The control method of photoelectricity calculating cell operation state
Technical field
The present invention relates to the control methods that a kind of photoelectricity calculates cell operation state, belong to optical field and semiconductor devices
Field.
Background technique
Photoelectricity computing unit is that one kind can be combined with operation independent or with current electronic computation technology to carry out operation
Calculating device, feature are as follows: " depositing-calculate a body function " can be realized in single device, and integrated level is good, and Energy Efficiency Ratio is high, and compatibility is strong.
When photoelectricity computing unit group is combined into large-scale photoelectricity computing array, it can be used to realize the operation acceleration of various complexity
Convolution algorithm in function, such as convolutional neural networks algorithm.If, can be very using photoelectricity computing array as convolution operator
The characteristic that convolution Nuclear Data is constant in multiple convolution operation is played well, utilizes the light input data " storage " of photoelectricity computing unit
Characteristic is able to achieve high-performance convolution algorithm.But how to be routed by limited in large-scale convolution operator array
Each unit is controlled and is carried out accurate light input, this problem also just becomes particularly critical.
Summary of the invention
Against the above technical problems, the purpose of the present invention is to provide the controlling parties that a kind of photoelectricity calculates cell operation state
Method.
The technical solution adopted by the invention is as follows:
Photoelectricity calculates the control method of cell operation state, and photoelectricity computing unit includes luminescence unit and computing unit,
Luminescence unit issue light be incident in computing unit, the computing unit includes three parts: as photo-generated carrier collect and
The P-type semiconductor substrate of read-out area;Charged Couple layer as coupled zone;Control grid as carrier control zone;Wherein,
P-type semiconductor substrate is divided into collecting region and read-out area, is equipped with shallow-trench isolation, N-type source and N-type drain terminal in read-out area, shallow slot every
It offs normal between collecting region and read-out area;
The control method specifically: by being applied in control grid, P-type semiconductor substrate and N-type source and N-type drain
Add different electrical conditions, changes the electrical state in the computing unit, so that computing unit is in different working stages,
Working stage includes: the light input phase for receiving photon and generating photo-generated carrier;It receives electronics and makes electronics and photoproduction current-carrying
Son generates the electric input phase of operation relation;Output is by photo-generated carrier and electricity input electronics coefficient reading stage;It wipes
Except the light of light input data inputs reseting stage.
Further, the control method of the smooth input phase specifically: apply positive pressure on the control gate, and in P
Apply negative pressure in type semiconductor substrate, while to N-type drain terminal and N-type source floating, so that P-type semiconductor substrate bottom generates energy
The depletion layer for enough absorbing incident photon, reaches the condition of light input, so that computing unit enters light input phase;By removing or
The positive pressure on control grid is reduced, or removes or promoted the negative pressure on P-type semiconductor substrate, so that computing unit disengaging light is defeated
Enter the stage;
The control method of the light input reseting stage specifically: by applying negative pressure on the control gate, and in p-type
Apply positive pressure or zero-bias in semiconductor substrate;Or apply negative pressure or zero-bias on the control gate, and partly lead in p-type
Apply positive pressure in body substrate, so that the photo-generated carrier that collecting region is collected is pumped from P-type semiconductor substrate and partly leads with p-type
Hole-recombination in body substrate, computing unit enter light input reseting stage and complete the reset of light input quantity;
The control method of the electricity input phase specifically: the size of electric input quantity is represented by applying on the control gate
Positive bias, make electric input quantity in the form of carrier in input control grid;Or pass through the N-type in P-type semiconductor substrate
Apply the bias for representing the size of electric input quantity between drain electrode and N-type source, makes electric input quantity in terms of input in the form of carrier
The photo-generated carrier for calculating unit is collected in read-out area, and computing unit enters electric input phase and inputs electric input quantity;
The control method in the reading stage specifically: when the carrier for representing electric input quantity passes through control grid input meter
When calculating in unit, by adding a constant bias between the N-type source and N-type drain of P-type semiconductor substrate, make read-out area ditch
It is exported in the form of electric current in road by light input quantity and the electric coefficient carrier of input quantity, computing unit, which enters, reads rank
Section simultaneously is completed to read work;Or when the carrier for representing electric input quantity is single by N-type source and N-type drain input photoelectricity calculating
When in member, by applying a constant bias on the control gate, read-out area surface is set to generate channel, channel carriers are through light
It is exported in the form of electric current again after input quantity and electric input quantity collective effect, computing unit enters the reading stage and completes to read work
Make.
Further, multiple photoelectricity computing units are arranged in the array to the convolution kernel size in convolution algorithm, battle array
The photo-generated carrier of all computing units is collected in column and the output end of read-out area is connected with each other and is aggregated into an output end;Control
Method processed specifically: luminescence unit is in light by controlling each computing unit using the method for single equal field light source light input
The length of time of input phase, to realize individually accurate light input;All computing units according to the duration of light input phase not
Enter light input phase simultaneously, but terminates light input phase simultaneously;Apply constant electricity to photoelectricity computing units all in array
Input quantity controls the single photoelectricity computing unit that some completes light input and enters the reading stage, and other photoelectricity calculate list
Member does not enter the reading stage;The accuracy of the value of active cell light input quantity is judged by reading result;It is tied when reading
Fruit records the deviation after there is deviation, reads the result of next computing unit later and records deviation, all until completing
The light input quantity of computing unit verifies;The reset of light input quantity is carried out to all computing units and is re-entered, and corrects light input
There is the light input phase duration of the computing unit of deviation in amount, and all computing units is finally made to obtain ideal light input quantity.
The present invention can accurately be made each by means of the electrical conditions being applied in photoelectricity computing unit different function area
Photoelectricity computing unit enters light input phase, electric input phase, reading stage or light input reseting stage.When multiple photoelectricity calculate
When unit forms array progress convolution algorithm, by controlling the working condition of each photoelectricity computing unit, array may be implemented
High-precision light input, substantially increase device calculating accuracy.
Detailed description of the invention
Fig. 1 is the multi-functional-area block diagram of computing unit.
Fig. 2 is the structural schematic diagram of computing unit in embodiment 1.
Fig. 3 is the electrical model figure of computing unit in embodiment 1.
Fig. 4 is the convolution operator structural schematic diagram being made of in embodiment 2 photoelectricity computing unit.
Specific embodiment
The present invention provides a kind of working state control method of photoelectricity computing unit, and is made of the photoelectricity computing unit
Convolution operator and control method.
As shown in Figure 1, the computing unit in photoelectricity computing unit is the multi-functional-area structure for including three zones area, wherein
Three zones area are as follows: carrier control zone, coupled zone, photo-generated carrier is collected and read-out area, and concrete function difference is as follows:
Carrier control zone: it is responsible for controlling and modulating the carrier in photoelectricity computing unit, and calculates list as photoelectricity
The electrical input mouth of member inputs one of operand as electric input quantity;Or it only controls and modulates in photoelectricity computing unit
Carrier, pass through other regions and input electric input quantity.
Coupled zone: it is responsible for the collection of connection photo-generated carrier and collecting region and read-out area in read-out area, so that photon is incident
The photo-generated carrier of generation acts on the carrier in photoelectricity computing unit, forms operation relation.
Photo-generated carrier is collected and read-out area: including collecting region and read-out area, wherein collecting region is responsible for absorbing incident light
Son and the photo-generated carrier for collecting generation, and the light input port as photoelectricity computing unit, input one of operand
As light input quantity;Read-out area can be used as the electrical input mouth of photoelectricity computing unit, input one of operand as electricity
Input quantity, and the output port as photoelectricity computing unit, output is by the carrier after light input quantity and electric input quantity effect
As unit output quantity;Or electric input quantity is inputted by other regions, read-out area is only used as the output end of photoelectricity computing unit
Mouthful, output is by the carrier after light input quantity and electric input quantity effect, as unit output quantity.
Single computing unit and a luminescence unit are combined into a complete photoelectricity computing unit, what luminescence unit issued
Light is collected as incident photoelectricity computing unit photo-generated carrier and the photon of read-out area, participates in operation.
Embodiment 1
As shown in Fig. 2, the computing unit of the present embodiment includes: as the control grid of carrier control zone, as coupling
The Charged Couple floor in area, and as the P type substrate of photo-generated carrier collecting region and read-out area, left side is divided into P type substrate and is received
Ji Qu and right side read-out area, wherein including shallow-trench isolation in the read-out area of right side, by the N-type source and N-type of ion implanting formation
Drain terminal.Shallow-trench isolation is located at the centre at semiconductor substrate middle part, collecting region and read-out area, and shallow-trench isolation is by etching and being packed into
Silica is formed, with the electric signal for collecting region and read-out area to be isolated.N-type source is located in read-out area and is situated between by near-bottom
The side of matter layer is adulterated by ion implantation and is formed.N-type drain terminal is located in semiconductor substrate close to underlying dielectric layer and N
The opposite other side of type source is equally doped method by ion implantation and is formed.It should be understood that left side mentioned in this article,
Right side, top and lower section, which are only represented, is changing change with observation visual angle by the relative position under view as shown in the figure
Change, and is not understood to the limitation to specific structure.
Apply the pulse that a voltage range is negative pressure on the substrate of collecting region, or applies a voltage on the control gate
Range is the pulse of positive pressure, so that generating the depletion layer collected for photoelectron in collecting region substrate, and passes through right side read-out area
Read the photoelectron quantity collected, the input quantity as light input end.When reading, applies a positive voltage on the control gate, make N
Conducting channel is formed between type source and collecting region N-type drain terminal, then by applying a biasing arteries and veins between N-type source and N-type drain terminal
Voltage is rushed, so that the electronics in conducting channel accelerates to be formed the electric current between source and drain.The load of electric current is formed between source and drain in channel
Stream is controlled the photoelectron quantity collective effect that gate voltage, source-drain voltage and collecting region are collected, as by light input quantity
Electronics with after electric input quantity collective effect, is exported in the form of electric current, and wherein control-grid voltage, source-drain voltage can be with
As the electric input quantity of device, photoelectron quantity is then the light input quantity of device.
The Charged Couple layer of coupled zone makes depletion region in collecting region substrate start to collect for connecting collecting region and read-out area
After photoelectron, the photoelectron quantity that collecting region substrate surface gesture just will receive collection influences;By the connection of Charged Couple layer,
So that read-out area semiconductor substrate surface gesture is influenced by collecting region semiconductor substrate surface gesture, and then between influence read-out area source and drain
Size of current, to read the photoelectron quantity of collecting region collection by judging electric current between read-out area source and drain;
The control grid of carrier control zone, to apply a pulse voltage on it, so that in P-type semiconductor substrate
It generates in read-out area for exciting photoelectronic depletion region, while can also be used as electrical input, input a wherein bit arithmetic amount.
In addition, there is the underlying dielectric layer for isolation between P-type semiconductor substrate and Charged Couple layer;Charged Couple layer
Also there is the top layer dielectric layer for isolation between control gate.
Control the technical thought that single computing unit is in various states are as follows: by control grid and P type substrate and N
Type source electrode and drain electrode applies different electrical conditions, changes the electrical state in computing unit, makes device be in difference with this
Working stage.It includes four that working stage, which has altogether: in reception photon, and generating the light input phase of photo-generated carrier;Place
In reception electronics, and electronics and photo-generated carrier is made to generate the electric input phase of operation relation;In output by photo-generated carrier
The electronics coefficient reading stage by as photoelectricity operation result read-out area electronics is inputted with electricity;Number is inputted in erasing light
According to light input reseting stage.The specific control method in each stage is as follows:
1, the method that single computing unit is in light input phase is controlled are as follows: by applying positive pressure on the control gate, and
And apply negative pressure in P type substrate, and to N-type source and drain electrode floating, so that P type substrate bottom, which generates, can absorb incident light
The depletion layer of son reaches the condition of light input, photoelectricity computing unit is made to enter light input phase.About the detailed of aforesaid operations mode
Thin derivation is as follows:
As shown in figure 3, left side collecting region is equivalent to a capacitor isMos capacitance, right side read-out area etc.
It imitates in the floating-gate MOS tube of a standard.When due to design, capacitor C2Much smaller than C1, therefore device work when read-out area to photosensitive area
The influence of generation is ignored.
Potential in one MOS- capacitor Si can be obtained by solving following Poisson's equation:
Wherein, εSIFor the dielectric constant of silicon, ρ is the density of volume charges of P type substrate.
When the substrate P as carrier collection and read-out area applies a negative pulse, or the control as carrier control zone
When applying a positive pulse on grid, substrate will be in spent condition, start to collect the photon as light input signal and generate photoelectricity
Son, ρ=qN for depletion regionA, wherein NAFor doping concentration.
It is available to solve above-mentioned Poisson's equation:
Wherein, the direction x is perpendicular to underlying dielectric layer downwardly direction, xdFor depletion region depth, q is electronic charge, V
For the potential that depth is at x.For MOS, P type substrate surface potential VSWhen as x=0 therefore the value of potential V can obtain:
The derivation formula can obtain:
Wherein ESFor surface field intensity, it is assumed that underlayer voltage is set as 0V, the control grid potential in such photo-process
Are as follows:
Wherein, VGTo control gate potential, solution can obtain depletion region depth xdAre as follows:
When there is photon incidence device, photoelectron is generated in depletion region and is collected in receipts under the action of gate electric field
Collect in area's channel, the total charge dosage Q on control gateCG=NA+ Q, Q are signal charge quantity (e-/cm2), because of signal charge thus
It is collected in collecting region under the action of electric field between control gate and P type substrate, and because of carrier in semiconductor substrate
It is compound need the regular hour, along with the presence of thermal excitation carrier in depletion region, therefore, this signal charge will be in disconnected light
It is still stored in arithmetic element in longer time afterwards, a body function is deposited-calculated in realization.
At this point,
Wherein, VQThe potential summation generated for signal charge:
By above formula it can be seen that with signal charge quantity Q increase, xdIt is gradually reduced, when the value of Q makes VQX when=0dI.e.
It is 0, surface potential V at this times=0, groove potential no longer changes, and device reaches full trap at this time.
For read-out area the floating-gate MOS FET, channel current I on right sidedIt can indicate are as follows:
Wherein W and L is respectively grid width grid length, VDSFor source-drain voltage, VFGFor Charged Couple layer potential, size by
Control gate potential VGWith P type substrate surface potential VsInfluence, may be expressed as:
When P type substrate doping concentration is lower (such as 2E15 is per cubic centimeter), the partial pressure of depletion region is much larger than capacitor C1And C3
Partial pressure, therefore formula (6) can simplify are as follows:
Formula (11) brings (3) into get P type substrate surface potential V is arrivedSWith control gate potential VGAnd signal charge generation
Potential summation VQIt is approximately equal, it may be assumed that
Vs≈VQ (12)
It brings formula (12), (8) into (10), then brings (9) into, can obtain:
And with incident light subnumber XphotonTo indicate the size of signal charge Q:
Q=Xphotontη (14)
Wherein t is time for exposure, XphotonFor the photon number of unit time incidence, η is device quantum efficiencies.
It can be used as the expression formula of multiplier work to just obtain device:
It is not difficult to find out from formula (15), the read-out area source-drain current I as output quantitydSimultaneously by as light input quantity
Xphoton, as the V of electric input quantityGAnd VDSEffect, using such interactively, which can be realized
A variety of different calculation functions.
According to derivation described above, device is allowed to be in light input state if necessary, then it must be in the photoproduction of P type substrate
In collecting region in carrier collection and read-out area, i.e., depletion layer is generated in the left side photosensitive area of P type substrate, in addition to institute in derivation
That states controls application positive pressure V on grid againG, depletion layer can also be generated by way of applying negative pressure on substrate, according to
The relativity principle of potential, because of N-type source and the equal floating that drains when photoelectricity computing unit is in light input state, at this time
Electrode port only control grid and substrate, if so control grid produces enough positive potential differences relative to substrate,
Enough depletion layers be can produce to collect the incident photon as light input quantity.According to formula (11), as long as producing control gate
Positive potential difference between pole and substrate will generate depletion layer in the photosensitive area of left side so that the photoelectricity computing unit into
Enter the condition of light input phase are as follows:
VG-VB> 0
Wherein VBFor P type substrate potential.
2, the method that single computing unit is in light input reseting stage is controlled are as follows: negative by applying on the control gate
Pressure, and apply positive pressure or zero-bias in P type substrate, or apply negative pressure or zero-bias on the control gate, and in P
Apply positive pressure on type substrate.According to analysis above, left side photosensitive area can enter light input phase and be received after generating depletion layer
Collect photon and generate photo-generated carrier, photo-generated carrier can be collected into the surface of left side photosensitive area and storage under the action of electric field
In this, the storage as " deposit and calculate one " is being stored in this in long-time for calculating the time, therefore when in P type substrate
When with applying the pressure drop opposite with light input phase on control grid, electronics will be under the action of electric field by from P type substrate
It takes away, completes the reset of light input quantity, in addition, after removing the electrical conditions being added on substrate and control grid completely, due to carrying
The compound action for flowing son, light input quantity also can voluntarily reset in compound completion after the long period, therefore the condition that light input quantity resets
It is considered:
VG-VB≤0。
3, the method that single photoelectricity computing unit is in electric input phase is controlled are as follows: by applying representative on the control gate
The positive bias of the size of electric input quantity makes electric input quantity input the carrier control zone of photoelectricity arithmetic element in the form of carrier
In the middle;Or apply the bias for representing the size of electric input quantity between the N-type drain and N-type source of P type substrate, make electric input
It measures and is inputted in the form of carrier in carrier collection and the read-out area of photoelectricity arithmetic element, photoelectricity computing unit is made to enter electricity
Input phase simultaneously inputs electric input quantity.
According to formula (15), if using control grid as the input port of electric input quantity, the electric input quantity inputted with
The relationship of electric current is then addition relationship between the source and drain finally exported as a result, along with light input quantity and output electric current are all to add
Method relationship, the then function that photoelectricity computing unit is realized at this time are considered add operation;Or electric input quantity is entered admittedly
Be set to 0 perhaps 1 binaryzation value then it is also assumed that at this time photoelectricity computing unit realize operation be 0 × Q or 1 × Q
Multiplying, wherein Q is light input quantity.If using the N-type source and N-type drain of read-out area in carrier collection and read-out area
As the input port of electric input quantity, then between the electric input quantity that inputs and the source and drain finally exported as a result, the relationship of electric current is then
For multiplication relationship, the progress of photoelectricity computing unit is multiplying.Because the voltage of control grid has to be larger than read-out area
The threshold value of MOSFET, otherwise operation result will be unable to read by read-out area electric current, thus control grid as electrical input when
Voltage have a lower limit.
Because electric input quantity does not have the function of " storage ", so if needing to be input to electric input quantity in unit to participate in
Operation must just make photoelectricity computing unit be in electric input phase for a long time.Therefore, photoelectricity computing unit is made to be in electricity input rank
The condition of section are as follows:
And VGRepresent electric input quantity or VDSIn the presence of the voltage for representing electric input quantity.
4, the method that single photoelectricity computing unit is in the reading stage is controlled are as follows: when the carrier for representing electric input quantity passes through
When controlling in grid input light electricity computing unit, by between the N-type source and N-type drain in P type substrate plus one constant
Bias makes in the read-out area channel of P type substrate right side by light input quantity and the electric coefficient carrier of input quantity with electric current
Form output allows photoelectricity computing unit to enter the reading stage and completes to read work;Or when the carrier for representing electric input quantity
It is constant inclined by applying one on the control gate when being inputted in photoelectricity computing unit by the source electrode and drain electrode in P type substrate
Pressure makes read-out area surface on the right side of P type substrate generate channel, by current-carrying in light input quantity and the coefficient channel of electric input quantity
Son exports in the form of electric current again, so that photoelectricity computing unit is entered the reading stage and completes to read work.
According to analysis above, when carrying out electricity input by control grid, photoelectricity computing unit realize addition or
Binaryzation multiplying needs to apply a constant bias between the source and drain of read-out area on the right side of the P type substrate at this time, so that it may drive
Electric current is generated by light input quantity and the electric coefficient carrier of input quantity in channel and is read;Equally, when by source electrode and
When drain electrode carries out electricity input, photoelectricity computing unit realizes multiplying, needs to apply one on the control gate at this time constant
And it is greater than the positive pressure of MOSFET threshold value, so that it may drive in channel by light input quantity and the coefficient carrier of electric input quantity
It generates electric current and reads.Therefore, photoelectricity computing unit is made to be in the condition in reading stage are as follows:
Or VDSThere are constant voltages.
Embodiment 2
Convolution operator as shown in Figure 4 can be formed using the photoelectricity computing unit of embodiment 1, is a convolution in the figure
Core is the convolution operator of 3*3 size, wherein each box for having a V, that is, represents the photoelectricity using embodiment 1
Computing unit, the lead that long side is drawn represent control grid lead, and two leads of short side represent drawing for source electrode and drain electrode
Line, 9 photoelectricity calculate the interface of one P type substrate of units shared.The working principle of convolution operator is as follows:
By matrix A be directed to the convolution algorithm of convolution kernel a for, the simple process for introducing lower convolution algorithm, wherein A be
10*10 matrix, convolution kernel of a for 3*3, step-length 1, such as formula (17):
The rule of convolution algorithm, be to convolution matrix under the mapping of convolution kernel and convolution kernel in element act on one by one, then
According to the mobile convolution kernel of corresponding step-length, mapped next time.The convolution algorithm in (17) is solved, needs to carry out following several
A step:
1) zero padding operates:
Will to convolution matrix A from 10*10 matrix-expand be 12*12 matrix, i.e., on 0 row, a left side for 0 column, under 10 rows
Again each with 10 column adds a row/column, the element all 0 in the ranks of addition, so zero padding is cried, through later, matrix A becomes
Matrix A0, such as formula (18):
2) initial convolution nuclear location is determined:
The initial position of convolution kernel and the most upper left corner of matrix A are overlapped, i.e. 3 rows 3 column of convolution kernel a respectively correspond matrix A0
The 0th, 1,2 rows and the 0th, 1,2 column, then by the element in convolution kernel and the matrix A with convolution kernel corresponding position0In element one
One is multiplied, and such as formula (19), becomes 9 multiplication results, then 9 multiplication results is all cumulative, obtains current convolution kernel position
The convolution algorithm set is as a result, be called R00, as completion (20) described operation:
(a00*0)+(a01*0)+(a02*0)+(a10*0)+(a11*A00)+(a12*A01)+(a20*0)+(a21*A10)+(a22*11)
=R00 (20)
3) position of mobile convolution kernel:
Because the step-length of predefined this time convolution algorithm is 1, the position of convolution kernel is moved to left into 1 column, that is, moves to left 1
3 rows 3 column of convolution kernel a respectively correspond matrix A after column0The 0th, 1,2 rows and the 1st, 2,3 column, later again under current location into
Convolution algorithm result is called R by row convolution algorithm01。
4) entire matrix A is traversed to convolution kernel0Afterwards, altogether available (10+2-2)2A convolution results, by the convolution
As a result it is arranged as matrix according to corresponding convolution nuclear location, obtains (21)
Above-mentioned matrix R, as band convolution matrix A, carry out the knot for the convolution algorithm that step-length is 1 under the action of convolution kernel a
Fruit.
As can be seen that convolution algorithm is two two-phase of multiple two matrix corresponding elements in the step of from above-mentioned convolution algorithm
Multiply operation cumulative again, two matrixes that wherein element is multiplied two-by-two, one of matrix is convolution kernel, in multiple operation
Constant amount, another matrix are band convolution matrix element corresponding with convolution nuclear location, are the amount of variation in multiple operation,
Therefore can use light input storage can store this advantages characteristic of data, input convolution Nuclear Data using light input end, and
Convolution algorithm is carried out by electrical input input tape convolution matrix data, Energy Efficiency Ratio and arithmetic speed can be greatly improved in this way.
Therefore, the electrical input of unit is convolution operator to convolution matrix data input pin, and light input end is convolution kernel input terminal.
Convolution operator can be divided into serial input and parallel two kinds of input, and the main distinction is the quantity and electricity using unit
For fan-in according to the mode of input, serial input scheme is as follows: according to convolution algorithm mode, element in usage quantity and convolution kernel
Unit is arranged in array identical with convolution kernel dimension by the comparable photoelectricity computing unit using the first scheme of quantity, and
Carrier collection is all connected with the output end of read-out area in read-out area, completes to be added by convergence.
Firstly, convolution Nuclear Data is input in computing unit, then will currently roll up in matrix one by one by light input end
The data of product core corresponding position are converted into binary system, then serially from input array, the result of output is converged on control grid
By AD conversion metering-in control system after addition, the convolution algorithm of current convolution nuclear location is arrived using shifting and adding up
As a result, inputting the convolution Nuclear Data pre-stored using light before in mobile convolution kernel later, electricity input number is directly re-entered
According to, can be obtained the corresponding convolution algorithm of next convolution nuclear location as a result, and so on, until convolution kernel is convenient entirely wait roll up
Then the convolution results of output are reassembled into matrix of consequence by product matrix, that is, complete whole convolution algorithms.
It illustrates how to control single photoelectricity computing unit in the convolution operator of a 3*3 convolution kernel in detailed below
It is filled in a variety of different working conditions and carries out accurate light input.
Totally 9 data for representing 3*3 convolution kernel are inputted in nine photoelectricity computing units respectively by light, utilize multiple volume
The constant characteristic of convolution Nuclear Data plays the advantage of photoelectricity computing unit " deposit and calculate one " in product operation;The number being convolved in matrix
It is sequentially input according to from 9 control grid leads.
If you need to make any one unit in 9 photoelectricity computing units in convolution operator enter light input phase, then
Need in P type substrate application -3V voltage, and need into light input phase device control grid on apply 0V, produce
The depletion layer of raw 3V potential difference, other do not need to make power-on and power-off into the control grid previous generation circle -3V of the device of light input phase
Potential difference is 0.
It is resetted if you need to make any one unit in 9 photoelectricity computing units in convolution operator enter light input data
Stage, then need in P type substrate holding -3V voltage, with the photoelectricity computing unit that ensures not needing to be resetted using substrate -
The light input data obtained under the conditions of 3V will not be destroyed, and in the control grid for the device for needing to enter light input reseting stage
Upper application -4V, other do not need to maintain 0V on the control grid of the device of light input reseting stage, and single light can be completed
The light input data of electric computing unit resets.
If you need to make any one unit in 9 photoelectricity computing units in convolution operator enter electric input phase, then
The voltage for representing electric input quantity need to be applied on the control grid of the unit, and apply 0V on other units, because of convolution thus
The wire laying mode of arithmetic unit is that the source of all units is connected with drain terminal, therefore unusable source-drain voltage is defeated as electricity
Enter end and electric input individually is carried out to some unit.
If you need to make any one unit in 9 photoelectricity computing units in convolution operator enter the reading stage, then need
Apply the bias of a 0.5V in the drain terminal of all element-interconn ections, source applies 0V, because for no photoelectricity for carrying out electric input
Control gate extremely 0V for computing unit, therefore the MOSFET of right side read-out area is in the electric current for closing short status, reading from source
The output electric current of the single photoelectricity computing unit of electric input is only carried out, it can be considered that only the unit enters reading rank
Section.
If it is defeated to be in light by each computing unit of control for the method that luminescence unit is inputted using single equal field light source light
Enter the length of time in stage, to realize individually accurate light input;All computing units are different according to the duration of light input phase
When enter light input phase, but terminate light input phase simultaneously.The convolution nuclear matrix for needing to input is
Then specific steps are as follows:
1, it opens equal field light source and irradiates convolution operator.
It 2, is the 0ms moment at the time of opening with light source, it is defeated that three photoelectricity computing units of the first row enter light in 0ms
Enter the stage, the second photoelectricity computing unit of row three does not enter light input phase, and three photoelectricity computing units of the third line are in 10ms
Into light input phase, when 20ms, all 9 photoelectricity computing units were all detached from light input phase, and closed equal field light source,
In, 2000 electronics should be entered by representing photoelectricity computing unit when light input quantity is 2, if light input 100% is accurate, then be existed
Equal field light source, which irradiates lower light input 20ms, should just input in 2000 electronics to photoelectricity computing unit.
3, the V in Fig. 4dPlace applies constant voltage values 0.5V, and lower end reads termination 0V.And it is calculated in upper left corner photoelectricity single
Apply 3V voltage on the control grid lead of member, the control grid of other units applies 0V, so that the photoelectrometer in the only upper left corner
Unit is calculated to enter the reading stage and export electric current.
If the read current of the unit should be 10uA, but real 4, when light input quantity is exactly equal to 2000 electronics
The unit read current in the border upper left corner is 9uA, then it is assumed that the input of this element light is insufficient, corrects the light input of the photoelectricity computing unit
Time becomes 22ms from 20ms, records the variation.Judge that the light of all photoelectricity computing units inputs with same method again
Departure simultaneously records.
5, reset all photoelectricity computing units, and to re-start light defeated according to the amendment light input duration recorded before
Enter, accurate light input quantity can be obtained.
It, can be complete after carrying out primary accurately light input because multiple operation convolution Nuclear Data is constant in convolution algorithm
At multiple convolution operation.