CN102118173A - High-speed coding method of LDPC and shortened codes thereof - Google Patents

Abstract

The invention discloses a high-speed coding method for LDPC (low-density parity-check) codes and shortened codes thereof. The method adopts the standard code types (8176 and 7154) established by CCSDS (consultative committee for space data system), realizes the coding of high-speed data stream on a hardware platform with lower cost, and enables low-density parity-check code with long length to be coded under the station of high code rate.

Description

The high speed decoding method of a kind of LDPC and shortening sign indicating number thereof

Technical field

The invention belongs to the coding and decoding technical field, relate to the high-speed coding device of a kind of LDPC and shortening sign indicating number thereof.

Background technology

The LDPC sign indicating number proposes in by the thesis for the doctorate of Gallager at him as far back as the sixties in 20th century, but be limited to technical conditions at that time, lack feasible decoding algorithm, after this 35 in the period of ignored by people basically, the figure that has promoted LDPC (low density parity check code) sign indicating number in 1981 by Tanner therebetween and provided the LDPC sign indicating number represents promptly alleged afterwards Tanner figure.People such as Berrou had found Turbo code in 1993, on this basis, people such as MacKay and Neal newly study the LDPC code weight before and after nineteen ninety-five, have proposed feasible decoding algorithm, thereby further found the superperformance that the LDPC sign indicating number is had, caused strong repercussion and greatly concern rapidly.Fields such as deep space communication, optical fiber communication, satellite digital video and audio broadcasting have been widely used at present.The LDPC sign indicating number has become the 4th strong competitor of generation communication system (4G), and is adopted by satellite digital video broadcast standard DVB-S2 of future generation based on the encoding scheme of LDPC sign indicating number.

LDPC channel coding technology first Application is the Chang'e-2 that launches on October 1st, 2010 in China's space industry.Its downstream data rate is 12M/S, and just the LDPC coder all is operated under the frequency of 12MHz.Along with the development of Chinese Space technology, the data volume that various payload need be descended to pass in the space exploration can be increasing, and this LDPC sign indicating number that just need have good error correcting capability can effectively adapt to the requirement of high bit rate.

The hardware of current LDPC decoding realizes being based on the short sign indicating number type of code length (tens code lengths to the hundreds of bit) more, and arithmetic speed lower (tens000000 to tens), and because the restriction of hardware resource, realize that the stronger ldpc decoder of function needs the good chip of performance even need be at the special algorithm specialized designs, the cost of hardware implementation platform is very high, is unfavorable for large-scale application.

Summary of the invention

Technology of the present invention is dealt with problems and is: overcome the deficiencies in the prior art, a kind of LDPC code translator and method are provided.The present invention is standard with CCSDS to (8176,7154) sign indicating number type, has realized the decoding to high-speed data-flow on the hardware platform of lower cost, has solved the long decoding problem of low density parity check code under high code check of long code.

Technical solution of the present invention:

Realize by following steps:

(1) i Frame of decoding data judged:

If be input as the LDPC data, then the true form of frame head data in the LDPC data is stored by the address of employed dateout register; And export the frame head data of Frame in the last execution cycle by the dateout register;

If be input as the shortening code data of LDPC, then the true form of frame head in the LDPC data and filler data stored by the address of employed dateout register; And export the frame head and the filler data of Frame in the last execution cycle by the dateout register;

(2) store i the valid data in the Frame according to the address of employed intermediate data register and iteration result data register; And export decoding data in the last execution cycle by the dateout register described in the step (1);

After (3) i Frame storages finish, i Frame carried out the interative computation operation of check-node and variable node; I+1 Frame to input begins to utilize different dateout registers, intermediate data register and iteration result data register to operate from step (1) simultaneously;

(4) after the operation of i Frame is finished and exported, adopt data register, intermediate data register and the newly arrived Frame of iteration result data register pair that i Frame operated to operate again;

In above-mentioned steps, simultaneously all Frames of input are operated, to a Frame store the used time of back operation equal to all the other each Frames of input after this individual data frame store adding up of used time and.

Check-node in the described step (3) and the operation of the interative computation of variable node are realized by check-node computing module and variable node module respectively, and can carry out multiplexing in the Frame that arrives continuously.

The Frame of operating simultaneously is 4; After the operation of finishing i Frame, i+4 data register, intermediate data register and the iteration result data register that input data frame adopts i Frame operated again.

The present invention compared with prior art has following advantage:

(1) adopts half parallel data processing mode of 2 check-node modules and 16 variable node modules, make the optimization that averages out of resource occupation and operation efficiency.With respect to the processing mode of complete serial, can be so that the interative computation number of times doubles under the situation of buffer memory 3 frame data; With respect to parallel fully processing mode, can save resources of chip in a large number and take, make and on than the low side chip, realize that the decoding of this scale becomes possibility.In addition,, be adjacent two frame data distributing independent arithmetic elements, under the situation of a small amount of increase hardware resource expenditure, can further improve the operating frequency of system based on this kind method.According to actual engineering demand, sequencing control is partly done suitable modification, the iterations that can also double again under the situation that reduces operating frequency 50 percent provides more excellent decoding effect.

(2) adopt half parallel data processing method also can produce a large amount of intermediate data.If under the situation that shorter LDPC sign indicating number is deciphered to code length, can adopt the interior register resources of sheet of FPGA to store temporarily.But the LDPC sign indicating number that (8176,7154) this code length is long is deciphered, and the register resources in the sheet will be not enough, and the frequent register to diverse location carries out the operating frequency that read and write access also can reduce system in the computing.Based on this, adopted the storage of BlockRam in the design, thereby effectively utilized the special circuit of FPGA inside to save logic and interconnection resource in a large number as intermediate data, solved above problem.In addition, because BlockRam addressing read-write has also increased the portability that designs.

(3) control mode of multilayer pipeline organization is actually a kind of optimization of the data processing method that half-and-half walks abreast.On data flow state, this parallel organization is dynamic, when improving the system works frequency, has solved the coherent problem of decoded data stream input, output.

Description of drawings

Fig. 1 is a structural representation of the present invention;

Fig. 2 is Min-sum algorithm check matrix Tanner figure;

Fig. 3 is a check-node computing schematic diagram;

Fig. 4 is a workflow diagram of the present invention.

Embodiment

Below just in conjunction with the accompanying drawings the present invention is described further.

The decoding algorithm of LDPC sign indicating number has:

1.SPA(Sum-Product?Algorithm)，Probability-domain；

2.SPA，Log-domain；

(3.Min-sum the shortcut calculation of Log-domain SPA), and various invulnerable release;

4.BF(Bit-Flipping)；

5.MLG (Majority-Logic), only can use to cyclic LDPC code.

Above listed algorithm, complexity and error performance reduce successively.Wherein, Probability-domainSPA calculates too complex, and BF algorithm and MLG algorithm can not provide good error performance, all should not adopt in the reality.In operable several algorithm, the Min-sum algorithm error performance of correction can have low relatively complexity again near SPA, therefore, has adopted the Min-sum algorithm of revising through taking all factors into consideration the design.

Min-sum algorithm decoding iteration execution mode:

$H=\left[\begin{array}{cccccccccc}1& 1& 1& 1& 0& 0& 0& 1& 0& 0\\ 1& 0& 0& 0& 1& 1& 1& 0& 0& 0\\ 0& 1& 0& 0& 1& 0& 0& 1& 1& 0\\ 0& 0& 1& 0& 0& 1& 0& 1& 0& 1\\ 0& 0& 0& 1& 0& 0& 1& 0& 1& 1\end{array}\right]$

The Min-sum algorithm carries out iterative decoding based on the Tanner figure of LDPC code check matrix H.We are the calculating process that example illustrates it with following check matrix.The Tanner figure of H as shown in Figure 2.Imagination c _iAnd f _jBe respectively the arithmetic element of check-node and variable node, c _iAnd f _jBetween line be duplex channel, the note from c _iAnd f _jLine (be c _iOutput, f _jInput) be q _IjFrom f _jTo c _iLine (be f _jOutput, c _iInput) be r _Ji

Min-sum is initialization f at first _jInput q _Ijq _IjThrough f _jExport r after the computing _Ji, again through c _iQ is upgraded in output after the computing _Ij, finish iteration one time.Up to solving correct numeral, perhaps reach predefined maximum iteration time, finish whole iterative process, decoding finishes.Concrete steps are as follows:

1. initialization q _Ij: what receive contains noise (Var=σ) ²Numeral γ=(γ ₀, γ ₁..., γ ₉), with q _i=2 γ _i/ σ ²As original input, and with all q _IjBe initialized as:

q _ij＝q _i＝2γ _i/σ ²

2. arithmetic element f _j: f ₀Be 5 inputs, 5 outputs, remaining f _jBe 4 inputs, 4 outputs.Output with the pass of input is:

$r_{\mathrm{ji}}={\mathrm{\Π}}_{i^{\′}\≠i}\mathrm{sign}\left(q_{\mathrm{ij}}\right)\·{\min }_{i^{\′}\≠i}\left|q_{i^{\′}j}\right|$

The symbol that is each output is for removing corresponding input q _IjOutward, the product of the symbol of other inputs; The size of each output is:

$r_{21}=\mathrm{\Π}(\mathrm{sign}\left(q_{42}\right),\mathrm{sign}\left(q_{72}\right),\mathrm{sign}\left(q_{82}\right)).\min \left(\right|q_{42}|,|q_{72}|,|q_{82}\left|\right)$

3. arithmetic element c _i: c ₇Be the arithmetic element of 3 inputs, 3 outputs, the c that all the other are all _iBe 2 inputs, 2 outputs.Physical relationship is said clearly,, input with the pass of output is:

$Q_i={\mathrm{\Σ}}_{j^{\′}}{r}_{{j^{\′}}_i}+q_i$

q _ij＝Q _i-r _ij

Be that each is output as and removes outside the corresponding input, other inputs and, add this arithmetic element c _iPairing original input q _i=2 γ _i/ σ ²Intermediate object program Q _iJudging numeral for next step prepares.With c ₇Be example, see Fig. 3, output to f ₀Q ₇₀For:

Q ₇＝r ₀₇+r ₂₇+r ₃₇+2γ ₇/σ ²

q ₇₀＝Q ₇-r ₀₇＝r ₂₇+r ₃₇+2γ ₇/σ ²

4. calculate numeral and judge stop condition: step 2 and 3 has been finished iteration one time, uses Q this moment _iMake the judgement numeral:

$c_i=\left\{\begin{array}{c}1,Q_i<0\\ 0,Q_i>0\end{array}\right.$

If c _iH ^T=0, promptly successfully decoded, perhaps reached predefined maximum iteration time, then stop iteration, otherwise, continue step 2 and 3.

NASA has proposed a kind of (8176,7154) LDPC sign indicating number by the finite geometry method construct in September, 2005, and in May, 2006 it is defined as replacing the normal channel coding standard that R-S adds convolution code.CCSDS (The Consultative Committee for Space Data Systems, CCSDS) technological document of delivering in August, 2006 is recommended the standard code (reference: " CCSDS 131.1-0-1, Aug 2006 ") that this yard used for near-earth.

This sign indicating number has following characteristics:

1, finite geometry sign indicating number can be at BER=10 ^-10The time error-floor do not appear;

2, finite geometry sign indicating number has very fast iterative decoding convergence rate;

3, quasi-cyclic code, systematic code, the utilogic circuit is realized, is not needed real arithmetic;

4, code check 7/8, not obvious increase transmission bandwidth.

Based on above characteristics, the present invention has adopted (8176,7154) the LDPC sign indicating number by the finite geometry method construct.

Hardware realizes adopting the programmable logic device of xilinx company, and the Ram resource that makes full use of wherein replaces look-up table.Adopt the control mode of streamline, the logic function part of key effectively improves operating frequency in the multiplexing algorithm, has reduced resources occupation rate.

Decoding adopts the Min-sum algorithm to carry out.Its core is check-node computing and variable node computing, and subject matter is when the LDPC sign indicating number long to code length deciphered, and the operand that carry out is very big, and producing a large amount of intermediate data simultaneously needs storage.With a frame length 8176bits is example, carry out an iteration and just need carry out 1022 check-node computings and 8176 variable node computings, and the decoding of finishing frame data need be carried out 10 iteration.

According to the characteristics of LDPC check matrix, adopted following method to realize continuous decoding to the 100Mbps data flow:

2 groups of check-node computings are parallel, compare with serial structure, reduce half check-node execution cycle;

16 groups of variable node computings are parallel, compare with serial structure, reduce half variable node execution cycle;

Streamline is controlled operation time sequence, has improved the disposal ability of data, improves system works speed;

Multiplexing one group of check-node computing module of two frame data and variable node computing module have reduced resources occupation rate;

Utilize the inner Ram of FPGA to replace the register-stored intermediate data;

Divide according to function, code translator of the present invention comprises data transaction and control module, data outputting module, a n intermediate data register, a n iteration result register, a n dateout register, a n/2 check-node computing module, a n/2 variable node computing module.Concrete structure below illustrates the realization function of each several part respectively as shown in Figure 1:

Data transaction and control module

The control core of LDPC decoding is finished the format conversion to original input data, input selection, address choice, the read-write operation of control intermediate data register, iteration result register and dateout register; Distribute data flow away to; The sequential of control all functions unit.

Data memory module

Data memory module comprises intermediate data register, iteration result register, dateout register three parts.For interative computation provides interim memory space, present frame is distributed to the next frame use after finishing computing, is the node of exchanges data.

Intermediate data register, iteration result register, dateout register adopt blcok ram to realize.Blcok ram is the specific resources among the FPGA, reads when can realize twoport, and does not take other look-up table resources (look-up table) among the FPGA.

The check-node computing module

Finish the check-node computing.Under (8176,7154) LDPC sign indicating number of CCSDS suggestion, its core is 32 7bits data inputs, calculate wherein minimum value and sub-minimum, sub-minimum is composed data to original minimum value position, minimum value is composed to remainder data, output is 32 7bits data still.Can one group in two groups of 32 7bits data of input be calculated according to the data input select signal that data transaction and control module provide.

When specific implementation, the check-node computing module is by comprising 4 submodules, and its effect is minimum value and the sub-minimum output of selecting in 8 input data wherein, and the position of output minimum value data.

The variable node computing module

Finish the variable node computing.(8176,7154) LDPC sign indicating number of same CCSDS suggestion, according to 5 7bits bit wide data calculating input, calculate the input data and and output; The difference and the output of calculating and value and each position data; Shorten the excessive data of numerical value.

The data output control module

MUX, the input data select signal that provides according to data transaction and control unit are exported data behind 4 frame codings successively continuously.

As shown in Figure 4, be apparatus of the present invention workflow diagram.The synchronous data flow of the demodulator output of code translator front end can be single-bit (a corresponding Hard decision decoding), also can be multi bit quantization (corresponding Soft decision decoding).After data entered code translator, time-sequence control module was changed the data form earlier, and the true form of initial data (coded data) is exported to intermediate data register and complement code is exported to the iteration result register.And according to the design feature of the check matrix of CCSDC suggestion, with the storage of a frame in 32 intermediate data register and 16 iteration result registers.When all data write finish after, set is carried out in address to middle data register and iteration result register, beginning simultaneously with the data read of 32 intermediate data register corresponding address to the check-node computing module, after node computing to be verified finishes, upgrade the data of input before with this result.The mode that adopts two groups of check-node computing modules to work is simultaneously carried out the check-node computing to 1022 groups of data altogether, finishes a verification computing of frame data is upgraded.And then the address of 32 intermediate data register and 16 iteration result registers carried out set, simultaneously with the data read of 32 intermediate data register and 16 iteration result register corresponding address to the variable node computing module, after treating that the variable node computing finishes, upgrade the data of input before with this result.Adopt 16 modes that the variable node computing module is worked simultaneously, altogether 8176 groups of data are carried out the variable node computing, finish a variable computing of frame data is upgraded.

More than be the process of frame data being carried out once complete interative computation, through 10 such iterative process, the decoding of frame data finishes.In order to reach higher speed, adopted the method for parallel, buffer memory, pipeline processes to carry out data processing simultaneously.When reading, the first frame initial data finishes when beginning to handle, the second frame initial data begins to read in, store in second group of 32 intermediate data register and 16 the iteration result registers, the read-in process of third and fourth frame initial data is identical, promptly be that first to fourth frame enters pipeline organization successively in the input process of initial data, every frame data take one group of intermediate data register and iteration result register separately.When first frame data were deciphered, from the operation that first frame data are carried out, other three frame data were equivalent to be buffered in intermediate data register and the iteration result register.In fact from the circulating treatment procedure of entire stream, the decode procedure of any frame data is all corresponding with the buffer memory of other three frame data.Next, second and third, the four frame data pipelined process that enters iterative decoding successively handles.When data flow was constantly imported in a steady stream, the check-node computing and the variable node computing of not only every frame inside walked abreast, and data processing also walks abreast between frame and the frame.As can be seen at any one time, all can have frame data to be in the I/O state, other three frame data are in the iterative decoding state.From the angle of sequencing control, the processing procedure of every frame data is controlled by state machine, from the output that is input to decode results of initial data, is total to about 50 states; With four frame data is that one-period is done circulation decoding, handles continuous data stream and need experience 200 kinds of combinations of states.

The unspecified part of the present invention belongs to general knowledge as well known to those skilled in the art.

Claims (3)

1. a LDPC and shorten the high speed decoding method of sign indicating number is characterized in that realizing by following steps:

(1) i Frame of decoding data judged:

If be input as the LDPC data, then the true form of frame head data in the LDPC data is stored by the address of employed dateout register; And export the frame head data of Frame in the last execution cycle by the dateout register;

If be input as the shortening code data of LDPC, then the true form of frame head in the LDPC data and filler data stored by the address of employed dateout register; And export the frame head and the filler data of Frame in the last execution cycle by the dateout register;

(2) store i the valid data in the Frame according to the address of employed intermediate data register and iteration result data register; And export decoding data in the last execution cycle by the dateout register described in the step (1);

After (3) i Frame storages finish, i Frame carried out the interative computation operation of check-node and variable node; I+1 Frame to input begins to utilize different dateout registers, intermediate data register and iteration result data register to operate from step (1) simultaneously;

(4) after the operation of i Frame is finished and exported, adopt data register, intermediate data register and the newly arrived Frame of iteration result data register pair that i Frame operated to operate again;

In above-mentioned steps, simultaneously all Frames of input are operated, to a Frame store the used time of back operation equal to all the other each Frames of input after this individual data frame store adding up of used time and.

2. the high speed decoding method of a kind of LDPC according to claim 1 and shortening sign indicating number thereof, it is characterized in that: check-node in the described step (3) and the operation of the interative computation of variable node are realized by check-node computing module and variable node module respectively, and can carry out multiplexing in the Frame that arrives continuously.

3. the high speed decoding method of a kind of LDPC according to claim 1 and shortening sign indicating number thereof, it is characterized in that: the Frame of operating simultaneously is 4; After the operation of finishing i Frame, i+4 data register, intermediate data register and the iteration result data register that input data frame adopts i Frame operated again.

Images