A High-Parallel GRAND-CRC Decoding Method and System Based on Pruning and Physical Multiplexing

By employing pruning algorithms and physical multiplexing techniques, the storage requirements for noisy error modes are reduced, solving the problem of excessive storage overhead in high-parallelism GRAND-CRC decoders and achieving high-efficiency decoding performance.

CN123074510APending Publication Date: 2026-11-10ZHEJIANG SHUANGHUAN DRIVELINE +1
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611581850.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies in high-parallelism GRAND-CRC decoders have excessive storage overhead for storing noise error modes, resulting in severe consumption of BRAM resources and making it difficult to achieve full coverage under limited resource conditions.

Method used

A high-parallel GRAND-CRC decoding method based on pruning and physical multiplexing is adopted. The storage amount of noisy error patterns is reduced by pruning algorithm, and high-throughput decoding is achieved by using parameterized data and time-series multiplexing physical evolution architecture.

Benefits of technology

It effectively reduces storage overhead, lowers computing resource consumption, improves the parallelism and throughput of the decoder, and solves the storage bottleneck problem.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074510A_ABST
    Figure CN123074510A_ABST
Patent Text Reader

Abstract

This invention discloses a highly parallel GRAND-CRC decoding method and system based on pruning and physical multiplexing. The decoding system includes the following for each decoding channel: a data parsing module for extracting morphological parameters, a first parameter set, and a second parameter set from parameterized data; a flipping module connected to the data parsing module for spatially flipping the first and second parameter sets according to the morphological parameters, outputting multiple sets of reconstructed permutations; a mapping module connected to the flipping module for mapping each set of reconstructed permutations to a corresponding noise mask of a preset bit width; a verification module connected to the mapping module for performing a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and calculating a corrector in the guessed sequence to perform CRC verification; and an output module connected to the verification module of each decoding channel for triggering a decoding completion signal when any verification module passes verification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a highly parallel GRAND-CRC decoding method and system based on pruning and physical multiplexing. Background Technology

[0002] According to the IEC 61784-3 protocol, in functional safety communication fields such as automotive electronics and industrial control, the data transmission framework in the underlying network typically adopts the black channel model, such as... Figure 1 As shown.

[0003] Assume that the errors in the transmission of binary digital information under black-channel conditions are symmetrical. Under electromagnetic interference in an industrial environment, the probability of any information bit flipping in a Security Protocol Data Unit (SPDU) is P. e To ensure that a sufficient amount of interference can be detected, IEC 61784-3 specifies the use of... Used for secure computation and evaluation of residual error probability (REP). Given a SPDU with a total length of N", the probability of k bit errors occurring in an N-bit channel noise is calculated. It strictly follows a binomial distribution, and its mathematical model is shown in the following equation:

[0004]

[0005] To ensure data integrity during black channel transmission, the sending end must sign the transmitted security information frame SPDU using Cyclic Redundancy Check (CRC). The receiving end performs error detection and integrity checks on the received SPDU. Only when the SPDU passes the CRC check is it submitted to the upper application layer. The core purpose of CRC signing is to ensure that the probability of residual errors caused by channel noise is lower than the given security level requirements, thereby constructing the underlying defense line for functionally safe communication.

[0006] Traditionally, the CRC mechanism only performs error detection, not error correction. If the SPDU fails the test, the frame is discarded and the sender is notified to retransmit. In harsh electromagnetic environments, high-frequency retransmissions can force data exchange in the communication system to stop, causing a surge in transmission latency. The GRAND (Guessing Random Additive Noise Decoding) algorithm is a general-purpose algorithm with excellent decoding performance that does not rely on a specific error correction code algebra structure. When the GRAND algorithm is used for CRC error correction, the receiver generates corresponding noise error patterns in descending order of probability based on the binomial distribution of error bits for verification testing. This mechanism endows CRC with powerful error correction capabilities, significantly reducing the system retransmission probability and significantly improving the reliability and security of SPDU transmission.

[0007] While the GRAND algorithm boasts excellent theoretical performance, the key to its practical application lies in designing an extremely efficient hardware decoder. In functional safety scenarios, the system demands exceptionally high real-time performance. Since GRAND requires testing a massive number of candidate error modes, a highly parallel architecture (assuming a system parallelism of Num_lanes) must be introduced to exponentially increase decoding throughput while maintaining decoding performance. Simultaneously, limited by FPGA silicon area and BRAM resources, the hardware resource consumption of the CRC-GRAND decoder must be minimized, necessitating a complete abandonment of traditional storage methods in the underlying hardware architecture.

[0008] In a high-parallelism GRAND decoding framework, to eliminate the need for storing full-length candidate noise error patterns, channel noise error patterns can be uniquely characterized by permutations of consecutive strings of "1"s (denoted as m blocks) and strings of "0"s (denoted as n blocks). Let the total length of the current Security Protocol Data Unit (SPDU) be N", and the total number of error bits (i.e., the total number of "1"s) be L. m The total number of "0"s is L n (and L) n =NL m Different beginning and ending patterns are determined by the combination signal C. α Decide.

[0009] For a maximum of 4 error bits (i.e., L) m For typical cases within ≤4), the complete set of characteristic parameters of channel noise error modes and the total number of corresponding modes. As shown in Table 1:

[0010] Table 1 Channel Noise Error Mode Parameters and Total Number of Arrangements

[0011]

[0012] It should be noted that the total number of 679,118 permutations listed in Table 1 is based on the assumption that... The length of the child elements must be greater than or equal to 1 (i.e., strong combination). This is based on an early theoretical foundation. However, in actual physical channels, in order to fully cover certain continuous burst error patterns, some... Child elements are allowed to have a length of 0. Therefore, the permutation generation algorithm must be modified to a weak combinatorial pattern (allowing...). ).

[0013] When using a true weak combinatorial expansion The formula for the number of combinations is derived from (Let the total number of combinations be...) ) leap to This causes the total number of noise error modes that the system actually needs to cover to expand dramatically from 679,118 to 857,876. If all 857,876 64-bit physical noise masks are directly stored in on-chip memory in the traditional way, the required storage capacity is as high as 52.36 Mbits. This poses a severe challenge to the BRAM resources of hardware platforms such as FPGAs and seriously restricts the deployment of high-parallelism GRAND decoders.

[0014] For example, patent CN121356602A discloses a high-speed parallel GRAND-CRC decoding method for functional safety communication, which includes a distributed memory design strategy and the advantage of unrestricted parallelism values. The parallelism can be set very conveniently according to hardware resources, and the decoder throughput increases linearly with the parallelism. The corrector is used as the basis for logical judgment, and the corrector signal processing unit triggers the correct decoding output. The feature is that there are different processing methods in three cases, especially when there is a contention, the guess sequence corresponding to the corrector with the smallest sequence number is selected as the decoding output.

[0015] However, the existing solutions mentioned above still have the following shortcomings:

[0016] 1. Existing technologies focus on task scheduling and allocation between memory cells, without fundamentally reducing the storage overhead of individual noisy error modes. As parallelism increases, the total number of noisy modes that need to be stored remains unchanged, and BRAM resource consumption increases linearly with parallelism, so the storage bottleneck still exists.

[0017] 2. While existing technologies have proposed scalable parallelism architectures, they have not provided a specific technical path for achieving full coverage of 857,876 massive noise modes with minimal storage cost under limited BRAM resources. When the system needs to support hundreds of parallelism levels, on-chip storage resources will face significant pressure.

[0018] To address the aforementioned issues, this invention proposes a high-parallelism GRAND-CRC decoding method and hardware architecture based on parameterized compressed storage and time-series multiplexing physical evolution. Summary of the Invention

[0019] The purpose of this invention is to address the shortcomings of existing technologies by providing a highly parallel GRAND-CRC decoding method and system based on pruning and physical multiplexing.

[0020] To achieve the above objectives, the present invention adopts the following technical solution:

[0021] A highly parallel GRAND-CRC decoding system based on pruning and physical multiplexing includes:

[0022] The storage module is used to store multiple parameterized data after offline pruning. The parameterized data includes a first parameter set and a second parameter set for reconstructing the noise error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters.

[0023] The global control module, connected to the storage module, is used to read D parameterized data from the storage module in parallel and distribute them to D decoding channels;

[0024] Each decoding channel includes:

[0025] The data parsing module is used to extract morphological parameters, the first parameter set, and the second parameter set from parameterized data.

[0026] The flipping module, connected to the data parsing module, is used to spatially flip the first parameter set and the second parameter set according to the morphological parameters and output multiple sets of reconstructed permutations.

[0027] The mapping module, connected to the flipping module, is used to map each group of reconstructed permutations to a noise mask of a corresponding preset bit width.

[0028] The verification module, connected to the mapping module, is used to perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and to calculate the corrector in the guessed sequence to perform CRC verification.

[0029] The output module is connected to the verification module of each decoding channel and is used to trigger a decoding completion signal when any verification module passes the verification.

[0030] Furthermore, the flipping module performs several spatial flips on the first parameter set and the second parameter set according to the morphological parameters, specifically:

[0031] Four types of space flips are performed in successive clock cycles from the first to the fourth, and four sets of reconstructed permutations are output in sequence.

[0032] Furthermore, the flipping module also includes a TDM state machine and a cross matrix;

[0033] The TDM state machine is used to issue corresponding gating control signals during the first to fourth clock cycles.

[0034] The cross matrix is ​​used to perform head-to-tail swapping on the first and second parameter sets according to the gating control signal.

[0035] Furthermore, the four sets of output reconstruction arrangements include:

[0036] During the first clock cycle, the TDM state machine controls the cross matrix to keep the first parameter set in its original order and the second parameter set in its original order.

[0037] During the second clock cycle, the TDM state machine controls the cross matrix to reverse the first parameter set from beginning to end while keeping the second parameter set in its original order.

[0038] During the third clock cycle, the TDM state machine controls the cross matrix to keep the first parameter set in its original order and reverse the order of the second parameter set.

[0039] During the fourth clock cycle, the TDM state machine controls the cross matrix to reverse the order of the first parameter set and the second parameter set.

[0040] Furthermore, the offline pruning process in the storage module specifically includes:

[0041] Based on either the first or second parameter set, construct a permutation set S. For any permutation A = (a1, a2, ..., a...) in the permutation set S... z ):

[0042] If the first element a1 of permutation A is less than the last element a z If so, then permutation A is removed;

[0043] If the first element a1 of permutation A is equal to the last element a z If the permutation A is not generated, then permutation A is retained and no subpermutations are generated.

[0044] If the first element a1 of permutation A is greater than the last element a1... z If so, then permutation A is retained.

[0045] Furthermore, the total number of valid permutations of the first or second parameter set retained after the offline pruning process is:

[0046]

[0047] in, X represents the total number of valid permutations of the first or second parameter set; P represents the total number of full permutations of the first or second parameter set; and P represents the number of palindromic permutations of the first or second parameter set.

[0048] Furthermore, the mapping module includes:

[0049] The accumulation unit is used to perform accumulation and summation on each length value in the reconstructed arrangement to obtain the absolute starting index of each length value in the preset bit-width space;

[0050] The mask generation unit is used to generate corresponding mask values ​​based on each length value;

[0051] The separation unit is used to perform bit-level shifting and byte alignment on the mask value according to the absolute start index;

[0052] The fusion unit is used to perform a bitwise OR operation on the shifted and byte-aligned mask values ​​to generate a noise mask with a preset bit width.

[0053] Furthermore, the separation unit performs bit-level shifting and byte alignment on the mask value according to the absolute start index, specifically as follows:

[0054] The mask value is bit-level shifted based on the low-order bits of the absolute start index, and multiplexer-based byte alignment is performed based on the high-order bits of the absolute start index.

[0055] Furthermore, the output module also includes: issuing a continue reading instruction to the global control module when all verification modules fail to pass verification.

[0056] Correspondingly, a highly parallel GRAND-CRC decoding method based on pruning and physical multiplexing is also provided, including:

[0057] S1. Parallel reading of D parameterized data from the storage module and distribution to D decoding channels; wherein the storage module stores multiple parameterized data after offline pruning, the parameterized data includes a first parameter set and a second parameter set for reconstructing the noise error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters;

[0058] S2. In each decoding channel, extract the morphological parameters, the first parameter set, and the second parameter set from the parameterized data;

[0059] S3. Perform spatial flipping on the first and second parameter sets according to the morphological parameters, and output multiple sets of reconstructed permutations;

[0060] S4. Perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and calculate the corrector in the guessed sequence to perform CRC check;

[0061] S5. When the CRC check of any decoding channel passes, trigger the decoding completion signal.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. Parameter mechanism based on BRAM interface: Completely abandons the storage of full-length noise error modes, and instead extracts m and A. i B j and combined signals It is pre-stored in BRAM. During decoding, it is read concurrently on demand through the interface, which greatly overcomes the on-chip storage bottleneck of high-parallel GRAND decoding on FPGA.

[0064] 2. Asymmetric timing "one-to-four" physical multiplexing evolution architecture: Front-end data distribution adopts high parallelism ( Within a single parallelism, a pure physical connection reverse mapping mechanism is used for 4-step serial evolution, which does not require the participation of complex arithmetic logic units (ALUs) throughout the process, achieving a 75% saving of computing resources with minimal clock cycle overhead.

[0065] 3. Globally Parallel CRC Early Stop Closed-Loop No-Blind-Zone Scheduling Network: A distributed CRC correction sub-array strictly matched with the asymmetric generation end was designed. This array synchronously processes massive candidate patterns in groups, terminating memory addressing and state machine evolution immediately upon hit, significantly eliminating control redundancy and power consumption waste caused by deep pipelines. Attached Figure Description

[0066] Figure 1 It is the black channel model of the IEC 61784-3 standard provided by the background technology;

[0067] Figure 2 This is a schematic diagram of a single 48-bit data combination provided in Embodiment 1;

[0068] Figure 3 This is a schematic diagram of the hardware architecture for generating noise error modes using a high-parallelism BRAM array and a state machine to control one to four states, as provided in Embodiment 1.

[0069] Figure 4 This is a schematic diagram of the distributed parallel comparator calculation and verification array design provided in this embodiment. Detailed Implementation

[0070] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0071] The purpose of this invention is to address the shortcomings of existing technologies by providing a highly parallel GRAND-CRC decoding method and system based on pruning and physical multiplexing.

[0072] Example 1

[0073] This embodiment provides a highly parallel GRAND-CRC decoding system based on pruning and physical multiplexing, including:

[0074] The storage module is used to store multiple parameterized data after offline pruning. The parameterized data includes a first parameter set and a second parameter set for reconstructing the noise error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters.

[0075] The global control module, connected to the storage module, is used to read D parameterized data from the storage module in parallel and distribute them to D decoding channels;

[0076] Each decoding channel includes:

[0077] The data parsing module is used to extract morphological parameters, the first parameter set, and the second parameter set from parameterized data.

[0078] The flipping module, connected to the data parsing module, is used to spatially flip the first parameter set and the second parameter set according to the morphological parameters and output multiple sets of reconstructed permutations.

[0079] The mapping module, connected to the flipping module, is used to map each group of reconstructed permutations to a noise mask of a corresponding preset bit width.

[0080] The verification module, connected to the mapping module, is used to perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and to calculate the corrector in the guessed sequence to perform CRC verification.

[0081] The output module is connected to the verification module of each decoding channel and is used to trigger a decoding completion signal when any verification module passes the verification.

[0082] In this embodiment, noise error patterns are stored in the storage module in a compressed parameterized form. During decoding, high-throughput decoding is achieved with minimal storage overhead through parallel reading and physical flipping multiplexing.

[0083] The storage module is used to store multiple parameterized data that have undergone offline pruning.

[0084] Each parameterized data set contains a first set of parameters and a second set of parameters for reconstructing noisy error patterns.

[0085] Suppose the total length of the current Security Protocol Data Unit (SPDU) is N bits, and the total number of error bits (i.e., the total number of "1"s) is L. m The total number of "0"s is L n And L n =NL m .

[0086] The first parameter set is the distribution of consecutive "1" strings in the channel noise error mode. Let m be the number of blocks in the consecutive "1" strings, then the first parameter set A i It can be represented as:

[0087] A i =(a i,1 ,a i,2 ,…,a i,m )

[0088] Among them, a i,r Let r represent the length of the r-th consecutive string of "1"s, where r∈{1, 2, …, m}.

[0089] In this embodiment, in A i In strong composition mode, for m blocks, perform strong composition, and require A to be strong. i any child element a in i,r ≥1.

[0090] The second parameter set represents the distribution of consecutive "0" strings in the channel noise error mode. Let n be the number of blocks in the consecutive "0" strings, then the second parameter set B... j It can be represented as:

[0091] B j =(b j,1 ,b j,2 ,…,b j,n )

[0092] Where b j,t Let represent the length of the t-th consecutive string of "0", where t∈{1, 2, …, n}.

[0093] In this embodiment, in B j In weak combination mode, for n blocks, due to the first and last shapes and the allocation of m blocks, the length of some adjacent child elements can be 0. Therefore, weak combination is performed, and B is required. j any child element b in j,t ≥0.

[0094] Morphological parameters To determine the relationship between the number of length values ​​n in the second parameter set and the number of length values ​​m in the first parameter set, in this embodiment, the morphological parameters... It includes the following three forms:

[0095] when When =1, n=m+1, indicating that the first and last bits of the noise error mode are both "1";

[0096] when When n=2, n=m, indicating that in the noise error mode, one of the first and last bits is "1" and the other is "0".

[0097] when When n=3, n=m-1, indicating that the first and last bits of the noise error mode are both "0".

[0098] It should be noted that the numbers 1, 2, and 3 are mapping relationships set for the convenience of hardware implementation. This embodiment is just an example and is not limited. As long as the three forms can be distinguished, it is acceptable.

[0099] In order to support the efficient operation of the underlying hardware, this embodiment does not calculate the permutation A of the string "1" in real time in the hardware. i Permutation B of the string "0" j Instead, they are generated in advance by an external computer program (such as a preprocessing script in Matlab, Python, or C language), and the generation of both depends on a set of pruning algorithms run by the external computer program.

[0100] To address the issue of a large number of existing noise error modes, this embodiment performs preprocessing (i.e., offline processing) on ​​the first parameter set A before hardware decoding begins. i Second parameter set B j Pruning is performed on all possible permutations.

[0101] A pruning algorithm is an optimization strategy that reduces computation and storage consumption by identifying and removing branches that are unlikely to produce valid or duplicate results during recursive search or combination processes.

[0102] For either the first or second parameter set, let S be the set of valid permutations without pruning constraints, and let X be the initial total number of permutations. Then |S| = X. The permutation set S is generated by obtaining all permutations of length m or n whose sum of elements equals L (L is the sum of the elements of the corresponding parameter set).

[0103] For any permutation A = (a1, a2, ..., a3) in the permutation set S,z ), where z represents the number of length values ​​in permutation A (i.e., the number of length values ​​contained in this parameter set, corresponding to m in the first parameter set or n in the second parameter set), a1 represents the length of the first element of permutation A (i.e., the length of the first consecutive string of "1"s or consecutive strings of "0"s), a z This represents the length of the last element in permutation A (i.e., the length of the last consecutive string of "1"s or "0"s), a1—a z All are positive integers (first parameter set) or non-negative integers (second parameter set).

[0104] Define a spatial reversal mapping operator R: S → S, for any permutation A = (a1, a2, ..., a...). z The space reversal mapping operator performs a head-to-tail position flipping mapping on A, completely reversing the head and tail of A, as shown below:

[0105]

[0106] Clearly, the spatial reversal mapping operator satisfies involution, meaning that two consecutive reversals equal the permutation of itself, expressed as:

[0107]

[0108] Based on the spatial reversal mapping operator R, the universal set S can be strictly divided into a set of palindromic permutations S. P With asymmetric permutation set S A ,at this time:

[0109]

[0110]

[0111] Among them, the palindrome permutation set S P Let R(A) represent the set of permutations whose sequences remain unchanged after being flipped, i.e., permutations satisfying R(A) = A. The characteristic of a palindrome permutation is: a1 = a... z a2=a z-1 ...; Let the number of permutations it contains be P, then:

[0112]

[0113] It should be noted that, in the context of pruning processing in this embodiment, the pruning determination only depends on the first element a1 and the last element a1. z By comparison, the determination of a palindrome can be simplified to satisfying a1=a z The arrangement.

[0114] Asymmetric permutation set S A Let the set of sequences that change after being flipped be represented as:

[0115]

[0116] The pruning algorithm uses two pointers to verify from both ends towards the middle. The specific rules are as follows:

[0117] First pruning rule: If the first element a1 of permutation A is less than the last element a z (i.e., a1) z If the permutation A is truncated, this is because the underlying hardware supports reverse evolution, and the reverse permutation R(A) (whose first element is a) is truncated. z The last element is a1, satisfying a z > a1) will be generated by the flip module during hardware execution, so the current permutation does not need to be stored.

[0118] Second pruning rule: If the first element a1 of permutation A is equal to the last element a z (i.e., a1 = a) z If the permutation A is not generated, then permutation A is retained and no subpermutations are generated. In this case, permutation A is a palindrome (i.e., the permutation is the same as its reverse). All permutations of this type are retained in the pruning process. The reason for blocking the generation of subpermutations is to avoid repeated calculations and reduce the subsequent processing overhead.

[0119] Third pruning rule: If the first element a1 of permutation A is greater than the last element a z (i.e., a1 > a) z If the permutation A is a, then the permutation A is retained because in R(A), which forms an inverse pair with A, the first element of R(A) is a. z The last element is a1, satisfying a z Since R(A) satisfies the first pruning rule, it is pruned, and A is retained as a representative of the inverse pair.

[0120] It should be noted that the above pruning process is performed independently for the first parameter set and the second parameter set, respectively.

[0121] According to the above rules, the asymmetric set S A Further divided into the retention set S keep With the discard set S prune , is represented as:

[0122]

[0123]

[0124]

[0125] Since the head-to-tail relationship of any asymmetric permutation A and its flip map R(A) is necessarily strictly inverse, this constitutes S keep With S prune ​A non-intersecting bijection between them (i.e., any retained term necessarily corresponds uniquely to a discarded term):

[0126]

[0127] Therefore, the cardinality of the two sets is absolutely equal:

[0128]

[0129] Combined with S A From the partition equation, we can derive the equation for the number of retainers:

[0130]

[0131]

[0132] The final closed-form solution generating the number and reduction amount is obtained by applying the second pruning rule of the two-pointer pruning algorithm, resulting in a palindrome set S with the head and tail being completely equal. P All will be retained (the current sequence is preserved, but the generation of sub-permutations is blocked), so the effective seed set S that is ultimately stored in the BRAM memory will be... final for:

[0133]

[0134] The actual number of valid permutations N generated pruned Represented as:

[0135]

[0136]

[0137] It should be noted that the total number of valid permutations generated above is performed independently for the first parameter set and the second parameter set, respectively.

[0138] After pruning, the number of valid "seed" arrangements N is extracted. pruned It equals half of the non-palindrome combination plus all palindrome combinations, that is... The generated A i and B j will and m and Data is pre-stored in the FPGA's BRAM, combining and merging data according to bit width, into 48-bit data entries. A single 48-bit data entry is combined as follows: Figure 2 As shown.

[0139] In the global control module, it is connected to the storage module and is used to read D parameterized data from the storage module in parallel and distribute them to D decoding channels.

[0140] Under the clock constraints of the global control, the system outputs D-channel 48-bit data (pattern seeds) in parallel through a single BRAM read operation, which are then sent to the subsequent D noise error mode generation paths. Each decoding channel independently completes the conversion from parameterized data to noise mask and CRC check, thereby achieving D-channel parallel decoding processing and significantly improving the system throughput.

[0141] Each decoding channel consists of a data parsing module, a flipping module, a mapping module, a verification module, and an output module, which in turn generate noise error modes, such as... Figure 3 As shown.

[0142] The data parsing module is used to extract morphological parameters from parameterized data. First parameter set A i The second parameter set B j .

[0143] After 48 bits of data are input, the valid data is extracted from the 48 bits, and then the 'm' in the data is extracted. A i =(a i,1 ,a i,2 ,…,a i,m ), B j =(b j,1 ,b j,2 ,…,b j,n ).

[0144] The flip module, connected to the data parsing module, is used to, within a continuous first to fourth clock cycle, determine the data based on the morphological parameters. For the first parameter set A i Second parameter set B j Perform four types of spatial flips to output multiple sets of reconstructed permutations.

[0145] During the hardware decoding execution phase, after a single hardware lane reads a string of 48-bit data from the BRAM in a single operation, it will rely on the TDM state machine to construct four specific noise error modes within four consecutive clock cycles.

[0146] The flip module (one-to-four evolution) consists of two parts: the TDM state machine and the cross matrix.

[0147] The TDM state machine is a time-division multiplexing state machine. It sequentially issues four corresponding gating control signals during the first to fourth clock cycles to control the connection state of the cross matrix.

[0148] The crossbar matrix is ​​a purely hardware-connected cross-connect matrix. Based on the gating control signals issued by the TDM state machine, the crossbar matrix connects the input signal lines to the output in a predetermined manner, thereby controlling the first parameter set A. i Second parameter set B j Perform the head-to-tail swap operation.

[0149] The operations performed in each of the four clock cycles are as follows:

[0150] First clock cycle: The TDM state machine issues the first strobe control signal, controlling the cross matrix to make the first parameter set A i Preserve the original order, second parameter set B j Maintaining the original order, the reconstructed arrangement of the output (A) i B j )for:

[0151] A i =(a i,1 ,a i,2 ,…,a i,m ), B j =(b j,1 ,b j,2 ,…,b j,n );

[0152] Second clock cycle: The TDM state machine issues the second strobe control signal, controlling the cross matrix to make the first parameter set A i First and last order reversed (only m is flipped), second parameter set B j Maintaining the original order, the output is reconstructed and rearranged ( B j )for:

[0153] =(a i,m ,a i,m−1 ,…,a i,1 ), B j =(b j,1 ,b j,2 ,…,b j,n );

[0154] Third clock cycle: The TDM state machine issues the third strobe control signal, controlling the cross matrix to make the first parameter set A i Preserve the original order, second parameter set B j Reverse order (only n is flipped), output the reconstructed permutation (A) i , )for:

[0155] A i =(a i,1 ,a i,2 ,…,ai,m ), =(b j,n ,b j,n−1 ,…,b j,1 );

[0156] Fourth clock cycle: The TDM state machine issues the fourth strobe control signal, controlling the cross matrix to make the first parameter set A i First and last order reversed (m flipped), second parameter set B j Reverse order (n flips), output reconstructed permutation ( , )for:

[0157] =(a i,m ,a i,m−1 ,…,a i,1 ), =(b j,n ,b j,n−1 ,…,b j,1 ).

[0158] In the fourth clock cycle, A i and B j The arrangement order was all reversed, and at the same time, based on the morphological parameters Adjust the polarity of the endpoint signal, that is, correctly set the first and last bits of the noise error mode to "1" or "0".

[0159] The first and last order reversal operations in the above four reconstruction arrangements are implemented entirely by the physical connection of the cross matrix at the hardware level. No iterative calculation or arithmetic logic unit is required. Therefore, the operation of the flip module within four clock cycles does not consume additional arithmetic operation resources. The four forms of output can be completed simply by switching the strobe signal.

[0160] The mapping module, connected to the flipping module, is used to map each group of reconstructed permutations to a noise mask of a corresponding preset bit width.

[0161] In this embodiment, the preset bit width is 64 bits, and the mapping module specifically includes an accumulation unit, a mask generation unit, a separation unit, and a fusion unit.

[0162] The accumulation unit is used to perform accumulation and summation on each length value in the reconstructed arrangement to obtain the absolute starting index of each length value in the preset bit-width space.

[0163] After receiving the data for the current clock cycle, each child element 'a' is computed using a parallel addition tree within each clock cycle. i,r and b j,t Absolute starting index in the 64-bit physical channel space.

[0164] For example, now the bit width N=64, and the shape parameters =1 (representing a string with both ends being "0", i.e., n=m+1), at this time m=2, n=3. Let the extracted first parameter set A be... i =(2,3) (i.e., a1=2, a2=3), the second parameter set B j =(3,4,52) (i.e., b1=3, b2=4, b3=52), the sum of the two sets and the elements is 2+3+3+4+52=64.

[0165] In a parallel addition tree, the specific process of calculating the absolute starting index of the string "1" over four clock cycles is as follows:

[0166] First clock cycle (positive): The cross matrix is ​​not flipped, and the reconstructed permutation is A. i =(2,3), B j =(3,4,52), the splicing order is b1, a1, b2, a2, b3.

[0167] The absolute starting index of the first string of "1"s (a1) is b1 = 3;

[0168] The absolute starting index of the second string of "1"s (a2) = b1 + a1 + b2 = 3 + 2 + 4 = 9;

[0169] Second clock cycle (only m flips): The fork matrix only makes A... i Reverse order, reconstructed arrangement as =(3,2), B j =(3,4,52), the splicing order is b1, a2, b2, a1, b3.

[0170] The absolute starting index of the first string of "1"s (a2 in this period) is b1 = 3;

[0171] The absolute starting index of the second string of "1"s (a1 in this period) is b1+a2+b2=3+3+4=10.

[0172] Third clock cycle (only n flips): The cross matrix only makes B... j Reverse order, reconstruct the arrangement to A i =(2, 3), =(52,4,3), the splicing order is b3, a1, b2, a2, b1.

[0173] The absolute starting index of the first string of "1"s (a1 in this period) is b3 = 52;

[0174] The absolute starting index of the second string of "1"s (a2 in this period) is b3+a1+b2=52+2+4=58.

[0175] Fourth clock cycle (full flip): The cross matrix makes A i With B j All in reverse order, reconstructed as follows =(3,2), =(52,4,3), the splicing order is b3, a2, b2, a1, b1.

[0176] The absolute starting index of the first string of "1"s (a2 in this period) is b3 = 52;

[0177] The absolute starting index of the second string of "1"s (a1 in this period) is b3+a2+b2=52+3+4=59.

[0178] The mask generation unit is used to generate the corresponding mask value based on each length value.

[0179] For a length of a i,r The mask value of a consecutive string of "1"s is (1 << a i,r )-1, where << represents the left shift operator.

[0180] The separation unit is used to perform bit-level shifting and byte alignment on the mask value according to the absolute start index.

[0181] In this embodiment, the separation unit first obtains the absolute start index calculated by the accumulation unit, performs bit-level shift on the mask value according to the low bits (e.g., the low 3 bits) of the absolute start index, and performs byte alignment based on the multiplexer according to the high bits (high 3 bits) of the absolute start index.

[0182] Taking a preset bit width of 64-bit (i.e., divided into 8 bytes, denoted as Byte7 to Byte0 from high to low) as an example, the absolute starting index is represented by a 6-bit binary number (within the range of 0-63). Assume that the initial mask value output by the mask generation unit is 64 bits of 000...00000111 (i.e., a continuous string of "1"s of length 3, all distributed in the lowest position Byte0), and the accumulation unit calculates that the absolute starting index of the mask is 27 (its 6-bit binary representation is 011011).

[0183] The specific execution process of the separation unit is as follows:

[0184] First stage (bit-level shift): The separation unit extracts the lower 3 bits 011 (i.e., decimal 3) of the absolute start index, controls the underlying shift logic to shift the initial mask value left by 3 bits, and obtains the intermediate mask value 00...0111000.

[0185] The second level (byte-level routing): The separation unit extracts the high 3 bits of the absolute starting index, 011 (i.e., the decimal 3, representing that the target position needs to cross 3 bytes upwards). This step no longer uses bit-by-bit shifting, but instead uses "byte (8-bit)" as the granularity. Through an 8-to-1 multiplexer (MUX) network, the Byte0 output from the first level, containing valid data, is directly routed (transferred) to the position of Byte3 (all other unrouted bytes are padded with zeros). After byte-level routing, the 64-bit data instantly becomes: [Byte7 to Byte4 all zeros] - [Byte3 = 00111000] - [Byte2 to Byte0 all zeros].

[0186] After the above two-stage separation operation, the mask value is effectively shifted left by a total of 27 bits (3-bit bit shift + 3×8-bit byte routing). This two-stage architecture of "low 3-bit local shift combined with high 3-bit MUX byte overall routing" cleverly avoids instantiating a large and high-latency 64-bit barrel shifter in the underlying hardware, and significantly reduces combinational logic latency and LUT resource consumption while achieving accurate position mapping.

[0187] The fusion unit is used to perform a bitwise OR operation on the shifted and byte-aligned mask values ​​to generate a noise mask with a preset bit width.

[0188] The mask is correctly placed into the corresponding 64-bit position and then combined by bitwise OR operation. Finally, the output latch provides the parallel noise error mode noise_out and the valid indicator valid_out.

[0189] The above steps correctly achieve A within a single clock cycle. i and B j The real-time reverse ordering reduces resource consumption during parallel generation.

[0190] In this embodiment, 48-bit is the bit width of the parameterized data stored in BRAM, i.e. the bit width of the compressed storage format. After the 48-bit data is expanded by the flip module and mapping module during the hardware decoding execution period, a 64-bit physical noise mask is generated. 64-bit is the total length of SPDU N=64 bits. Therefore, the storage of 64-bit noise error modes is changed to storing their arrangement and performing pruning during the generation stage, which reduces the on-chip storage capacity.

[0191] If, under the traditional mechanism, these 857,876 64-bit physical noise masks are directly stored in on-chip memory, the required storage capacity is:

[0192]

[0193] The storage not only reduces the bit width to 48 bits, but more importantly, it achieves a significant reduction in depth through a one-to-four multiplexing logic. Let's assume a certain parameter group... The total number of combinations is The number of palindromic permutations is . The number of weak combinations is The number of palindromic permutations is Data stored in BRAM needs to be generated by pairing the pruned sets using two pointers, and the effective depth actually occupied by a single set of parameters is [not specified]. for:

[0194]

[0195] Expanding, we get:

[0196]

[0197] Because palindromic permutations form indivisible symmetrical residuals, and for highly symmetrical forms (such as m=4, L...), m =4"), the one-to-four reuse actually degenerates into one-to-two, so the total number of data actually stored must be greater than the simple 857876 / 4 (i.e. 214,469).

[0198] Adding up the 26 regular combinations listed in Table 1 and the 5 special boundary state combinations listed in Table 2, including the supplementary all-1 sequences (such as the boundary state with n=0), a total of 31 seed combinations are accumulated. The total number of 48-bit combinations that the system actually needs to store is 406,551.

[0199] Table 2. Special Boundary State Combination Parameters for Channel Noise Error Mode

[0200]

[0201] After adopting the 48-bit parameterized storage in this embodiment, the actual required storage capacity is:

[0202]

[0203] In contrast, if all 857,876 64-bit types are compared using the traditional method through G and G1 values, it can be seen that: under the premise of ensuring complete coverage of 857,876 real physical noise modes, this embodiment uses a pruning algorithm and state machine to achieve a one-to-four mechanism, reducing the required storage capacity from 52.36 Mbits to 18.61 Mbits, a reduction of 64.5%, in order to increase the upper limit of high-parallelism GRAND-CRC decoding stored on the FPGA.

[0204] In the verification module, connected to the mapping module, a bitwise XOR operation is performed between the noise mask and the data to be decoded to obtain a guessed sequence. A corrector is then calculated from the guessed sequence to perform CRC verification. The specific calculation and implementation process is as follows:

[0205] Let Y be the data to be decoded (i.e. the received sequence affected by channel interference) obtained by the receiver from the black channel, and E be the noise mask with a preset bit width (i.e. the guessed channel error mode) generated by the mapping module.

[0206] First, an XOR denoising operation is performed: the verification module performs a bitwise XOR operation on Y and E through a parallel XOR logic gate network in the hardware to obtain the denoised guess sequence (i.e., candidate codeword). The XOR calculation formula is as follows:

[0207]

[0208] Subsequently, the specific implementation of CRC verification involves calculating the synthesizer. Let H be the check matrix of the CRC check space used by the Security Protocol Data Unit (SPDU), and the obtained guessed sequence... The input is fed into the independent CRC check tree at the back end, and the corrector S is calculated using the matrix operation formula:

[0209]

[0210] in, This represents the transpose of the parity check matrix H.

[0211] The verification module internally includes all-zero detection logic (such as a NOR gate with a preset bit width). If the corrector S=0 (i.e., an all-zero vector) is detected, the guessed sequence is proven. This is a set of valid codewords under the current channel coding; at this point, the CRC check passes. This is the recovered original correct data; if S≠0, it indicates that the currently generated noise mask E is incorrect and the CRC check fails, and the system will continue to test the subsequently generated noise mask.

[0212] like Figure 4 As shown, the candidate noise error modes generated in parallel by D channels within any clock cycle are immediately XORed with the data to be decoded (SPDU) for noise reduction and then sent to the D independent CRC check trees in the back end.

[0213] In the output module, it is connected to the verification module of each decoding channel to collect the verification results of each channel and perform global early stop determination. The early stop mechanism can effectively avoid invalid decoding search and reduce system power consumption and latency.

[0214] When the verification module of any decoding channel passes the verification, the output module triggers a decoding completion signal, indicating that the decoding has been successful.

[0215] When all the verification modules of the decoding channels fail to pass the verification (i.e., there is still no successful result after completing a set of tests for four clock cycles), the output module sends a continue reading instruction to the global control module, triggering the next round of parallel reading of D parameterized data from the BRAM.

[0216] This embodiment has the following beneficial effects:

[0217] 1. Parameter mechanism based on BRAM interface: Completely abandons the storage of full-length noise error modes, and instead extracts m and A. i B j and combined signals It is pre-stored in BRAM. During decoding, it is read concurrently on demand through the interface, which greatly overcomes the on-chip storage bottleneck of high-parallel GRAND decoding on FPGA.

[0218] 2. Asymmetric timing "one-to-four" physical multiplexing evolution architecture: Front-end data distribution adopts high parallelism ( Within a single parallelism, a pure physical connection reverse mapping mechanism is used for 4-step serial evolution, which does not require the participation of complex arithmetic logic units (ALUs) throughout the process, achieving a 75% saving of computing resources with minimal clock cycle overhead.

[0219] 3. Globally Parallel CRC Early Stop Closed-Loop No-Blind-Zone Scheduling Network: A distributed CRC correction sub-array strictly matched with the asymmetric generation end was designed. This array synchronously processes massive candidate patterns in groups, terminating memory addressing and state machine evolution immediately upon hit, significantly eliminating control redundancy and power consumption waste caused by deep pipelines.

[0220] Correspondingly, a highly parallel GRAND-CRC decoding method based on pruning and physical multiplexing is also provided, including:

[0221] S1. Parallel reading of D parameterized data from the storage module and distribution to D decoding channels; wherein the storage module stores multiple parameterized data after offline pruning, the parameterized data includes a first parameter set and a second parameter set for reconstructing the noise error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters;

[0222] S2. In each decoding channel, extract the morphological parameters, the first parameter set, and the second parameter set from the parameterized data;

[0223] S3. Perform spatial flipping on the first and second parameter sets according to the morphological parameters, and output multiple sets of reconstructed permutations;

[0224] S4. Perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and calculate the corrector in the guessed sequence to perform CRC check;

[0225] S5. When the CRC check of any decoding channel passes, trigger the decoding completion signal.

[0226] Example 2

[0227] The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing provided in this embodiment differs from that in Embodiment 1 in the following ways:

[0228] To verify the effectiveness of this system architecture, a physical implementation was performed on a Xilinx Zynq UltraScale+XCZU15EG-ffvb1156-2-i FPGA platform, using an SPDU length of N=64 bits and a CRC32 encoding scheme as an example. The system pre-stores 406,551 types of 48-bit data, and performs synthesis and placement / routing using three parallelism schemes D∈{128, 256, 512}. The system's main clock frequency is 96MHz, and the PS communicates with Matlab via UART (115200 baud).

[0229] Table 3 shows the resource usage of the three parallelism schemes on the FPGA after placement and routing.

[0230] Table 3 Comparison of FPGA resource consumption under different parallelism levels (XCZU15EG)

[0231]

[0232] As shown in Table 3: (1) BRAM consumption is negatively correlated with parallelism. The lower the parallelism, the greater the BRAM depth carried by each lane (3,177 lines / lane for 128 lanes, and only 795 lines / lane for 512 lanes), but the total bit width of BRAM is reduced. The overall result is that the BRAM consumption is the highest for 128 lanes (573 tiles) and the lowest for 512 lanes (504.5 tiles); (2) LUT consumption increases approximately linearly with parallelism, from 128 lanes to 256 lanes by about 1.95 times, and from 256 lanes to 512 lanes by about 2.10 times, mainly due to the multiple replication of the CRC check array and noise generation logic.

[0233] The power consumption of different parallelism schemes after placement and routing is shown in Table 4. The power consumption analysis was performed at a main clock frequency of 96MHz.

[0234] Table 4. Power consumption comparison under different parallelism levels (96MHz, XCZU15EG)

[0235]

[0236] As shown in Table 4, dynamic power consumption exhibits an approximately linear relationship with the increase in parallelism (approximately 2.18 times the increase from 128 channels to 512 channels), which is consistent with the linear growth trend of LUT resources. Static power consumption varies little between different parallelism levels (0.830-0.890 W), mainly determined by the inherent leakage current of the device, and is only slightly affected by the scale of logic resources.

[0237] The static timing analysis (STA) results of the three parallelism schemes after placement and routing are shown in Table 5. The target clock constraint is 96 MHz (period 10.417 ns).

[0238] Table 5. Timing convergence results under different parallelism levels (target 96MHz)

[0239]

[0240] All three schemes achieved zero-violation timing convergence. The 128-channel and 256-channel schemes had ample timing margin (WNS > 1.7 ns), while the 512-channel scheme reduced its WNS to 0.375 ns, but remained within the constraints and had no violations at any endpoint. The highest operable frequency decreased with increasing parallelism, mainly because the combinational logic fan-out of the CRC check array increased under high parallelism, leading to increased interconnect latency.

[0241] The main processing flow of the system is as follows: each decoding round (batch) contains the evolution and verification of 4 test_forms. The precise cycle consumption of each stage within a single batch (based on cycle-by-cycle analysis of the state machine) is shown in Table 6.

[0242] Table 6. Cycle Consumption of Each State within a Single Decoding Round

[0243]

[0244] A single decoding round without a match consumes 35 clock cycles. Upon a match, an additional 5-8 clock cycles are added for the hit mode fetch state (1 cycle), data output lock state (3-6 cycles), and decoding completion state (1 cycle).

[0245] Table 7 shows the decoding delays of the three parallelism schemes in the worst case (traversing all BRAM addresses and finally matching), with the clock frequency calculated at 96MHz.

[0246] Table 7. Worst-case decoding delay (96 MHz) under different parallelism levels

[0247]

[0248] Worst case decoding delay T worst The calculation formula is as follows:

[0249]

[0250] Where, N batches This represents the total number of batches. 35 represents the number of periods without a match in a single batch, and 43 represents the total number of periods with a match in a batch (35+8).

[0251] The performance and resource trade-offs of the three parallelism schemes are shown in Table 8.

[0252] Table 8. Comprehensive Comparison of Three Parallelism Schemes

[0253]

[0254] In practical engineering applications, it is recommended to choose the appropriate solution based on the specific scenario: for power consumption and resource-sensitive scenarios, a 128-channel solution is recommended (lowest power consumption, LUT usage is only 20.3%); for balanced scenarios, a 256-channel solution is recommended (0.58 ms latency, moderate resource usage); for ultra-low latency scenarios, a 512-channel solution should be selected, which can achieve a worst-case decoding latency of 0.29 ms.

[0255] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A highly parallel GRAND-CRC decoding system based on pruning and physical multiplexing, characterized in that, include: The storage module is used to store multiple parameterized data after offline pruning. The parameterized data includes a first parameter set and a second parameter set for reconstructing the noise error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters. The global control module, connected to the storage module, is used to read D parameterized data from the storage module in parallel and distribute them to D decoding channels; Each decoding channel includes: The data parsing module is used to extract morphological parameters, the first parameter set, and the second parameter set from parameterized data. The flipping module, connected to the data parsing module, is used to spatially flip the first parameter set and the second parameter set according to the morphological parameters and output multiple sets of reconstructed permutations. The mapping module, connected to the flipping module, is used to map each group of reconstructed permutations to a noise mask of a corresponding preset bit width. The verification module, connected to the mapping module, is used to perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and to calculate the corrector in the guessed sequence to perform CRC verification. The output module is connected to the verification module of each decoding channel and is used to trigger a decoding completion signal when any verification module passes the verification.

2. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 1, characterized in that, The flipping module performs several spatial flips on the first parameter set and the second parameter set according to the morphological parameters, specifically: Four types of space flips are performed in successive clock cycles from the first to the fourth, and four sets of reconstructed permutations are output in sequence.

3. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 2, characterized in that, The flip module also includes a TDM state machine and a cross matrix; The TDM state machine is used to issue corresponding gating control signals during the first to fourth clock cycles. The cross matrix is ​​used to perform head-to-tail swapping on the first and second parameter sets according to the gating control signal.

4. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 3, characterized in that, The four sets of output reconstruction arrangements include: During the first clock cycle, the TDM state machine controls the cross matrix to keep the first parameter set in its original order and the second parameter set in its original order. During the second clock cycle, the TDM state machine controls the cross matrix to reverse the first parameter set from beginning to end while keeping the second parameter set in its original order. During the third clock cycle, the TDM state machine controls the cross matrix to keep the first parameter set in its original order and reverse the order of the second parameter set. During the fourth clock cycle, the TDM state machine controls the cross matrix to reverse the order of the first parameter set and the second parameter set.

5. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 1, characterized in that, The offline pruning process in the storage module specifically involves: Based on either the first or second parameter set, construct a permutation set S. For any permutation A = (a1, a2, ..., a...) in the permutation set S... z ): If the first element a1 of permutation A is less than the last element a z If so, then permutation A is removed; If the first element a1 of permutation A is equal to the last element a z If the permutation A is not generated, then permutation A is retained and no subpermutations are generated. If the first element a1 of permutation A is greater than the last element a1... z If so, then permutation A is retained.

6. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 5, characterized in that, The total number of valid permutations of the first or second parameter set retained after the offline pruning process is: ; in, X represents the total number of valid permutations of the first or second parameter set; P represents the total number of full permutations of the first or second parameter set; and P represents the number of palindromic permutations of the first or second parameter set.

7. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 1, characterized in that, The mapping module includes: The accumulation unit is used to perform accumulation and summation on each length value in the reconstructed arrangement to obtain the absolute starting index of each length value in the preset bit-width space; The mask generation unit is used to generate corresponding mask values ​​based on each length value; The separation unit is used to perform bit-level shifting and byte alignment on the mask value according to the absolute start index; The fusion unit is used to perform a bitwise OR operation on each mask value after shifting and byte alignment to generate a noise mask with a preset bit width.

8. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 7, characterized in that, The separation unit performs bit-level shifting and byte alignment on the mask value according to the absolute start index as follows: The mask value is bit-level shifted based on the low-order bits of the absolute start index, and multiplexer-based byte alignment is performed based on the high-order bits of the absolute start index.

9. The high-parallel GRAND-CRC decoding system based on pruning and physical multiplexing according to claim 1, characterized in that, The output module also includes: issuing a continue reading instruction to the global control module when all verification modules fail to pass verification.

10. A decoding method based on the pruning and physical multiplexing high-parallel GRAND-CRC decoding system according to any one of claims 1-9, characterized in that, include: S1. Parallel read D parameterized data from the storage module and distribute them to D decoding channels; The storage module stores multiple parameterized data that have undergone offline pruning. The parameterized data includes a first parameter set and a second parameter set for reconstructing the noisy error mode, and the number of length values ​​contained in the first parameter set and the second parameter set is determined by the morphological parameters. S2. In each decoding channel, extract the morphological parameters, the first parameter set, and the second parameter set from the parameterized data; S3. Perform spatial flipping on the first and second parameter sets according to the morphological parameters, and output multiple sets of reconstructed permutations; S4. Perform a bitwise XOR operation between the noise mask and the data to be decoded to obtain a guessed sequence, and calculate the corrector in the guessed sequence to perform CRC check; S5. When the CRC check of any decoding channel passes, trigger the decoding completion signal.

Citation Information

Patent Citations

  • High-speed parallel GRAND-CRC decoding method for functional safety communication

    CN121356602A