 3D structure lights algorithm does not meet scheduled frame
When structure, the embedded platform frame cannot be utilized.
That is, in the method for this embodiment of the present invention, in the step 2002112, judgement obtains institute
The described image signal processing in 3D structure light algorithm steps, smothing filtering are stated, convolutional filtering can utilize described embedded flat
The platform framework of platform is subdivided into the platform framework part.In the step 2002113, judgement obtains the 3D structure lights
Normalization, characteristic point filtering and the filtering of Node points of algorithm cannot utilize the platform framework of the embedded platform, cut-in non-flat
Table frame part.
In the step 2003, the multinuclear processing mode of multi-core DSP is distributed, such as to the 3D structure lights algorithm flow
Middle step distribution is handled using monokaryon or is handled using multinuclear, and what dividing mode constantly used using multinuclear, vertically
Partitioning scheme or horizontal segmentation mode.
In some cases, the task of the internuclear distribution of DSP is oneself completion inside platform framework, therefore described flat
Table frame part is suitable for needing then further can specifically being judged by the processing of carry out multinuclear rather than platform framework part.
Further, when the no platform frame part, which is assigned, opens multinuclear, the multinuclear point for judging the DSP is needed
The direction cut, such as horizontal segmentation, vertical segmentation, when the adhesive in the ranks of the image of the 3D structure lights algorithm process is big
It waits, image needs by horizontal segmentation, and when the adhesive in the ranks of image is small, then it needs by vertical segmentation.The adhesive is
The correlation that pixel shows in algorithm calculating between upper and lower several rows of the pixel of finger image.
After being adjusted to the 3D structure lights algorithm, the 3D structure lights are calculated in some embodiments on citing ground
Each step in method flow is adjusted to platform framework part and no platform frame part, can be carried out to the platform framework part
The unlatching of multinuclear, that is to say, that the platform framework part is handled by the multinuclear cooperation of DSP.And for the no platform frame
Can the judgement of frame part open multinuclear, if cannot, it distributes monokaryon and writes operator；If can open, need to judge multinuclear at this time
The direction of separation, the big distribution horizontal segmentation of the adhesive of pixel, the small distribution vertical segmentation of the adhesive of pixel.And it is logical
Cross this method so that the DSP multi-core resources utilize maximization.
The step 2003 further comprises step as a result,：
20031：To the platform framework part, the processing of DSP multinuclears is opened；
Specifically, in the step 10031, to the platform framework part in the 3D structure lights algorithm, i.e. institute
The processing of image model, the smothing filtering are stated, the convolutional filtering opens the processing of DSP multinuclears.
20032：Judge that the no platform frame part multinuclear can be opened；
The step 20032 basis for estimation citing ground but be not limited to, the maximum of the image pixel by seeking algorithm process,
Minimum value judges correlation size, that is, judges the influence size of mutual handling result.When correlation is big, cannot open
Multinuclear can open multinuclear when correlation is small.Citing ground, when the latter handling result depends on previous handling result,
Correlation is big.When the latter processing structure is independent of previous handling result, correlation is small.It is of course also possible to use its
His basis for estimation.
The step 20032 further comprises step：
200321：When judging that multinuclear cannot be opened, distribution is handled using monokaryon, and writes operator；
200322：When judging that multinuclear can be opened, the 3D structure lights algorithm process image is further judged in the ranks
Whether adhesive is big；When the adhesive in the ranks for handling image is big, using horizontal segmentation；When the adhesive in the ranks of the image of processing
When little, using vertical segmentation.
Specifically, the no platform frame part after 3D structure lights algorithm adjustment, that is, the normalization, institute
Characteristic point filtering is stated, in Node points filtering, characteristic point filter step is due to the in the ranks adhesive of data when looking for neighbours' point
Greatly, it is therefore desirable to horizontal segmentation is carried out to it, and the filtering of Node points can then be divided in a vertical direction.
In the step 2004, spilling judgement is carried out to the DSP memories, if judging, obtaining memory has spilling,
The memory of the DSP core is adjusted by the way of code segment and data segment separation, code segment is linked to the low memory of rate
In section, and data segment is linked in the high application heap of rate；When judging that memory does not overflow, can directly carry out in next step.
At runtime due to algorithm, code segment is typically to load once, and data segment is then constantly to be read out and refreshing, therefore
Algorithm is distributed to different application heaps, so as to more efficiently utilize memory.
The step 2004 further comprises step as a result,：
20041：Judge whether memory overflows；
20042：When judging that obtaining memory overflows, the code segment of the algorithm of the inside and data terminal are separately linked；
In the step 20042, code segment is linked in the low application heap of rate, it is high that data segment is linked to rate
In application heap.
In the step 2005, by the interface of the 3D structure lights algorithm flow frame RISC control core on carry out it is whole
It closes.For example, the platform framework part and the no platform frame part are integrated after being handled by DSP respectively, from
And it realizes and RISC frames is integrated in the 3D structure lights algorithm flow frame system.
In the step 2006, the optimization of details is carried out to the 3D structure lights algorithm, it is described in order to further increase
Operational efficiency of the 3D structure lights algorithm in the embedded platform frame.
Specifically, following methods may be used to optimize：
1, the calculating of correlation is placed on as possible under multinuclear and is handled.
2, the time sequence spacing of thread is multiplexed.
3, the optimization of register level is carried out (as collected) using substrate.
4, it is pre-processed in the way of inline etc. in the compilation phase.
5, the partial software filter operator in algorithm is substituted for hardware filtering.
Specifically, in the 3D structure lights algorithm, by Fig. 5 it can be seen that, carry out Node points calculate when, due to original
Image data needs are imported again, and testing algorithm, which takes, to be increased, to reduce operational performance, in the shifting of the present invention
In plant method, processing under multinuclear, time sequence spacing multiplexing, assembly code optimizing, compiler pretreatment and hard are put using correlation calculations
The various ways such as part filtering advanced optimize the algorithm frame of embedded end, improve operational efficiency.
In the step 2007, operation is assessed the 3D structure lights algorithm and is transplanted to the calculation after the embedded platform
Method operational effect.In other words, whether the 3D structure lights algorithm is tested and assessed by successful implantation and transplantation effect.
In the step 2008, embedded platform host computer described in the 3D structure lights algorithm typing will be transplanted.
For example carry out flash (flash memory) burning file using (chip interior test compiler).
It should be understood by those skilled in the art that the embodiment of the present invention shown in foregoing description and attached drawing is only used as illustrating
And it is not intended to limit the present invention.The purpose of the present invention has been fully and effectively achieved.The function and structural principle of the present invention exists
It shows and illustrates in embodiment, under without departing from the principle, embodiments of the present invention can have any deformation or modification.
Claims (36)
1. an algorithm implant system, which is characterized in that including：
One obtains assessment unit, for obtaining assessment algorithm；
One algorithm flow adjustment unit is used for adjustment algorithm flow；
One multinuclear allocation unit, for handling algorithm flow distribution multinuclear；
One framework integration unit, treated for will pass through multinuclear, and the algorithm flow carries out framework integration；With
One typing unit is used for the algorithm typing embedded platform, to transplant the algorithm in described embedded flat
Platform.
2. algorithm implant system according to claim 1, wherein the acquisition assessment unit is after obtaining the algorithm,
It is assessed in such a way that assembler language is into the maximized Cycle calculating of line frequency on DSP core.
3. algorithm implant system according to claim 1, wherein the performance of algorithm described in the algorithm evaluation unit evaluation
Whether reach expected standard, when reaching standard, can into be advanced into next step；When assessment is not up to expected standard, carry out
Theory of algorithm optimizes, and is assessed again.
4. algorithm implant system according to claim 1, wherein the algorithm flow adjustment unit judge whether to need into
The adjustment of every trade process flow, when judgement needs, into when the adjustment of every trade process flow, flow adjustment to be carried out to the algorithm；When
Judge to adjust the algorithm flow without flow into when the adjustment of every trade process flow, and directly opens multinuclear
Processing.
5. algorithm implant system according to claim 1, wherein when the algorithm flow adjustment unit adjustment algorithm flow
When the algorithm flow is divided into platform framework part and no platform frame part.
6. algorithm implant system according to claim 5, wherein the multinuclear allocation unit is to the platform framework part
Open multinuclear processing.
7. algorithm implant system according to claim 5, wherein the multinuclear allocation unit is to the no platform frame section
Divide and judge whether to need to open multinuclear, when judging that multinuclear need not be opened, distributes monokaryon；When judging to need to open multinuclear,
Distribute multinuclear.
8. algorithm implant system according to claim 7, wherein the multinuclear allocation unit is to the no platform frame section
The part for opening multinuclear is needed to judge adhesive size in the ranks in point, when in the ranks adhesive is big, using horizontal segmentation, when in the ranks
Adhesive hour, vertical segmentation.
9. algorithm implant system according to claim 1, wherein the algorithm implant system includes a Memory adjustments unit,
The Memory adjustments unit is used to adjust Memory Allocation after the multinuclear allocation processing in the algorithm, to improve operation effect
Rate.
10. algorithm implant system according to claim 9, wherein the Memory adjustments unit carries out the DSP memories
It overflows and judges, when judging to overflow, the memory of DSP core is divided into code segment and data segment, and it is low that code segment is linked to rate
Application heap, part data segment is linked to the high application heap of rate.
11. algorithm implant system according to claim 1, wherein the framework integration unit piece will pass through at DSP multinuclears
The algorithm after reason carries out framework integration on RISC control cores.
12. algorithm implant system according to claim 1, wherein the algorithm implant system includes an algorithm optimization list
Member, the algorithm optimization unit are used to optimize the algorithm after the framework integration unit is integrated.
13. algorithm implant system according to claim 12, wherein the optimization method of the algorithm optimization unit be selected from
Lower method：The calculating of correlation is placed on as possible under multinuclear and is handled；The time sequence spacing of thread is multiplexed；Utilize bottom
Language carries out the optimization of register level；It is pre-processed in the way of inline in the compilation phase；By the partial software in algorithm
Filter operator is substituted for hardware filtering.
14. algorithm implant system according to claim 1, wherein the algorithm implant system includes that an operation assessment is single
Member is transplanted for assessing the algorithm to the operational effect after the embedded platform, after assessment passes through, the typing unit
By embedded platform described in the algorithm typing.
15. algorithm implant system according to claim 1, wherein the typing unit carries out flash file using JTAG
It is burned onto the embedded platform.
16. algorithm implant system according to any one of claims 1 to 15, wherein the algorithm is 3D structure light algorithms.
17. the algorithm transplantation method based on embedded platform, it is characterised in that including step：
(A) assessment algorithm；
(B) DSP is distributed；
(C) conformable frame；With
(D) to algorithm described in embedded platform typing.
18. the algorithm transplantation method according to claim 17 based on embedded platform, wherein the step (A) includes
Step：After acquisition algorithm, commented in such a way that assembler language is into the maximized Cycle calculating of line frequency on DSP core
Estimate.
19. the algorithm transplantation method according to claim 17 based on embedded platform, wherein the step (A) includes
Step：Whether the performance for assessing the algorithm reaches expected standard, when reaching standard, can into be advanced into next step；When commenting
When estimating not up to expected standard, theory of algorithm optimization is carried out, is assessed again.
20. the algorithm transplantation method according to claim 17 based on embedded platform, wherein the step (B) includes step
Suddenly：Judge whether to need the adjustment into every trade process flow, when judging to need into when the adjustment of every trade process flow, to the calculation
Method carries out flow adjustment；When judge need not be into when the adjustment of every trade process flow, to the algorithm flow without flow tune
It is whole, and directly open multinuclear processing.
21. the algorithm transplantation method according to claim 20 based on embedded platform, wherein the step (B) includes step
Suddenly：Judging whether can be using the platform framework of the embedded platform, when judging that the platform framework can be utilized, by institute
It states algorithm flow step and is subdivided into platform framework part；It, will when judging that the platform framework of the embedded platform cannot be utilized
The algorithm flow step is subdivided into no platform frame part.
22. the algorithm transplantation method according to claim 21 based on embedded platform, wherein the step (B) is further
Including step：Multinuclear processing is opened to the platform framework part.
23. the algorithm implant system according to claim 21 based on embedded platform, wherein the step (B) is further
Including step：The no platform frame part is judged whether to need to open multinuclear, when judging that multinuclear need not be opened, distribution
Monokaryon；When judging to need to open multinuclear, multinuclear is distributed.
24. the algorithm implant system according to claim 23 based on embedded platform, wherein the algorithm is 3D structures
Light algorithm, wherein normalization step distribute monokaryon processing, and multinuclear is opened in characteristic point filtering and the filtering of Node points.
25. the algorithm transplantation method according to claim 21 based on embedded platform, wherein the step (B) is further
Including step：To needing the part for opening multinuclear to judge adhesive size in the ranks in the no platform frame part, glued when in the ranks
When Lian Xing great, using horizontal segmentation, when in the ranks adhesive is small, vertical segmentation.
26. the algorithm transplantation method according to claim 25 based on embedded platform, wherein the algorithm is 3D structures
The filtering of light algorithm, wherein characteristic point uses horizontal segmentation, the filtering of Node points to use vertical segmentation.
27. according to any algorithm transplantation method based on embedded platform of claim 21 to 26, wherein the algorithm
For 3D structure light algorithms, wherein picture signal is handled, and smothing filtering and convolutional filtering are subdivided into platform framework part；It is described to return
One changes, characteristic point filtering and the filtering of Node points are subdivided into no platform frame part.
28. according to claim 17 is the algorithm transplantation method of platform based on insertion, wherein after the step (B)
Including step (E)：Adjust DSP multi-core internal memories.
29. the algorithm transplantation method according to claim 28 based on embedded platform, wherein to institute in the step (E)
It states DSP memories and carries out spilling judgement, when judging to overflow, the memory of DSP core is divided into code segment and data segment, and by code segment
It is linked to the low application heap of rate, part data segment is linked to the high application heap of rate.
30. the algorithm transplantation method according to claim 17 based on embedded platform, wherein in the step (C), it will
By DSP multinuclears treated the algorithm framework integration is carried out on RISC control cores.
31. the algorithm implant system according to claim 17 based on embedded platform, wherein being wrapped after the step (D)
Include step (F)：Optimize the algorithm.
32. according to being the algorithm implant system based on embedded platform described in claim 31, wherein the optimization of the step (F)
Method is selected from following methods：The calculating of correlation is placed on as possible under multinuclear and is handled；The time sequence spacing of thread is answered
With；The optimization of register level is carried out using substrate；It is pre-processed in the way of inline in the compilation phase；It will be in algorithm
Partial software filter operator be substituted for hardware filtering.
33. according to claim 17 be based on Embedded algorithm transplantation method, wherein including step before the step (D)
Suddenly：Transplantation effect is assessed, when assessment passes through, executes step (D).
34. it is according to claim 17 be based on Embedded algorithm transplantation method, wherein the step (D) in using JTAG into
Row flash file is burned onto the embedded platform.
35. any described based on Embedded algorithm transplantation method according to claim 17 to 34, wherein algorithm is 3D structure lights
Algorithm.
36. a 3D structure light algorithm transplantation methods, wherein the 3D structure lights algorithm flow includes smothing filtering, normalization is special
Sign point filtering, Node points filtering, which is characterized in that the wherein described 3D structure lights algorithm is any described by claim 17 to 35
Method migrate to embedded platform.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611256319.6A CN108614703B (en) | 2016-12-30 | 2016-12-30 | Embedded platform-based algorithm transplanting system and algorithm transplanting method thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611256319.6A CN108614703B (en) | 2016-12-30 | 2016-12-30 | Embedded platform-based algorithm transplanting system and algorithm transplanting method thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108614703A true CN108614703A (en) | 2018-10-02 |
| CN108614703B CN108614703B (en) | 2022-04-19 |
Family
ID=63658332
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611256319.6A Active CN108614703B (en) | 2016-12-30 | 2016-12-30 | Embedded platform-based algorithm transplanting system and algorithm transplanting method thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108614703B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674935A (en) * | 2019-09-24 | 2020-01-10 | 中国航空工业集团公司沈阳飞机设计研究所 | Method for transplanting intelligent algorithm to airborne embedded platform and intelligent computing platform |
| CN112162797A (en) * | 2020-10-14 | 2021-01-01 | 珠海格力电器股份有限公司 | Data processing method, system, storage medium and electronic device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101118499A (en) * | 2006-08-04 | 2008-02-06 | 深圳市研祥智能科技股份有限公司 | System for software transplantation between isomerization hardware systems |
| CN103631632A (en) * | 2013-11-29 | 2014-03-12 | 华为技术有限公司 | Transplantation method and source to source compiler |
| US20140282372A1 (en) * | 2013-03-15 | 2014-09-18 | ArtinSoft Corporation | Code separation with semantic guarantees |
-
2016
- 2016-12-30 CN CN201611256319.6A patent/CN108614703B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101118499A (en) * | 2006-08-04 | 2008-02-06 | 深圳市研祥智能科技股份有限公司 | System for software transplantation between isomerization hardware systems |
| US20140282372A1 (en) * | 2013-03-15 | 2014-09-18 | ArtinSoft Corporation | Code separation with semantic guarantees |
| CN103631632A (en) * | 2013-11-29 | 2014-03-12 | 华为技术有限公司 | Transplantation method and source to source compiler |
Non-Patent Citations (1)
| Title |
|---|
| 张海: "毫米波成像***DSP处理单元设计及算法移植与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674935A (en) * | 2019-09-24 | 2020-01-10 | 中国航空工业集团公司沈阳飞机设计研究所 | Method for transplanting intelligent algorithm to airborne embedded platform and intelligent computing platform |
| CN110674935B (en) * | 2019-09-24 | 2022-12-20 | 中国航空工业集团公司沈阳飞机设计研究所 | Method for transplanting intelligent algorithm to airborne embedded platform and intelligent computing platform |
| CN112162797A (en) * | 2020-10-14 | 2021-01-01 | 珠海格力电器股份有限公司 | Data processing method, system, storage medium and electronic device |
| CN112162797B (en) * | 2020-10-14 | 2022-01-25 | 珠海格力电器股份有限公司 | Data processing method, system, storage medium and electronic device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108614703B (en) | 2022-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE102020110688A1 (en) | HARDWARE-AGNOSTIC COMPILER FOR DEEP NEURAL NETWORKS | |
| CN112163601B (en) | Image classification method, system, computer device and storage medium | |
| CN111783937A (en) | Neural network construction method and system | |
| CN110148157A (en) | Picture target tracking, device, storage medium and electronic equipment | |
| CN100511135C (en) | Unit test system and method thereof | |
| CN110516817A (en) | A kind of model training data load method and device | |
| CN106682702A (en) | Deep learning method and system | |
| CN115249315B (en) | Heterogeneous computing device-oriented deep learning image classification method and device | |
| CN104813307A (en) | Data processing device, data processing method, and program | |
| CN110750298B (en) | AI model compiling method, equipment and storage medium | |
| CN110689136B (en) | Deep learning model obtaining method, device, equipment and storage medium | |
| CN114641041B (en) | Internet of vehicles slicing method and device oriented to edge intelligence | |
| CN108614703A (en) | Algorithm implant system based on embedded platform and its algorithm transplantation method | |
| CN102855153A (en) | Flow compilation optimization method oriented to chip multi-core processor | |
| CN105718601A (en) | Business dynamic integration model and application method thereof | |
| CN107851002A (en) | A kind of code compiling method and code encoder | |
| CN114461978B (en) | Data processing method and device, electronic equipment and readable storage medium | |
| CN108875914A (en) | The method and apparatus that Neural Network Data is pre-processed and is post-processed | |
| CN113190345B (en) | Method and device for deploying neural network model oriented to software defined satellite | |
| CN111831354A (en) | Data precision configuration method, device, chip array, equipment and medium | |
| Cai et al. | Optimus: towards optimal layer-fusion on deep learning processors | |
| CN108008959A (en) | A kind of Software Development Kit SDK cut-in methods, system and device | |
| CN112200310B (en) | Intelligent processor, data processing method and storage medium | |
| CN103559574B (en) | Method and system for operating workflow | |
| CN117608596A (en) | Neural network model deployment method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |