HardwareEfficient Low Density Parity Check Code for Digital Communications
A network element receiving signals from the network over a communications channel via transceiver circuitry. The network element has a host interface for communicating to a host system, decoded signals corresponding signals received from the network. Demodulator circuitry demodulates the signals into a data stream. Circuitry for decoding the data stream according to a sequence of operations is provided. The sequence of operations includes receiving a set of input values corresponding to input nodes of the macro parity check matrix. Estimating a check node value using values of other input nodes contributing to the parity check sum. Evaluating a probability value using the estimates of the check node values for that input node. The The operations are repeated until termination point is reached.
Latest TEXAS INSTRUMENTS INCORPORATED Patents:
This application is a Divisional of and claims priority under 35 U.S.C. §120 to U.S. application Ser. No. 11/463,236, filed on Aug. 8, 2006; which is a Continuation of U.S. application Ser. No. 10/329,597, filed on Dec. 26, 2002—now U.S. Pat. No. 7,178,080. Said Continuation claims priority, under 35 U.S.C. §119(e), to Provisional Application No. 60/403,668, filed Aug. 15, 2002. All said applications incorporated herein by reference.
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENTNot applicable.
BACKGROUND OF THE INVENTIONThis invention is in the field of digital data communications, and is more specifically directed to redundant coding for error detection and correction in such communications.
Highspeed data communications, for example in providing highspeed Internet access, is now a widespread utility for many businesses, schools, and homes. At this stage of development, such access is provided according to an array of technologies. Data communications are carried out over existing telephone lines, with relatively slow data rates provided by voice band modems (e.g., according to the current v.92 communications standards), and at higher data rates using Digital Subscriber Line (DSL) technology. Another modern data communications approach involves the use of cable modems communicating over coaxial cable, such as provided in connection with cable television services. The Integrated Services Digital Network (ISDN) is a system of digital phone connections over which data is transmitted simultaneously across the world using endtoend digital connectivity. Localized wireless network connectivity according to the IEEE 802.11 standard has become very popular for connecting computer workstations and portable computers to a local area network (LAN), and often through the LAN to the Internet. Wireless data communication in the Wide Area Network (WAN) context, which provides cellulartype connectivity for portable and handheld computing devices, is expected to also grow in popularity.
A problem that is common to all data communications technologies is the likelihood of corruption of data due to noise. As is fundamental in the art, the signaltonoise ratio for a communications channel is a degree of goodness of the communications carried out over that channel, as it conveys the relative strength of the signal that carries the data (as attenuated over distance and time), to the noise present on that channel. These factors relate directly to the likelihood that a data bit or symbol received over the channel will be in error relative to the data bit or symbol as transmitted. This likelihood is reflected by the error probability for the communications over the channel, commonly expressed as the Bit Error Rate (BER) ratio of errored bits to total bits transmitted. In short, the likelihood of error in data communications must be considered in developing a communications technology. Techniques for detecting and correcting errors in the communicated data must be incorporated for the communications technology to be useful.
Error detection and correction techniques are typically implemented through the use of redundant coding of the data. In general, redundant coding inserts data bits into the transmitted data stream that do not add any additional information, but that indicate whether an error is present in the received data stream. More complex codes provide the ability to deduce the true transmitted data from a received data stream, despite the presence of errors.
Many types of redundant codes that provide error correction have been developed. One type of code simply repeats the transmission, for example repeating the payload twice, so that the receiver deduces the transmitted data by applying a decoder that determines the majority vote of the three transmissions for each bit. Of course, this simple redundant approach does not necessarily correct every error, but greatly reduces the payload data rate. In this example, a predictable likelihood remains that two of three bits are in error, resulting in an erroneous majority vote despite the useful data rate having been reduced to onethird. More efficient approaches, such as Hamming codes, have been developed toward the goal of reducing the error rate while maximizing the data rate.
The wellknown Shannon limit provides a theoretical bound on the optimization of decoder error as a function of data rate. The Shannon limit provides a metric against which codes can be compared, both in the absolute and relative to one another. Since the time of the Shannon proof, modern data correction codes have been developed to more closely approach the theoretical limit. An important type of these conventional codes are “turbo” codes, which encode the data stream by applying two convolutional encoders. One convolutional encoder encodes the datastream as given, while the other encodes a pseudorandomly interleaved version of the data stream. The results from the two encoders are interwoven to produce the output encoded data stream.
Another class of known redundant codes is the Low Density Parity Check code. According to this class of codes, a sparse matrix H defines the code, with the encodings t of the payload data satisfying:
Ht=0 (1)
over Galois field GF(2). Each encoding t consists of the source message s combined with the corresponding parity check bits for that source message s. The encodings t are transmitted, with the receiving network element receiving a signal vector r=t+n, n being the noise added by the channel. Because the decoder at the receiver knows matrix H, it can compute a vector z=Hr. However, because r=t+n, and because Ht=0:
z=Hr=Ht+Hn=Hn (2)
The decoding process thus involves finding the sparsest vector x that satisfies the equation:
Hx=z (3)
over GF(2). The vector x becomes the best guess for noise vector n, which can be subtracted from the received signal vector r to recover encodings t, from which the original source message s is recoverable. There have been many examples of LDPC codes that are known in the art, and these LDPC codes have been described as providing code performance that approaches the Shannon limit, as described in Tanner et al., “A Class of GroupStructured LDPC Codes”, ISTCA2001 Proc. (Ambleside, England, 2001).
In general, highperformance LDPC code decoders are difficult to implement into hardware. In contrast to Shannon's adage that random codes are good codes, it is regularity that allows efficient hardware implementation. To address this difficult tradeoff between code irregularity and hardware efficiency, the technique of belief propagation provides an iterative implementation of LDPC decoding can be made somewhat efficient, as described in Richardson, et al., “Design of CapacityApproaching Irregular LowDensity Parity Check Codes,” IEEE Trans. on Information Theory, Vol. 47, No. 2 (February 2001), pp. 619637; and in Zhang et al., “VLSI ImplementationOriented (3,k)Regular LowDensity ParityCheck Codes”, IEEE Workshop on Signal Processing Systems (September 2001), pp. 25.36. Belief propagation decoding algorithms are also referred to in the art as probability propagation algorithms, message passing algorithms, and as sumproduct algorithms.
In summary, belief propagation algorithms are based on the binary parity check property of LDPC codes. As mentioned above and as known in the art, each check vertex in the LDPC code constrains its neighboring variables to form a word of even parity. In other words, the product of the LDPC code word vector with each row of the parity check matrix sums to zero. According to the belief propagation approach, the received data are used to represent the input probabilities at each input node (also referred to as a “bit node”) of a bipartite graph having input nodes and check nodes. Within each iteration of the belief propagation method, bit probability messages are passed from the input nodes to the check nodes, updated according to the parity check constraint, with the updated values sent back to and summed at the input nodes. The summed inputs are formed into log likelihood ratios (LLRs) defined as:
where c is a coded bit received over the channel.
In its conventional implementation, the belief propagation algorithm uses two value arrays, a first array L(q_{mj}) storing the LLRs for the input nodes, and the second array R_{mj }storing the results of the parity check node updates, with m being the parity check row index and j being the column (or input node) index. The general operation of this conventional approach determines, in a first step, the R_{mj }values by estimating, for each check sum (row of the parity check matrix) the probability of the input node value from the other inputs used in that checksum. The second step of this algorithm determines the LLR L (q_{mj}) probability values by combining, for each column, the R_{mj }values for that input node from parity check matrix rows in which that input node participated. A “hard” decision is then made from the resulting probability values, and is applied to the parity check matrix. This twostep iterative approach is repeated until the parity check matrix is satisfied (all parity check rows equal zero, GF(2)), or until another convergence criteria is reached, or a terminal number of iterations have been executed.
By way of further background, the code design approach described in Boutillon et al., “DecoderFirst Code Design”, Proc.: Int'l Symp. on Turbo Codes and Related Topics (Brest, France, September 2001) defines the decoder architecture first, and uses this architecture to constrain the design of the LDPC code itself. Sridhara, et al., “Low Density Parity Check Codes from Permutation Matrices”, 2001 Conference on Information Sciences and Systems (Johns Hopkins University, Mar. 2123, 2001) describes the LDPC code as constructed from shifted identity matrices (i.e., permutation matrices).
However, it has been observed in connection with this invention, that these prior approaches are somewhat limited, in that these approaches are limited to a single code or a small selection of codes. Practically useful communications receivers require some amount of flexibility in code rates, and in optimizing their operation for varying noise levels and channel conditions.
BRIEF SUMMARY OF THE INVENTIONIt is therefore an object of this invention to provide an LDPC decoding scheme which can be efficiently implemented in an integrated circuit.
It is a further object of this invention to provide such a scheme that is flexible over a wide range of code rates.
It is a further object of this invention to provide such a scheme having the capability of parallelism, to provide further efficiencies in operation and construction.
Other objects and advantages of this invention will be apparent to those of ordinary skill in the art having reference to the following specification together with its drawings.
The present invention may be implemented in connection with a network receiver, having a decoder that implements a LowDensity ParityCheck (LDPC) code for retrieving the transmitted message. The LDPC code is implemented according to a parity check matrix consisting of an irregular arrangement of cyclically shifted identity matrices, resulting in an irregular LDPC code that provides performance near the Shannon limit. A decoder architecture for this code includes a group of column sum memories that receive the received input data, and that accumulate and store updated values for the input node predictions. A reversible router block forwards these column, input node, values to a parity check update block, at which multiple predictions are generated for each input node, one prediction for each parity check (row) in which the input node is involved; a prediction memory is also provided for storing these predictions. The outputs of the parity check update block are forwarded through the router, and accumulated in the column sum memories.
According to another aspect of the invention, the invention is implemented by encoding a datastream by applying a systematic block code corresponding to an irregular arrangement of circularly shifted identity matrices.
The present invention will be described in connection with an example of its implementation in an exemplary transceiver, for example a wireless network adapter such as according to the IEEE 802.11 wireless standard. It will be apparent to those skilled in the art having reference to this specification that this invention is particularly wellsuited for use in such an application. However, it is also contemplated that this invention will be of similar benefit in many other applications that involve error correction coding, including communications according to orthogonal frequency division multiplexing (OFDM), discrete multitone modulation (DMT) for example as used in conventional Digital Subscriber Line (DSL) modems, and other modulation and communication approaches, whether carried out as land line or wireless communications. It is therefore to be understood that these and other alternatives to and variations on the embodiment described below are contemplated to be within the scope of the invention as claimed.
As shown in
The input bitstream is received by LDPC encoder function 11, according to this embodiment of the invention. LDPC encoder function 11 digitally encodes the input bitstream by applying a redundant code for error detection and correction purposes. According to this embodiment of the invention, the redundant LDPC code applied by encoder function 11 is selected in a manner that facilitates implementation and performance of the corresponding decoder in receiving tranceiver 20. The specifics of the code will become apparent from the description of this decoder function, presented below relative to the description of the construction and operation of receiving tranceiver 20. In general, the coded bits include both the payload data bits and also code bits that are selected, based on the payload bits, so that the application of the codeword (payload plus code bits) to the sparse LDPC parity check matrix equals zero for each parity check row. After application of the LDPC code, bit to symbol encoder function 11 groups the incoming bits into symbols having a size, for example, ranging up to as many as fifteen bits. These symbols will modulate the various subchannels in the OFDM broadband transmission.
The encoded symbols are then applied to inverse Discrete Fourier Transform (IDFT) function 14. IDFT function 14 associates each input symbol with one subchannel in the transmission frequency band, and generates a corresponding number of time domain symbol samples according to the Fourier transform. These time domain symbol samples are then converted into a serial stream of samples by paralleltoserial converter 16. Functions 11 through 16 thus convert the input bitstream into a serial sequence of symbol values representative of the sum of a number of modulated subchannel carrier frequencies, the modulation indicative of the various data values, and including the appropriate redundant code bits for error correction. Typically, for an input of N/2 complex symbols, IDFT function 14 outputs a block of N realvalued time domain samples. Those skilled in the art having reference to this specification will readily recognize that each of functions 11 through 16 may be carried out, and preferably actually are carried out, as digital operations executed by a digital signal processor (DSP).
Filtering and conversion function 18 then processes the datastream for transmission. Function 18 applies the appropriate digital filtering operations, such as interpolation to increase sample rate and digital low pass filter for removing image components, for the transmission. The digitallyfiltered datastream signal is then converted into the analog domain and the appropriate analog filtering is then applied to the output analog signal, prior to its transmission.
The output of filter and conversion function 18 is then applied to transmission channel C, for forwarding to receiving tranceiver 20. The transmission channel C will of course depend upon the type of communications being carried out. In the wireless communications context, the channel will be the particular environment through which the wireless transmission takes place. Alternatively, in the DSL context, the transmission channel is physically realized by conventional twistedpair wire. In any case, transmission channel C adds significant distortion and noise to the transmitted analog signal, which can be characterized in the form of a channel impulse response.
This transmitted signal is received by receiving tranceiver 20, which, in general, reverses the processes of transmitting transceiver 10 to recover the information of the input bitstream.
Transceiver 20 in this example includes spread spectrum processor 31, which is bidirectionally coupled to bus B on one side, and to radio frequency (RF) circuitry 33 on its other side. RF circuitry 33, which may be realized by conventional RF circuitry known in the art, performs the analog demodulation, amplification, and filtering of RF signals received over the wireless channel and the analog modulation, amplification, and filtering of RF signals to be transmitted by transceiver 20 over the wireless channel, both via antenna A. The architecture of spread spectrum processor 31 into which this embodiment of the invention can be implemented follows that of the TNETW1100 singlechip WLAN medium access controller (MAC) available from Texas
Instruments Incorporated. This exemplary architecture includes embedded central processing unit (CPU) 36, for example realized as a reduced instruction set (RISC) processor, for managing high level control functions within spreadspectrum processor 31. For example, embedded CPU 36 manages host interface 34 to directly support the appropriate physical interface to bus B and host system 30. Local RAM 32 is available to embedded CPU 36 and other functions in spread spectrum processor 31 for code execution and data buffering. Medium access controller (MAC) 37 and baseband processor 39 are also implemented within spreadspectrum processor 31 according to the preferred embodiments of the invention, for generating the appropriate packets for wireless communication, and providing encryption, decryption, and wired equivalent privacy (WEP) functionality. Program memory 35 is provided within tranceiver 20, for example in the form of electrically erasable/programmable readonly memory (EEPROM), to store the sequences of operating instructions executable by spreadspectrum processor 31, including the coding and decoding sequences according to the preferred embodiments of the invention, which will be described in further detail below. Also included within wireless adapter 20 are other typical support circuitry and functions that are not shown, but that are useful in connection with the particular operation of tranceiver 20.
According to the preferred embodiments of the invention, LDPC decoding is embodied in specific custom architecture hardware associated with baseband processor 39, and shown as LDPC decoder circuitry 38 in
Alternatively, it is contemplated baseband processor 39 itself, or other computational devices within tranceiver 20, may have sufficient computational capacity and performance to implement the decoding functions described below in software, specifically by executing a sequence of program instructions. It is contemplated that those skilled in the art having reference to this specification will be readily able to construct such a software approach, for those implementations in which the processing resources are capable of timely performing such decoding.
Referring back to the functional flow of
LDPC decoder function 28 reverses the encoding that was applied in the transmission of the signal, to recover an output bitstream that corresponds to the input bitstream upon which the transmission was based. This output bitstream is then forwarded to the host workstation or other recipient.
LDPC Decoding
The theory of operation of the preferred embodiment of the invention will now be described, following which its implementation into LDPC decoding function 28 in tranceiver 20, in the form of LDPC decoder circuitry 38 operating in cooperation with baseband processor 39, will then be described.
By way of nomenclature, the LDPC code is fundamentally contained within an m×j parity check matrix H_{pc }that, when multiplied by the true transmitted code word vector c equals zero:
H_{pc}·c=0 (5)
over Galois Field (2). For a single one of the m rows in parity check matrix H_{pc}, this parity check amounts to:
H_{1}c_{1}+H_{2}c_{2}+ . . . +H_{j}c_{j}=0 (6a)
over GF(2). In the LDPC code according to the preferred embodiments of the invention, the parity check matrix H_{pc }is formed from a composite of circularly shifted identity matrices represented by a macro matrix H. Each entry in macro matrix H represents a permutation matrix (e.g., a circularly shifted identity matrix), and in this example takes either a 1 or a 0 value. As will be described below, an entry with a 1 value in macro matrix
H symbolizes a p×p permutation matrix at that position within parity check H_{pc}, while entries with a 0 value symbolize a p×p zero matrix. The paritycheck equation thus logically becomes, for an exemplary row of matrix H_{pc }having a “1” in its columns 1, 3, 4, and 7:
c_{1}⊕c_{3}⊕c_{4}⊕c_{7}=0 (6b)
Once the coding matrix H_{pc }is defined, the encoding of a message frame is relatively straightforward, as known in the art, and can easily be performed by conventional programmable integrated circuits such as digital signal processors and the like. According to the preferred embodiments of the invention, the circularly shifted identity matrices are tiled within macro matrix H in an irregular manner, as will be described below, to provide excellent coding performance.
On the decoding side, one can define a set N(m) as the set of all bit indices (columns) in a given row m for which codeword bits contribute to the checksum (i.e., all bit indices for which the entries of parity check matrix H_{pc }in row m are 1). The checksum equation for a row of the parity check can be expressed as:
over GF(2) or, logically, the exclusiveOR of the input bits c_{j }that correspond to column bits in the row having a 1 value. One can thus determine, for a given codeword vector c, whether an error is present by determining whether this equation is true for each row of the parity check matrix H_{pc}.
In practice, however, the actual input bit values r_{j }that are recovered after demodulation and that are to be interpreted as codeword vector c by a decoder, for example by decoding function 28 in transceiver 20 of
which can of course take negative and positive values, corresponding to 1 and 0 being more likely, respectively. For this description of the preferred embodiment of the invention, one can assume that the incoming LLRs (i.e., the received data) have the form −2r_{j}/σ^{2 }where σ^{2 }represents channel noise variance.
Fundamentally, the LDPC decoding process according to the preferred embodiments of the invention involves an iterative twostep process:

 1. Estimate a value R_{mj }for each of the j input nodes, for each of the m rows of the checksum, using the current probability values from the other input nodes, setting the result of the checksum for the row to 0; and
 2. Update the sum L(q_{j}) for each of the j input nodes from a combination of the m values of R_{mj }in the same column.
The iterations continue until a termination criterion is reached. A preferred termination criteria is the earlier of (i) evaluation of the matrix operation H_{pc}·c=0 (mod 2), using “hard” decisions from the LLRs L(r_{j}) as the codeword vector c, and (ii) completion of a specified number of iterations.
Mathematically, for the first step of estimating values R_{mj }for each of the j input nodes, for each of the m rows of the checksum, one can derive an amplitude A_{mj }and a sign value s_{mj }as follows:
where the function Ψ is defined as:
Ψ(x)≡log(tan h(x/2))=log(tan hx/2) (10)
The function Ψ is its own negative inverse: Ψ(Ψ(x))=−x. For computational ease, one can express tan h(L/2) as:
The sign is determined from:
which is simply an odd/even determination of the number of negative probabilities, excluding each row's own contribution. The updated estimate of values R_{mj }is thus:
R_{mj}=−s_{mj}Ψ(A_{mj}) (13)
The negative sign of value R_{mj }contemplates that the function Ψ is its own negative inverse. The value R_{mj }thus corresponds to an estimate of the LLR for input node j as derived from the other input nodes in the mth row of the parity check matrix, except input node j itself.
An alternative computation of the estimate values R_{mj }can be defined as a special summation:
where the LLR addition [+] is defined as:
This alternative determination of the estimate values R_{mj }may be easier to implement into some integrated circuit architectures. The selection of the computations may be made by those skilled in the art having reference to this specification, and confronted by a particular technology and decoding application.
In the second step of each decoding iteration, the LLR estimates for each input node are updated. For each column (i.e., each input node):
where the set M(j) is the set of all check sum indices (rows) for a given column j of check sum equations to which input bit j contributes to the checksum (i.e., all row indices for which the entries of parity check matrix H_{pc }in column j are 1). This operation effectively sums the estimated values R_{mj }over the jth column, and adds in the original received input node value
to form the best full estimate of the LLR for input node j in this iteration.
This column estimate will be used in the hard decision check. In preparation for the next iteration, the perrow (or extrinsic) LLR probabilities are then derived:
L(q_{mj})=L(q_{j})−R_{mj } (17)
for each column j in each row m. The perrow probabilities thus amount to an estimate for the probability of the input value, excluding the contribution to the estimate for each row from the row itself.
As noted above, the determination of whether the iterations have converged to an error free decoding is based on the percolumn LLRs L(q_{j}):
L(q_{j})≧0, c_{j}=0 (18a)
L(q_{j})<0, c_{j}=1 (18b)
The codeword vector c={c_{0}, c_{1}, . . . c_{N}} is then applied to the parity check matrix H to determine if the product is zero, which as noted above indicates that codeword vector c is a valid codeword, from which the payload data portion can be readily extracted.
In practice, for those value arrays R_{mj}, L(q_{mj}), A_{mj }that are used in the algorithm, the computations performed and the nonzero array values occur only at those index positions (m,j) in parity check matrix H_{pc }, where a “1” value appears. Also in practice, the initialization of the array L(q_{mj}) can be arbitrary. For example, each of the L(q_{mj}) values can be initialized to zero, or to the input values −2r_{j}/σ^{2}, as desired.
According to the preferred embodiment of the invention, an LDPC code is used for encoding data to be transmitted which has the benefits of performing near the Shannon limit, while being implementable in a very efficient manner. The encoding function according to this preferred embodiment of the invention will be described in detail below. However, it is the decoding operation that requires significant computational power, considering that the received datastream is expected to include some frequency of errors, as evident by the use of redundant coding to begin with. Accordingly, the computational complexity and performance of decoding circuitry has significant practical importance, and is often the deciding factor in whether a particular code is practically useful. It will therefore be useful, for purposes of this description, to first describe the code in terms of the decoding algorithm and architecture, prior to describing the encoding function itself, following a description of the code derivation in general terms.
In process 42, the largest input node degree, or variable degree, for the code is selected. As known in the art, this maximum input node degree corresponds to the largest number of checksum rows that involves a given input node; for LDPC codes such as used in this embodiment of the invention, this input node degree corresponds to the maximum number of parity checks that any input node contributes to. While larger input node degrees are desirable, for better error correction capability, the input node degree is ultimately constrained by the hardware implementation. In the examples of the preferred embodiments of the invention described in this specification, the largest variable degree is contemplated to be on the order of ten to twenty.
In process 44, the degree distributions within parity check matrix H_{pc }are optimized. These degree distributions refer to (i) the input node degree distribution, which is the number of check nodes that each input node contributes to, and (ii) the check node degree distribution, which is the number of input nodes that each check node receives. Regular LDPC codes are those for which all nodes of the same type have the same degree. According to the preferred embodiment of the invention, however, irregular LDPC codes are used, to improve error rate performance, and as such each code will have a distribution of degrees over its input nodes, and possibly also over its check nodes. A preferred example of an optimization tool for process 46 is a webbased tool known as “ldpcopt”, which is readily available at http://lthcwww.epfl.ch/research/ldpcopt/ and is described in Richardson et al., “Design of CapacityApproaching Irregular LowDensity ParityCheck Codes”, Transactions on Information Theory, Vol. 47, No. 2 (IEEE, February 2001), pp. 619637; and in Chung, et al., “Analysis of sumproduct decoding of lowdensity paritycheck codes using a Gaussian approximation”, Transactions on Information Theory, Vol. 47, No. 2 (IEEE, February 2001), pp. 657670; both incorporated herein by this reference. The optimization criteria used by this exemplary process minimizes the bit error rate for the code, by varying the variable degrees of the LDPC parity check matrix. The results of the optimization provide the optimum numbers of “1” values (but not position) in parity check matrix H_{pc}, in the form of optimum numbers of “blocks” within its defining macro matrix H.
According to the preferred embodiments of the invention, as briefly mentioned above, the parity check matrix H_{pc }is determined through the use of a macro matrix H. Macro matrix H is a matrix, of dimensions c by d, in which each entry (“block”) indicates the position of a p×p matrix in parity check matrix H_{pc}, and in which the value of each entry symbolizes the form of the corresponding p×p matrix at that position. As a result, parity check matrix H_{pc }has a total number of rows m=p×c and a total number of columns j=p×d. Each row of macro matrix H thus defines a “block row” of p rows in parity check matrix H_{pc}; conversely, each column of macro matrix H defines a “block column” of p columns in parity check matrix H_{pc}. As such, this description will refer to rows of macro matrix H as “block rows”, and columns of macro matrix H as “block columns”. In the preferred embodiments of this invention, a zero value of an entry in macro matrix H symbolizes a zerovalued p×p matrix (i.e., all entries in the p×p matrix are zero). A nonzero value (e.g., 1) of an entry in macro matrix H_{pc }symbolizes the location of a permutation matrix, which in the preferred embodiments of the invention is a cyclically (or circularly) shifted identity matrix.
The optimization of process 44 determines, for a given macro matrix H of c block rows and d block columns, the distribution of the “1” values within the matrix. Specifically, the distribution determines the number of block columns of macro matrix H that have each of a range of numbers of “1” values, and the number of block rows of macro matrix H that have each of a range of numbers of “1” values. By way of definition, the block rows and block columns refer to the illustration of macro matrix H that indicates the locations of the permutation matrices by “1” values. Once the optimization of process 44 is performed, the specific size of c block rows and d block columns is selected in process 46 as shown in
A first example has the following parameter values for a code rate of ½:
In this example, the optimization of process 44, performed by way of the “ldpcopt” program, provides the following distributions of column blocks:
In this table, the value g is the number of “1” bits in a given block column. As shown in this example, the optimization indicates that the optimum distribution of its twentyfour block columns (d=24) would have 10.80 block columns with two “1” bits, 8.9 block columns with three “1” bits, 0.74 block columns with four “1” bits, and 3.56 block columns with eleven “1” bits. Of course, this arrangement is not possible for a single macro matrix H having discrete rows and columns. As such, the “Useful” values in this table have rounded the optimum value to integer values.
Optimization process 44 also provides an optimized allocation of row blocks, by providing an optimized selection of the number of block rows that have varying numbers of “1” values. In this same first example, the optimization process provided the following results:
As evident from this table, the optimum distribution provided for 5.56 block rows with seven “1” bits in each row, and 6.64 block rows with eight “1” bits in each row, for a total of c=12 rows. The wide variation for the “Useful” allocation from the optimum is due to the number of filled “1” blocks defined for the columns, which is incompatible with the number of filled blocks defined by the optimized allocation of block rows. In this example, the “Useful” column allocation establishes that there are eightysix filled block matrix entries; the only combination of row allocations that satisfies that optimized column allocation, while maintaining either seven or eight filled “1” blocks in each row, is the “Useful” allocation of ten block rows with seven “1” bits each, and two block rows with eight “1” bits each, as shown in this table. Alternatively, one could arrange the matrix to satisfy the optimum row allocation and adjust the column allocation.
According to a second example, for a code rate of ⅓, the parameters are as follows:
In this example, the optimization of process 44, performed by way of the “ldpcopt” program, provides the following distributions of column blocks:
This optimum distribution of the twentyfour block columns (d=24) is approximated, in this case, by relatively close rounding of the optimum counts to the nearest integer. The row allocation in this example is as follows:
In this case, the optimum distribution of the sixteen (c=16) block rows is also closely approximated by rounding to the nearest integers, as shown in the “Useful” row of the table. In this case, the optimized block row and block column allocations are compatible.
In process 46, if not previously selected prior to optimization process 44, the multiplicative orders c, d are defined, in which order value c and order value d correspond to the number of block rows and block columns, respectively, in macro matrix H, as described above.
In process 48, the code designer next constructs the particular arrangement of macro matrix H according to the optimization results of process 44, and according to additional constraints according to this embodiment of the invention. According to the preferred embodiments of the invention, the macro matrix H is arranged in groups of block columns, so that, for any given block row in macro matrix H, at most one block column within each group has a filled “1” entry. This arrangement must take into account the manner in which the constituent matrices are developed.
As noted above, generator a is selected to have multiplicative order d, and generator b has multiplicative order c. In process 50, these cyclically shifted identity matrices are then generated for each of the filled “1” locations of macro matrix H, to produce parity check matrix H_{pc}. As mentioned above, parity check matrix H_{pc }thus has m=p×c rows and j=p×d columns, with the “1” values determined by generators a, b. The manner in which the permutation matrices are formed is described in Sridhara, et al., “Low Density Parity Check Codes from Permutation Matrices”, 2001 Conference on Information Sciences and Systems (Johns Hopkins University, March 2123, 2001), incorporated herein by this reference.
In contrast to the Sridhara approach, however, macro matrix H according to the preferred embodiments of this invention is irregular, in that it is itself relatively sparse, including many zerovalued entries. As will be evident from the following description, the irregularity included by the code designer in process 48, in combination with grouping of block columns of macro matrix H, provides a high performance LDPC code that can be efficiently implemented in hardware.
It is also contemplated, in connection with this invention, that the cyclic shifts of the identity matrices corresponding to the row and column position of macro matrix H need not follow this particular approach. Indeed, the offset P_{s,t }of the cyclic shift within a particular permutation matrix can be an arbitrary selection. Such arbitrary assignment, if followed, may affect the ultimate dimensions of macro matrix H.
The examples of
Referring now to
As shown in
Router circuitry 58f is a bank of multiplexers and demultiplexers, as will be described in further detail below, that forwards the appropriate estimate terms R_{mj }to the corresponding bit update circuits 60. Bit update circuits 60 are effectively accumulators, by way of which current values of the LLRs of the input nodes are maintained from iteration to iteration. The number of bit update circuits 60 depends upon the maximum number of groups of block columns of macro matrix H. In the particular code; for the code example of
Bit update blocks 60 receive the input node data values, prior to the first iteration of the belief propagation. As mentioned above, the received input node data values are typically based on a multiple bit fractional value, for example expressed in six or seven bits, as produced after demodulation and recovery after frequency domain equalization (e.g., by functions 26 and 27 of
the value r_{j }being the received data value. Bit update blocks 60 each forward an output to reverse router circuitry 58r, which in turn applies the output value to parallel adder 54, as minuends for the subtraction performed by that circuit. In addition, the outputs of bit update blocks 60 are also applied by reverse router circuitry 58r to parity check function 61, which performs a slicing function on these estimates, and after converting these values to “hard” decisions, determines whether the parity check equation is satisfied by the current estimates for each row of parity check matrix H_{pc}.
Referring now to
the value r_{j }being the received data value, stored in the appropriate memory locations as the initial estimate of the LLRs for the columns.
This description of the operation of LDPC decoder 38, and the detailed construction of its functional blocks, for this exemplary iteration, will begin at parallel adder 54. Parallel adder 54 receives the current estimates R^{i}_{mj }for the participating columns in a given row m of macro matrix H, from memory 52. These current estimates R^{i}_{mj}, which were generated in the previous iteration i of the process, are subtracted from the current LLR values L(q_{j}) of the input nodes that participate in the current row m of parity check matrix H_{pc}, thus generating the LLR values L(q_{mj}) according to Equation (17) described above. These values L(q_{mj}) are forwarded to parity check update circuitry 56, the construction of which will be described in detail relative to
The updating process as carried out by parity check update circuitry 56 begins with the application of each of the values L(q_{mj}) in the input word received from parallel adder 54 to lookup tables (LUTs) 80, by way of which the Ψ function of Equation (10) is evaluated for each of the values L(q_{mj}) within the current row m. The outputs of LUTs 80 are forwarded to augmented adder tree 82, which performs the summing of the values Ψ(L(q_{mj})) over all of the columns participating in the current row m of macro matrix H. Augmented adder tree 82 effects this summation in a manner that can be readily implemented by those skilled in the art having reference to this specification. This overall sum result is applied to an input of adders 86, one adder 86 associated with each of the columns j contributing to the current row m. Each adder 86 also receives, at a negative input, the output of its corresponding LUT 80, and thus subtracts the column's own contribution from the overall sum. The outputs of adders 86 thus present the set of amplitude values A_{mj }corresponding to the result of Equation (9), each associated with one of the columns j that are participating in this row. The outputs of adders 86 are then again applied to corresponding LUTs 88, to again apply the Ψ function to the amplitude values A_{mj}, according to Equation (13). Sign correction functions 90 apply the appropriate sign to the output of LUTs 88, based on the logical combination of the sign bit outputs of LUTs 80 for the corresponding column according to an odd/even determination of the number of negative probabilities, excluding each row's own contribution, as described above relative to Equation (12), and effecting the negative sign applied according to Equation (13). It is this handling of the sign bit outputs of LUTs 80 that corresponds to the augmented addition performed by augmented adder tree 82. Sign correction functions 90 thus present each of the updated estimate values R^{i+1}_{mj }as updated for this, the i+1 iteration.
Alternatively, parity check update circuitry 56 may instead follow the approach described above in connection with Equations (14) and (15). According to this approach, LUTs 80, 88 for applying the W function are not used, but instead a special addition function [+] of Equation (15) is applied to the L(q_{mj}) values, and these values are summed according to Equation (14), to derive the updated estimate values R^{+1}_{mj }for the i+1 iteration. Variations of these parity check update approaches, and other alternative parity check update approaches, may also be realized within parity check update circuitry 56, within LDPC decoding circuitry 38 according to this invention. The selection of the particular arithmetic approach will depend upon the available circuitry and performance of the specific implementation.
Referring back to
The second step of the belief propagation algorithm, in which the input node estimates (in the form of the LLRs) are updated, begins with the application of these updated estimate values R^{i+1}_{mj }to the appropriate bit update circuit 60, via router and reverse router circuitry 58. The construction of router and reverse router circuitry 58 preferably depends upon the particular code arrangement, either by way of hard wiring or alternatively by way of a software controlled logic arrangement.
Routing circuitry 58f thus consists of a set of multiplexers 92, which effect the forwarding of the values R_{mj }of the output word to the appropriate bit update circuit 60. Knowledge of the particular code arrangement within macro matrix H defines the control of these multiplexers 92 because, according to the preferred embodiments of this invention, macro matrix H is constructed with column block grouping, by way of which only one possible column of parity check matrix H_{pc }is involved within each column block group, for any given row. The example of routing circuitry 58f illustrated in
In this example, the leftmost column block group of macro matrix H of
In any event, router circuitry 58f forwards the most recent iteration of estimates R^{i+1}_{mj }to the appropriate bit update circuits 60. Bit update circuits 60 accumulate these estimates R^{1+1}_{mj }with the estimates for the same input node in the same column j, from different rows, as will now be described relative to
As shown in
In operation, with reference to Equation (16), column sum memories 66A, 66B of bit update circuits 60 accumulate the sum of the estimated values R_{mj }for its corresponding input node, which is associated with a corresponding one of the blocks of columns in the appropriate code. This accumulation is carried out by adder 62 receiving the most recent estimate R^{i+1}_{mj }at one input, and receiving the current accumulation of estimates R_{mj }for the same column j, but for different rows m, from one of column sum memories 66A, 66B, selected via crossswitching multiplexer 72. Adder 62 combines these values, and forwards the sum back to the selected column sum memory 66A, 66B, which rewrites the accumulated sum for that row and column position, expressed as the sum
for iteration i+1, following Equation (16). This value is retained in the selected one of column sum memories 66A, 66B, at a memory location associated with the corresponding input node, as addressed by address generator circuit 68.
Address generator circuit 68 includes the appropriate logic and memory circuitry for maintaining and applying memory address values associated with the input nodes managed by bit update circuit 60. According to this preferred embodiment of the invention, each of the permutation matrices involved in generating parity check matrix H_{pc }from macro matrix H are circularly shifted identity matrices, with the particular position of the identity diagonal varying with the position of the permutation matrix within macro matrix H. Because the rows within parity check matrix H_{pc }are accessed in sequence and because the permutation matrices are sliding identity matrices, address generator circuit 68 can address column sum memories 66A, 66B also in a sequence, using relative addressing techniques. Preferably, address generator circuit 68 includes a register for storing an offset address value for the first column position within a particular block row of macro matrix H, with each successive row then having an address that is simply incremented from this offset value, with wraparound, modulo m. For example, if the first row (m=0) has a contribution from input node column position e, the next row (m=1) will have a contribution from column position e+1, and so on. Column sum memories 66A, 66B can thus store the accumulated values in a linear fashion, which greatly simplifies the construction of address generation circuit 68. Of course, the offset values will change when a new block row of macro matrix H is operated upon.
Alternatively, the permutation matrices within macro matrix H can be arranged differently than as circularly shifted identity matrices, so long as each permutation matrix only has a single column entry involved in each row. Address generation circuit 68 will, of course, have to comprehend the particular construction of these permutation matrices so that the memory locations in column sum memories 66 correspond to the proper column locations.
As mentioned above, the use of two column sum memories 66A, 66B, permits the updating of the acculated estimates
for one set of columns associated with one row of parity check matrix H_{pc }while the accumulated estimates
for the previous (i^{th}) iteration are forwarded back to parallel adder 54 for another set of columns associated with a different row of parity check matrix H_{pc}, for example the next successive row of parity check matrix H_{pc }to be updated by parity check update circuitry 56. The architecture of LDPC decoder 38 according to this preferred embodiment of the invention thus provides a high degree of performance efficiency, by way of the pipelining effected by the pingpong arrangement of column sum memories 66A, 66B. In addition, this arrangement greatly improves the utilization of the memory and logic resources of LDPC decoder 38, improving the efficiency with which it can be implemented.
This description will continue for the same row m and its associated columns that was updated in the prior discussion, and as such corresponds to some time in the future, after all updates have been completed for all rows in parity check matrix H_{pc }in iteration i+1. Crossswitching multiplexer 72 forwards the new value of accumulated sums
from the selected one of column sum memories 66 to adder 74. A second input to adder 74 receives the original received input node data value for the column, in the form of the ratio
the value r_{j }being the received data value, as stored in received data memory 70. According to Equation (16), adder 74 therefore generates the current LLR value L^{i+1}(q_{j}), corresponding to the current Log Likelihood Ratio probability for the corresponding input nodes, using the value
for iteration i+1. This value is forwarded to the reverse router circuitry 58r, along with the values for the other corresponding columns that contribute to row m, to parallel adder 54, in the proper order for updating of the R_{mj }estimates. Parallel adder 54 then generates updated R_{mj }estimates as described above.
Because the corresponding updated column LLRs L^{i+1}(q_{j}) are now available, it is convenient to use these values to evaluate this iteration i+1 to determine whether the decoding process has converged upon a valid result, satisfying the codeword equation Hc=0, GF(2). As such, the output word from reverse router circuitry 58r is also applied to parity check function 61, which performs the function of “hard slicing” the current LLR values L(q_{mj}) to digital bit values 0 and 1, and then exclusiveORing these bits to determine whether a 0 result is returned for that row m. It is contemplated that the circuitry or program instructions (e.g., executable by baseband processor 39) for performing this determination can be readily derived by those skilled in the art, having reference to this specification. The results of function 61 are retained over all rows in macro matrix H, so that the ultimate termination decision can be made.
This iterative process then continues, until parity check function 61 determines that the current LLRs of the codeword input nodes match a valid decoding, by returning a “0” value when applied to all of the rows of parity check matrix H_{pc}, or until another termination criterion is met. According to the preferred embodiments of the invention, any validly decodable codeword c explicitly contains the payload data itself, i.e., the code is systematic. Baseband processor 39 or other circuitry in transceiver 20 then retrieves the payload data from the hard decision codeword, and the process is repeated for the next codeword or data block.
As noted above, this embodiment of the invention provides for excellent LDPC coding performance, with error rates approaching the Shannon limit, while still being readily and efficiently implementable in custom logic or in a sequence of program instructions. For example, several codes constructed in this manner have been investigated for their performance. For examples of rate ½ codes corresponding to
According to a second preferred embodiment of the invention, this construction of the LDPC codes enables the use of parallel circuitry to greatly increase the performance of the LDPC decoding. More specifically, this parallelism is attained because the permutation matrices are constructed from a diagonal matrix (the identity matrix), so that incrementing of a row position within the permutation matrix means that the column position of the nonzero value also increments. The circularly (or “cyclically”) shifted identity matrix is an example of permutation matrices with this property. The parallelism effectively performs the single row parity check update and column bit update steps in parallel for multiple rows within the same row block.
According to this embodiment of the invention, multiple R_{mj }memories 152 are provided. Each of memories 152 are dual port RAMs, and in this case have thirtysix output bits, and are thirtytwo in number (RAMs 152_{1 }through 152_{32}). The number of RAMs 152 need not correspond to the number of rows of parity check matrix H_{pc }that are being operated in parallel, but rather is preferably selected for convenience of the integrated circuit layout and manufacturing technology; indeed, the number of RAMs 152 is entirely arbitrary. In this example, twentyfour rows of parity check matrix H_{pc }will be simultaneously processed, from thirtytwo RAMs 152. Accordingly, the bit width of the output of each of RAMs 152 need not match that of the number of bits to be processed for a given row. In this example, each estimate occupies six bits, and up to eight columns may contribute to a row of parity check matrix H_{pc}. As such, the thirtysix output bits from RAM 152_{1 }and twelve output bits from RAM 152_{2 }are applied to first parallel adder block 154_{1}, while the remaining twentyfour output bits from RAM 152_{2 }are applied to second parallel adder block 154_{2}.
In this example, where twentyfour rows are processed in parallel, twentyfour instances of parallel adder blocks 154_{1 }through 154_{24 }are provided, each receiving eight possible R_{mj }values from the corresponding R RAMs 152, along with a corresponding output of reverse router circuitry 158r corresponding to the LLR L(q_{j}) as described above. Parallel adder blocks 154 are constructed similarly as described above relative to
Parity check update blocks 156_{1 }through 156_{24 }are constructed similarly as described above relative to
Router circuitry 158f routes the corresponding estimates R_{mj }to the appropriate bit update unit within its corresponding bit update block 160_{1 }through 160_{9}. In this example, nine bit update blocks 160_{1 }through 160_{9 }are provided, for updating and maintaining the sums
for the respective columns. According to this embodiment of the invention, all of the twentyfour rows being processed in parallel are from the same block row of macro matrix H, and as such the routing for all of the rows is identical. As such, the output value R_{mj }for corresponding bit update positions from each of parity check update blocks 156_{1 }through 156_{24 }are all routed to the same bit update block 1601. For example, the first one of the outputs of each of parity check update blocks 156_{1 }through 156_{24 }may be routed by router circuitry 158f to first bit update block 160_{1}, the second one of the outputs of each of parity check update blocks 156_{1 }through 156_{24 }may be routed to second bit update block 160_{2}, and so on. The number of bit update blocks 160 (nine in this instance) is typically larger than the number of possible outputs from parity check update blocks 156, and corresponds to the number of block column groups of macro matrix H. The number of bit update blocks 160 is preferably not much larger than the number of outputs from parity check update blocks 156, however, so that router circuitry 158f can be relatively simple and efficient. This efficiency is a result of the construction of the LDPC code, according to the preferred embodiments of the invention, in which the order of terms, for example from left to right, in the outputs of any parity check update block 156 remains fixed after routing to bit update blocks 160.
Each of bit update blocks 160 is constructed and operates similarly as described above relative to
In the example where circularly shifted identity matrices are used in macro matrix H as described above, the memory address offset e of the first column contributing to the first row of a block row will vary from permutation matrix to permutation matrix. These changes in the address offset e must of course be considered in this embodiment of the invention, as it is in the single row case described above. However, the parallelism of this embodiment of the invention provides an additional complexity that must be comprehended by the address generator circuitry within each of bit update blocks 160, as will now be described.
While the offsets e associated with each of the nonzero entries in macro matrix H will vary, the number of rows being processed in parallel is constant, namely twentyfour in this example. Because the offset e does not necessarily vary according to the parallelism multiple (e.g., twentyfour), the same input node column will likely appear in a different position from one nonzero permutation matrix entry to the next. This situation is illustrated in
However, considering the construction of the permutation matrices within macro matrix H, the column offset e for the column input node associated with the first row of a second successive block row (x+1 in
The belief propagation algorithm requires the summing of the R_{mj }estimates for the same column nodes, though. Because the multiple parallel sums are being operated on simultaneously by bit update blocks 160, the sums
for a column node j that is updated in different block rows will thus not align, and thus will not be readily summable by bit update blocks 160. Accordingly, the address generator circuit must contemplate this mispositioning of the columns within the parallel group. From the viewpoint of memory addressing, this problem amounts to a misalignment of the data words corresponding to the columnbased values. Other hardware must also comprehend this nonalignment. An example of such hardware includes input aligner 63 and output aligner 75 as shown in
Referring now to
In process 214, bit update block 160 retains data word portion 200A(0) in a local register or store; as mentioned above, this portion 200A(0) is that portion of data word 200(0) that is ahead of the word boundary for block row x. In process 216, bit update block 160 retains data word portion 200B(0) in a separate (or separately accessible) local register or store; also as mentioned above, this portion 200B(0) is that portion of data word 200(0) that follows the block row x word boundary.
In process 218, next data word 200(k+1) is then received, which, at this point, is data word 200(1) because index k=0. Data word 200(k+1) is illustrated in
According to this embodiment of the invention, in process 222, data word portion 200A(1) is combined with data word portion 200B(0) that was retained in process 216. This combined data word 200B(0):200A(1) constitutes an aligned data word, and is aligned with the same data word alignment as that of block row x. Because, according to this embodiment of the invention, all block rows are similarly processed, this results in all data word alignments among all of the block rows being coaligned. As such, the same column input node will have a consistent data word position, over segments of rows within the block row of macro matrix H in which it participates. In process 224, this combined data word is then processed, and its contributions to the various column input nodes summed, as described above relative to
In decision 225, LDPC decoder 138 determines whether additional rows remain within the current block row (x+1) remain to be processed. If so (decision 225 is NO), index k is incremented in process 226, and processes 218 et seq. are repeated for the next row. If the current block row is complete (decision 225 is YES), then only the initial leading data word portion 200A(0) and the most recent trailing data word portion 200B(k) remain to be processed. These two data word portions are combined into data word 200A(0):200B(k+1) in process 228, and is processed for update of the column input nodes, in similar manner as described above, in process 230.
Similar management of the misaligned data words must also be performed in handling data words as they are being output by bit update circuits 160, following the same concept as that described above for the incoming data words. It is contemplated that those skilled in the art having reference to this specification will be readily able to effectively reverse the process for the outputting of results from bit update circuits 160.
In addition, another source of misalignment is also present within each permutation matrix. As discussed above, the permutation matrices are square matrices, of dimension p by p. Because p is prime, however, the parallelism factor will necessarily not evenly divide into the number of rows of the permutation matrix. This means that a partial data word will necessarily result at some point within the processing of a block row; this partial data word may occur at any point within the block row, or permutation matrix. Preferably, the overall architecture of LDPC decoder 38 is constructed and controlled to efficiently deal with these occasional partial words. It is contemplated that those skilled in the art having reference to this specification will be readily able to manage these partial word events.
Referring back to
According to further alternative embodiments of the invention, additional flexibility and memory space efficiency can be provided by arranging memory 52, 152 in multiple blocks. The efficiency of the memory implementations is somewhat complicated by the irregular nature of macro matrix H, considering that the weights, or degree, of the parity check rows varies from row to row; in other words, the number of columns in each row that participate in the parity check equation varies from block row to block row. In the example of
According to many variations of the preferred embodiments of the invention, however, the variations in row degree can be readily incorporated into the implementation and access of memories 52, 152. Specifically, the following description will refer to alternative implementations of memory 52 as used for the single rowatatime operation described above relative to
According to one alternative implementation, memory 52 may be simply sized to correspond to the largest possible degree, times the number of bits used to store each R_{mj }value. For example, if the largest row degree to be contemplated is nine, and if six bits are used for each value, the width of memory 52 may be set to fiftyfour bits. For those rows that have fewer than nine values, the remaining entries in the particular memory word in memory 52 may be filled with the next row's values. This approach may be generalized so that the data values for a first row of a block row of macro matrix H may be stored in a row in memory 52; the remainder of that row of memory 52 may be used to at least partially store data values for the next row of the block row of macro matrix H, and the remainder of the data values in that next row of the block row of macro matrix H are stored from the beginning of the next row in memory 52. It is contemplated that the memory access control of the row values stored in memory 52 according to these alternatives can be effected by an arrangement of multiplexers and registers, as will be apparent to those skilled in the art having reference to this specification. Indeed, if the values for each row of macro matrix H fits, on the average, within a single row of memory 52, it is contemplated that this improved packing may be implemented without involving extra memory cycles to be used in operation.
Further in the alternative, memory 52 may be sized to correspond to typical row degrees, but with the understanding that some rows of macro matrix H may have an extra term that does not fit within a single row of memory 52. According to this additional alternative, an auxiliary memory may be provided to store the additional value or values for those rows of macro matrix H. This alternative approach may be particularly beneficial to provide code flexibility in LDPC decoder 38, where memory 52 is sized to handle most code sizes but where the flexibility to handle some codes that involve large row degrees for some rows in macro matrix H is desired.
An important benefit of this invention is its flexibility in applying the LDPC decoding circuitry to a variety of code sizes. In general, two different approaches may be followed in order to handle a code in which the maximum row size (i.e., number of nonzero entries in a given row of parity check matrix H_{pc}) is greatly increased. In one approach, the operation of the decoding circuitry is timesequenced to handle the doublesized rows in two cycles. In another approach, portions of the decoding circuitry are doubled and their results handled in parallel, permitting each row to be handled in a single cycle. According to the timesequencing, or time doubling, approach, the construction and operation of an example of parity check update circuitry 256 will now be described relative to
At the output of adder tree 282, twostage accumulator 284 accumulates the sequential partial sums through LUTs 280 and adder tree 282 into a full sum for the matrix row. The construction of twostage accumulator 284 is illustrated in further detail in
Subtracting adders 286 receive sequential inputs from LUTs 280 to generate each subtracted value that is applied in the two input data words. As noted above, the full sum is generated over two successive cycles by adder tree 282 and twostage accumulator 284. Two sets of LUT 280 results are used for sequential subtractions, and as such LUT output registers 285 are provided to store the LUT outputs in sequence. Two output registers 292a, 292b are optionally provided to store the two portions of the output result, for forwarding in parallel to corresponding inputs of the bit update circuitry 60, 160 with registers 292a, 292b being logically in parallel with one another. Alternatively, the sequential outputs can be sequentially forwarded in two separate portions, in which the column input nodes are updated in separate cycles.
As a result, in this embodiment of the invention in which a single parity check update block 256 operates over two passes, the data path is made effectively wider, and capable of handling a wider set of values from the parallel adder block. This permits a great deal of flexibility in handling a wide range of code rates with the same LDPC circuitry, thus permitting programmability in the encoding and decoding processes.
Referring now to
In the example of
Of course, more than two such parity check update functions 56 could be used in parallel to handle a single ultrawide row, if desired. For example four parity check update functions 56 could be grouped in this manner, thus providing a x4 parallel capability. In such an arrangement, it is contemplated that the corresponding control circuitry could selectably process four shorter rows at once (one from each parity check update unit), one ultrawide row by combining the sum from all four parity check update units, or a by2 operation in which the parity check update units are paired, with each pair handling a wide row of up to 2× width. It is contemplated that the detailed construction and control of these and other alternative implementations will be apparent to those skilled in the art having reference to this specification.
As mentioned above, an important benefit of this invention is its great flexibility in applying the LDPC decoding circuitry to a variety of code sizes, either by timesequencing or by the simultaneous use of multiple portions of the decoding circuitry. In the context of parallel construction such as shown in
For the time sequencing approach, referring back to
In this time sequencing approach, there may not be enough bit update blocks 160 to handle a full row, but it is contemplated that there will be at least half of the required number. Considering that each separate data item from parity check update blocks 156 originate from a different column grouping, and that each such item will likely be routed to a different bit update block 160, bit update blocks 160 will also be time sequenced in this approach. More specifically, in the first of the two paired cycles, router circuitry 158f will route the first set of partial row outputs from parity check update blocks 156 to their correct bit update units 160, and in the second of the two cycles, router circuitry 158f will route the second set of partial row outputs from parity check update blocks 156 to their correct bit update units 160, which are likely to differ from those for the first set of partial row outputs. In turn, bit update units 160 will be operating on entries from one nonzero submatrix of macro matrix H in one cycle, and another nonzero submatrix in the next. In order for bit update units 160 to toggle in this fashion, it is contemplated that some circuitry, such as control circuitry and data registers, may be duplicated in bit update units 160.
The construction and operation of the output paths of bit update unites 160 will similarly be constructed and operate, to effect time sequencing on the outputs passing through reverse router circuitry 158r.
According to the other approach, in which the parallelism factor Z is reduced, consider an increase in the code size (row size) by a factor of two. The new parallelism factor Z′ then becomes Z/2. In each cycle, Z′ full rows will be processed in a cycle. According to this embodiment of the invention, neighboring parity check update blocks 156 are paired and combined as described above relative to
In this approach, it is contemplated that the number of bit update blocks 160 are also less than that required by the wider row. As such, bit update blocks 160 are again time sequenced, similarly as described immediately above for the timesequenced operation. In this approach, bit update blocks 160 must additionally have sufficient storage to be able to combine data across cycles to process the data with a higher parallelism factor (Z=2Z′), to match the data flow rate with the computation rate.
More specifically, partial data from each of two submatrices A, B are stored by bit update blocks 160 in a first cycle. In the next cycle, the remaining data for submatrix B are stored and the remaining data for submatrix A are grouped with the partial data for submatrix A from the previous cycle, so that this combined data appears as data having the full parallelism factor (Z=2Z′). In the next cycle, the two stored data items for submatrix B are combined and similarly processed by bit update blocks 160; in addition, because this cycle is the first cycle in the next cycle pair, the next row's first partial data for two submatrices A, B are stored. This pipelinedlike process continues to match the full row data rate.
Again, the output side of bit update blocks 160 are similarly constructed to correspond to the computation rate.
According to these two approaches, therefore, the LDPC decoding circuitry is capable of handling a wide range of code widths. While the described examples handle increased widths up to double the size of one another, it is contemplated that this concept can be readily extended in a similar manner to greater multiples, by those skilled in the art having reference to this specification. The configuration and operation of the LDPC decoding circuitry to accomplish this change in code size is contemplated to be under software control, thus providing transceivers having a wide range of field flexibility.
Referring now to
In the examples of
It is contemplated, relative to the examples illustrated in
In
Also according to this embodiment of the invention, the data widths of blocks 252a through 252d need not be uniform among one another.
According to this set of alternative implementations, it is contemplated that a high degree of code flexibility is provided by this invention, in a manner that maximizes the efficiency of the memory utilization.
LDPC Encoding
As mentioned above relative to
As known in the art, LDPC codes do not immediately result in simple and efficient encoding methods, particularly for real world computation and hardware implementation. Structured approaches to encoder construction may or may not be available, considering that the parity check matrices according to this invention are intended to provide desirable performance characteristics on the decoding side. In general, therefore, one can consider the encoding process to be based on matrix multiplication of the an input vector x (having K information bits) with a generator matrix G to produce the codeword c:
c=G^{T}x (19)
As such, a straightforward encoding process derives generator matrix G for a given macro matrix H and corresponding parity check matrix H_{pc}, and performs the above matrix multiplication. In general, though, this requires NK multiplications (N being the number of bit nodes of codeword c), and nearly that many additions, for the encoding of a single codeword. Efficiencies in that regard are known, and will be mentioned below.
A preferred approach to the derivation of generator matrix G from a known parity check matrix will now be described. It is known in the art that:
H_{pc}·G^{T}=0 (20)
By placing matrices G, H_{pc }in systematic form, the derivation of one from the other is simple. For systematic matrices G^{T}_{sys }and H_{sys}:
where the I_{z }matrices are identity matrices of dimension z. Conversion of parity check matrix H_{pc }into its systematic form is quite straightforward, especially using the MATLAB computing environment available from The Mathworks, Inc., by adding (modulo 2) a copy of any row to another row, and by swapping any two columns; neither operation changes the code or rank of the matrix. A simple modification of the row reduction function of MATLAB can also be used to form an identity matrix at the front of parity check matrix, by using row additions until the process is finished or a zero pivot is reached. If a zero pivot is found, column swaps are performed on the remaining matrix to locate a nonzero pivot. These operations result in the derivation of matrix P.
Once matrix P is derived, then generator matrix G can be readily produced using equation (21). The partitioning of systematic generator matrix G_{sys }that is apparent in equation (21) also defines a partitioning in codeword c that is generated from c=G_{sys}^{T}x , i.e., the first K bits are the K information bits. Also, because of the column swapping that is done in the reducing of parity check matrix H_{pc}, the order of the codeword bits is not correct for the original parity check matrix H_{pc}. Because LDPC decoding circuitry 38 is based on this original form of parity check matrix H_{pc}, the order of bits in codeword c must be reordered according to the order in systematic matrix H_{sys}. Alternatively, this could be accomplished by reordering the rows of systematic generator matrix G_{sys}.
In some cases, the result of the row reduction will cause parity check matrix H_{pc }to not be of full row rank, as including some number of dependent rows. The systematic matrix H_{sys }will simply have all zeros in its bottom rows after the reduction process. This will result in new code size parameters with a larger effective K and a smaller effective M (their sum still equaling N). Generator matrix G_{sys }will still be a valid generator matrix, although it will be encoding fewer information bits (and the true code rate will be slightly increased); parity check matrix H_{pc }will still be a valid parity check matrix, and used for decoding.
In practice, the derivation of generator matrix G_{sys }need only be done once per code, and as such this generation does not enter into the realtime encoding flow. However, as mentioned above, the actual encoding of payload data can be quite complex, if performed by way of a straightforward matrix multiplication of the datastream by generator matrix G_{sys}. This is due in part to systematic generator matrix G_{sys }being typically quite dense, even though parity check matrix H_{pc }is quite sparse. Therefore, while this approach will properly generate the desired codewords, it is contemplated, in connection with this invention, that some amount of efficiency can be implemented in this encoding.
One approach follows that described in T. J. Richardson and R. L. Urbanke, “Efficient Encoding of LowDensity ParityCheck Codes”, IEEE Trans. on Information Theory, Vol. 47, No. 2 (February 2001), pp. 638656, incorporated herein by this reference. In this approach, in its construction, and by way of row and column reorderings, parity check matrix H_{pc }is arranged to be lower triangular over the upper M_{t}×M portion that will multiply against the parity bits in the codeword. Because M_{t}<M, this lower triangular property will end prior to the last few check equations or rows. The parity bits can then be recursively determined over the lower triangular portion, beginning with the first parity bit that is the only remaining unknown in the first parity equation in the first row of parity check matrix H_{pc}. Beginning with this first parity bit, the second parity bit in the second row can be similarly determined, and so on. Upon reaching the end of the lower triangular portion, while more parity bits remain to be derived, this lower portion of the parity check matrix is quite small, and can be inverted to solve the remaining parity bits. Of course, this technique requires the situation in which parity check matrix H_{pc }can be arranged in this form.
Another approach is described in MacKay, et al., “Comparison of
Constructions of Irregular Gallager Codes”, IEEE Trans. on Comm., Vol. 47, No. 10 (IEEE, 1999), pp. 14491454, incorporated herein by this reference. This approach extends the Richardson et al. concept described immediately above, by forming a different form of lower triangular parity check matrix H_{pc}. In this case, the parity bit portion of parity check matrix H_{pc }is lower triangular only over the upper right M_{t}×M portion. While the recursion approach cannot be followed in this form, parity check matrix H_{pc }can be viewed in terms of its major block matrices along natural divisions, from which sparse computations of matrix equations can be used to derive the encoding.
It is contemplated that those skilled in the art having reference to this specification will comprehend the general nature of the encoding process, especially considering these examples. It is also contemplated that those skilled artisans can implement the encoding functions either in terms of a general software implementation or, considering that the modulo 2 multiplications and additions can be trivially implemented in hardware by AND and XOR gates, in terms of custom logic hardware.
In Conclusion
According to all of the preferred embodiments of the invention, a hardware efficient LDPC coding and decoding method and circuitry is provided. The code involves a high degree of irregularity, thus providing excellent performance, nearing the Shannon limit in some cases. This performance is attained while enabling relatively simple and robust hardware implementation, despite the irregular codes.
While the present invention has been described according to its preferred embodiments, it is of course contemplated that modifications of, and alternatives to, these embodiments, such modifications and alternatives obtaining the advantages and benefits of this invention, will be apparent to those of ordinary skill in the art having reference to this specification and its drawings. It is contemplated that such modifications and alternatives are within the scope of this invention as subsequently claimed herein.
Claims
133. (canceled)
34. A network element, comprising:
 transceiver circuitry for receiving signals from the network over a communications channel; and
 processor circuitry, comprising: a host interface for communicating, to a host system, decoded signals corresponding to the signals received from the network; demodulator circuitry, for demodulating signals received from the network into a data stream; and circuitry for decoding the data stream according to a sequence of operations comprising: receiving a set of input values corresponding to input nodes of the macro parity check matrix; for each of the input nodes, over each of a plurality of parity check sums of the LDPC code, estimating a check node value using values of other input nodes contributing to the parity check sum; for each of the input nodes, evaluating a probability value using the estimates of the check node values for that input node; and repeating the estimating steps until reaching a termination criterion.
35. The network element of claim 34, wherein the decoding circuitry comprises programmable logic circuitry;
 and further comprising: program memory for storing a sequence of program instructions; wherein the programmable logic circuitry executes the sequence of program instructions to decode the data stream to perform the estimating and repeating operations.
36. The network element of claim 34, wherein the decoding circuitry comprises:
 a check node memory for storing estimates of check node values associated with each of the input nodes over each of a plurality of parity check sums of the LDPC code;
 a parallel adder coupled to the check node memory, for combining a group of check node values associated with a row of the parity check matrix with probability value estimates for input nodes corresponding to the group of check node values, to produce a plurality of extrinsic estimates;
 a parity check update circuit, for updating the estimates of the check node values using the extrinsic estimates, each updated estimate of the check node values associated with an input node;
 a plurality of bit update circuits, each for updating a probability value estimate corresponding to an input node;
 routing circuitry, for routing each updated estimate of the check node values to the one of the plurality of bit update circuits associated with its corresponding input node; and
 rerouting circuitry, for routing each updated probability value from the bit update circuits to the parallel adder.
37. The network element of claim 34, wherein the transceiver circuitry comprises:
 RF circuitry for performing analog demodulation, amplification, and filtering of RF signals received over the communications channel.
38. The network element of claim 37, further comprising:
 an antenna coupled to the RF circuitry.
39. The network element of claim 34, wherein each permutation matrix corresponding to a nonzero entry of the macro matrix is a cyclically shifted identity matrix.
40. The network element of claim 39, wherein an offset for each of the cyclically shifted identity matrices corresponds to the block row and block column of the permutation matrix in the macro matrix.
4157. (canceled)
58. A node in an orthogonal frequency division multiplexing (OFDM) communications system, comprising:
 a receiver for receiving an OFDM signal stream containing data encoded according to a low density parity check (LDPC) code represented by a macro matrix having zerovalued and nonzerovalued entries arranged in block rows and block columns and in which each zerovalued entry corresponds to a p×p zerovalued matrix and each nonzerovalued entry corresponds to a p×p permutation matrix that has at most a single “1” entry in each row and each column and “0” entries elsewhere to define a parity check matrix, wherein the block columns of the macro matrix are grouped so that at most one column has a “1” entry in any row, and wherein the columns of the parity check matrix correspond to input nodes and the rows of the parity check matrix correspond to parity check sums, the receiver comprising: a demodulator for demodulating signals received from the network into a data stream; and decoder circuitry, coupled to receive the data stream from the demodulator, and comprising: a check node memory for storing estimates of check node values associated with each of the input nodes over each of a plurality of parity check sums of the LDPC code; a parallel adder coupled to the check node memory, for combining a group of check node values associated with a row of the parity check matrix with probability value estimates for input nodes corresponding to the group of check node values, to produce a plurality of extrinsic estimates; a parity check update circuit, for updating the estimates of the check node values using the extrinsic estimates, each updated estimate of the check node values associated with an input node; a plurality of bit update circuits, each for updating a probability value estimate corresponding to an input node; routing circuitry, for routing each updated estimate of the check node values to the one of the plurality of bit update circuits associated with its corresponding input node; and rerouting circuitry, for routing each updated probability value from the bit update circuits to the parallel adder.
59. The network node of claim 58, wherein each of the plurality of bit update circuits is associated with a group of the block columns of the macro matrix;
 and wherein each of the plurality of bit update circuits comprises: first and second column sum memories; a received data memory; an incoming adder, having a first input coupled to the routing circuitry; a demultiplexer, having an input coupled to the output of the incoming adder, and having outputs coupled to the first and second column sum memories; a crossswitching multiplexer, having inputs coupled to outputs of the first and second column sum memories, and having a first output coupled to a second input of the incoming adder; an outgoing adder, having a first input coupled to a second output of the crossswitching multiplexer, and having an output coupled to the rerouting circuitry; and control circuitry, for controlling the addressing of the memories and for controlling the demultiplexer and the crossswitching multiplexer so that incoming data from the routing circuitry is being accumulated by the incoming adder in one of the first and second column sum memories, while the other of the first and second column sum memories is presenting an output to the outgoing adder that is being combined with corresponding contents of the received data memory.
60. The network node of claim 58, wherein the parity check update circuit comprises a plurality of parity check update circuits, for updating the estimates of the check node values over a plurality of rows of the parity check matrix in parallel.
61. The network node of claim 58, wherein successive portions of the extrinsic estimates for a parity check matrix row are applied to the parity check update circuit in successive cycles;
 wherein the parity check update circuit is for combining the successive portions of the extrinsic estimates to produce updated estimates of the check node values for the parity check matrix row;
 and wherein one or more of the plurality of bit update circuits processes updated estimates of check node values for a first portion of the parity check matrix row in a first cycle, and processes updated estimates of check node values for a second portion of the parity check matrix row in a later cycle.
62. The network node of claim 58, wherein successive portions of the extrinsic estimates for a parity check matrix row are applied to the parity check update circuit in successive cycles;
 wherein the parity check update circuit comprises: a first lookup table for producing first function values from extrinsic estimates for a parity check matrix row; an augmented adder tree for generating a sum of the first function values; a plurality of adders for applying corresponding ones of the first function values to the sum; a second lookup table for producing second function values from the outputs of the plurality of adders; sign correction functions for correcting the sign of the second function values from the sum, to produce the parity check values for the parity check matrix row; and a twostage accumulator, at the output of the augmented adder tree, for accumulating successive sums into a full sum; wherein the first lookup table, the second lookup table, the plurality of adders, and the sign correction functions operate on successive data portions for the matrix row in successive cycles, the plurality of adders using the full sum from the twostage accumulator; so that the parity check update circuit generates successive portions of the parity check values for the matrix row in successive cycles.
63. The network node of claim 58, wherein the check node memory is arranged in rows and columns;
 and wherein the check node memory is for storing the estimates of check node values for a first row of the parity check matrix in a first row of the check node memory, and also for storing at least some of the estimates of check node values for a second row of the parity check matrix in the first row of the check node memory.
64. A low density parity code stored by a storage media, comprising:
 a macro matrix having zerovalued and nonzerovalued entries arranged in block rows and block columns and in which each zerovalued entry corresponds to a p×p zerovalued matrix and each nonzerovalued entry corresponds to a p×p permutation matrix that has at most a single “1” entry in each row and each column and “0” entries elsewhere to define a parity check matrix,
 wherein the block columns of the macro matrix are grouped so that at most one column has a “1” entry in any row,
 and wherein the columns of the parity check matrix correspond to input nodes and the rows of the parity check matrix correspond to parity check sums.
65. The code of claim 64, wherein each permutation matrix corresponding to a nonzero entry of the macro matrix is a cyclically shifted identity matrix.
66. The code of claim 66, wherein an offset for each of the cyclically shifted identity matrices corresponds to the block row and block column of the permutation matrix in the macro matrix.
Type: Application
Filed: Sep 22, 2009
Publication Date: Mar 3, 2011
Applicant: TEXAS INSTRUMENTS INCORPORATED (Dallas, TX)
Inventor: Dale E. Hocevar (Plano, TX)
Application Number: 12/512,506
International Classification: H03M 13/11 (20060101); G06F 11/10 (20060101);