A flexible and efficient Polar code decoding method and decoder
Through the Fast-CASCL algorithm and semi-parallel hardware architecture, the problem of high latency of traditional Polar code decoding algorithms is solved, efficient decoding performance and low resource consumption are achieved, and it is suitable for high-speed communication scenarios.
Patent Information
- Application Number
- CN202410563645.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-08
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-05-08
AI Technical Summary
Traditional Polar code decoding algorithms have high computational complexity, high latency and low throughput, making it difficult to adapt to fast-responsive communication scenarios.
A low-latency fast Polar decoding algorithm Fast-CASCL is designed. By summarizing four types of special decoding structures and fast nodes, combining CRC auxiliary strategies, it optimizes decoding efficiency and reduces latency, and realizes semi-parallel in the hardware architecture. Architecture to improve efficiency.
It significantly reduces the decoding delay, improves the decoding efficiency and performance, is suitable for hardware implementation in short code scenarios, maintains high decoding performance, and is suitable for high-speed communication scenarios.
Smart Images

Figure CN118399979B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of communication technology, and in particular relates to a flexible and efficient Polar code decoding method and decoder, which are used to reduce Polar code decoding delay while maintaining excellent decoding performance. Background Art
[0002] Due to the influence of noise or interference in the channel environment, random errors may occur during information transmission. The mainstream defense method in the current communication field is to use channel coding technology to correct the transmission information at the expense of a certain transmission efficiency. The main indicator for evaluating channel coding technology is whether it has reached the Shannon transmission limit. Turbo codes and LDPC, which have been widely used in 4G and 5G respectively, are only close to the Shannon limit in simulation. The Polar code proposed by Arikan in 2009 theoretically proves that the code type can reach the Shannon limit, and with the support of some decoding algorithms, the performance of Polar code in the short code field is better than Turbo code and LDPC. At the same time, the encoding method and decoding calculation structure of Polar code are simple, which is very suitable for the hardware implementation of low-complexity codecs, and has been applied in the short code control channel of 5G NR.
[0003] The performance of Polar codes under traditional SC decoding algorithms is very poor, and the coding gain is far behind that of LDPC. Subsequent scholars proposed decoding algorithms such as SCL (Successive Cancellation List), SCS, and CASCL for this algorithm, which gradually approached the optimal decoding performance of Polar codes. Based on the optimal CASCL, the performance of Polar codes in the field of short code communication exceeds that of LDPC. Therefore, the comprehensive performance of Polar is very suitable for short frame communication scenarios.
[0004] Although SC decoding algorithms such as SCL, SCS, and CASCL have good performance, their decoding efficiency is very poor, resulting in high actual decoder delay, which is not suitable for some fast-response communication scenarios. Later, some researchers proposed some special decoding structures such as Rate-0, Rate-1, REP, Type-I, Type-II, and Type-III to optimize Polar's SC decoding process and reduce the decoding delay to a certain extent. However, the decoding structures proposed by these algorithms are cumbersome, and some decoding structures do not give update rules for list decoding, resulting in the inability to optimize delay on CASCL with the best performance. In addition, the hardware calculation complexity of the above special decoding is too high, making it difficult to implement in engineering. Summary of the invention
[0005] In view of the high computational complexity, high latency and low throughput of the traditional SC-type Polar decoding algorithm, the present invention designs a low-latency fast Polar decoding algorithm Fast-CASCL (FastNode-based CRC-aidedSuccessive Cancellation List, based on fast node CRC-aided list continuous elimination). This algorithm uses the clustering effect of the decoding bits on the decoding tree to summarize four types of special decoding structures: Rate-0, Rate-1, REP and SPC, and forms four corresponding fast decoding nodes. This Fast-CASCL algorithm not only designs a decoding calculation method and path update strategy suitable for hardware implementation for each node, but also realizes the fast decoding of Polar codes, which is particularly suitable for the implementation of hardware decoders in short code scenarios, optimizes decoding efficiency and improves performance.
[0006] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:
[0007] S1. Based on the decoding tree structure of the SC decoding algorithm, a fast decoding structure is summarized using this structure. This structure facilitates the improvement of the Fast-CASCL algorithm from the CASCL (CRC-Assisted Successive Cancellation List) algorithm with better decoding performance.
[0008] S2. Summarize and design four special fast decoding nodes, namely
[0009] (1) Rate-0 node: Processes a subtree with all leaf nodes as frozen bits. The decoding rule is simplified to directly output 0, which greatly reduces the computational complexity.
[0010] (2) Rate-1 node: processes a subtree whose leaf nodes are all information bits, and improves efficiency by simplifying the calculation instead of the complex decoding process;
[0011] (3) REP node: processes a subtree containing only one information bit and simplifies the decoding process through addition and threshold judgment;
[0012] (4) SPC node: a subtree that processes a single parity check code. This subtree contains only one frozen bit and uses low-complexity maximum likelihood (ML) decoding.
[0013] S3. Design a list decoding update rule for each fast decoding node to obtain the final Fast-CASCL algorithm
[0014] (1) For the four fast decoding nodes described in S2, the decoding calculation methods and path metric update rules are designed respectively, and the PM update algorithm based on hardware optimization is adopted to reduce the decoding delay and maintain excellent decoding performance;
[0015] (2) The CRC-assisted SCL strategy is introduced for path selection. The CASCL algorithm generates CRC before encoding at the transmitting end and encodes the check code and information bits simultaneously. The receiving end uses CRC check combined with PM to select the optimal path, which greatly enhances the decoding performance and ensures high decoding efficiency and accuracy.
[0016] S4. Hardware architecture implementation based on Fast-CASCL algorithm
[0017] (1) Semi-parallel architecture: The present invention adopts a semi-parallel architecture based on SC operation, and implements semi-parallel operation through a reasonable number of processing units (PEs), thereby achieving a balance between resources and efficiency. This architecture also limits the size of special nodes on the decoding tree by defining the degree of parallelism, ensuring the flexibility and efficiency of hardware implementation.
[0018] (2) Hardware design: The decoder hardware design includes storage units and logic calculation units, including unit factor calculation core, channel layer LLR memory, intermediate layer LLR memory, feedback bit memory, path metric memory, address mapping memory, CRC unit and path memory, etc. These units cooperate to perform LLR calculation, hard bit value back propagation, path splitting, special processing of four fast nodes and CRC-assisted path selection.
[0019] (3) Decoding control: The decoding controller is responsible for controlling the entire decoding process, including the forward calculation of LLRs, the reverse propagation of the hard bit value B, and managing operations such as special node processing, path updating and screening. It guides the decoding process by storing the topological structure of the decoding tree after pruning (special node index).
[0020] S5. Further theoretical delay analysis of Fast-CASCL algorithm
[0021] (1) Fully parallel decoding assumption: In an ideal situation, assuming that the hardware can fully parallelize the calculations in the SC decoding structure, the decoding delay for Polar with a specific code length can be greatly reduced.
[0022] (2) Delay influencing factors: The decoding delay of the Fast-CASCL algorithm is mainly affected by the number of special nodes. Especially under low bit rate conditions, the proportion of certain low-latency type nodes increases, which further reduces the decoding delay.
[0023] Practical hardware parallelism limitations: Although the parallelism is limited in the actual hardware implementation, the Fast-CASCL algorithm can still significantly reduce latency through effective management and processing of special nodes, especially in various short code scenarios, achieving significant latency reduction compared to CASCL decoding.
[0024] S6. Build a baseband software simulation and test environment to verify the rationality of Fast-CASCL and test the performance indicators of the Fast-CASCL hardware decoder.
[0025] (1) By building a simulation environment on the Matlab platform, we simulated the performance of several common short and medium-sized Polar codes. By comparing the results, we confirmed that Fast-CASCL has significantly improved the performance compared with the SC and SCL algorithms while maintaining the performance level of CASCL.
[0026] (2) The Fast-CASCL hardware decoder was deployed on Xilinx's ZYNQ SoC XCZU9EG using the Vivado tool, and the hardware board verification was completed by building a prototype verification system. The system can simulate the complete communication process, including from signal source generation to BPSK modulation and demodulation, to Polar encoding and decoding, and finally performance testing.
[0027] The test results show that the Fast-CASCL hardware decoder works effectively at high clock frequencies, demonstrating the advantages of low resource consumption and high throughput, fully verifying its application potential in actual hardware implementation.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] Compared with traditional methods, this invention proposes an efficient Polar code decoding method based on Fast-CASCL in the field of communication technology, which improves the decoding efficiency and performance of Polar codes, and also meets the actual needs in high-speed communication scenarios through algorithm optimization, successfully overcoming many challenges faced by existing technologies. This innovative solution has the following significant advantages:
[0030] 1. The Fast-CASCL algorithm simplifies the decoding process and reduces the computational complexity through special node design (Rate-0, Rate-1, REP, SPC). Compared with the SC, SCL and other algorithms in the background technology, it significantly reduces the decoding delay and is especially suitable for communication scenarios with low delay requirements.
[0031] 2. In terms of hardware architecture design, a semi-parallel implementation architecture and a reasonable number of processing units are adopted to effectively balance resource consumption and decoding efficiency, making the Fast-CASCL algorithm have low computational complexity and high throughput, greatly improving the actual applicability of the algorithm in hardware decoders in short code scenarios, and has high engineering significance.
[0032] 3. Compared with the algorithms in the prior art, the Fast-CASCL algorithm designs a PM update calculation method based on hardware optimization for four fast decoding nodes, and extends these optimizations to the CASCL and other list algorithms with the best performance, further reducing the decoding delay while maintaining high decoding performance.
[0033] The practicality and effectiveness of the Fast-CASCL algorithm have been verified through simulation verification on Matlab and hardware implementation test on FPGA platform. The comprehensive verification from theory to practice shows the innovation and foresight of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 The design and test verification flow chart of the Fast-CASCL decoding algorithm.
[0035] Figure 2 This is the SC decoding diagram for N=8.
[0036] Figure 3 is the unit factor graph in the SC decoding structure.
[0037] Figure 4 This is the SC decoding tree structure diagram.
[0038] Figure 5 This is the structural diagram of the CASCL decoding algorithm.
[0039] Figure 6 It is a special structure diagram on the code length and code rate SC decoding tree.
[0040] Figure 7 This is the tree-type SC decoding structure diagram of the code length and code rate after pruning.
[0041] Figure 8 This is a schematic diagram of the Rate1 node path processing.
[0042] Fig. 9 It is the overall structure diagram of the Fast-CASCL decoder hardware design.
[0043] Fig.10 It is the state transition diagram implemented by the decoding controller.
[0044] Fig.11 Schematic diagram of the Fast-CASCL algorithm performance baseband test environment.
[0045] Fig.12 This is a performance comparison test chart between Fast-CASCL and CASCL algorithms (List=8,CRC4).
[0046] Fig.13 This is a performance comparison test chart of Fast-CASCL, SC, SCL and CASCL decoding algorithms.
[0047] Fig.14 It is the prototype verification diagram of Fast-CASCL hardware decoder.
[0048] Fig.15 This is the waveform result of the Polar codec baseband loopback hardware test.
[0049] Fig.16 This is a comparison chart between the Polar codec hardware test performance results and software simulation performance. DETAILED DESCRIPTION
[0050] The present invention is further described in detail below in conjunction with the test examples and specific implementation methods. However, this should not be understood as the present invention. The scope of the above subject is limited to the following embodiments, and all technologies implemented based on the content of the present invention belong to the scope of the present invention.
[0051] Example 1
[0052] The present invention is further described in detail below in conjunction with the test examples and specific implementation methods. However, this should not be understood as the present invention. The scope of the above subject is limited to the following embodiments, and all technologies implemented based on the content of the present invention belong to the scope of the present invention.
[0053] The detailed implementation scheme of the Fast-CASCL algorithm proposed in the present invention comprises the following steps:
[0054] S1. Determine the decoding tree structure: Based on the SC decoding algorithm, the Fast-CASCL algorithm is improved and developed using its decoding tree structure as a starting point. The present invention first analyzes the decoding tree structure in the SC algorithm in detail, especially focusing on its complexity in hardware implementation, and proposes a simplified, hardware-optimized calculation method, including forward and reverse hard bit propagation and simplified logical operations, to meet the needs of hardware implementation.
[0055] S2. Summarize and design four special fast decoding nodes, namely
[0056] (1) Rate-0 node: Processes a subtree with all leaf nodes as frozen bits. The decoding rule is simplified to directly output 0, which greatly reduces the computational complexity.
[0057] (2) Rate-1 node: processes a subtree whose leaf nodes are all information bits, and improves efficiency by simplifying the calculation instead of the complex decoding process.
[0058] (3) REP node: processes a subtree containing only one information bit and simplifies the decoding process through addition and threshold judgment.
[0059] (4) SPC node: a subtree that processes a single parity check code. This subtree contains only one frozen bit and uses low-complexity maximum likelihood (ML) decoding.
[0060] S3, designing a list decoding update rule for each fast decoding node to obtain the final Fast-CASCL algorithm;
[0061] (1) For the four fast decoding nodes described in S2, special PM update rules are designed, and a PM update algorithm based on hardware optimization is adopted to reduce the decoding computation complexity and delay while maintaining excellent decoding performance;
[0062] (2) The CRC-assisted SCL strategy is introduced for path selection. The CASCL algorithm generates CRC before encoding at the transmitting end and encodes the check code and information bits simultaneously. The receiving end uses CRC check combined with PM to select the optimal path, which greatly enhances the decoding performance and ensures high decoding efficiency and accuracy.
[0063] S4. Hardware architecture implementation based on Fast-CASCL algorithm
[0064] (1) Semi-parallel architecture: The present invention adopts a semi-parallel architecture based on SC operation, and implements semi-parallel operation through a reasonable number of processing units (PEs), thereby achieving a balance between resources and efficiency. This architecture also limits the size of special nodes on the decoding tree by defining the degree of parallelism, ensuring the flexibility and efficiency of hardware implementation.
[0065] (2) Hardware design: The decoder hardware design includes storage units and logic calculation units, including unit factor calculation core, channel layer LLR memory, intermediate layer LLR memory, feedback bit memory, path metric memory, address mapping memory, CRC unit and path memory, etc. These units cooperate to perform LLR calculation, hard bit value back propagation, path splitting, data preprocessing of four fast nodes and CRC-assisted decoding path selection.
[0066] (3) Decoding control: The decoding controller is responsible for controlling the entire decoding process, including the forward calculation of LLRs, the reverse propagation of the hard bit value B, and managing operations such as special node processing, path updating and screening. It guides the decoding process by storing the topological structure of the decoding tree after pruning (special node index).
[0067] S5. Further theoretical delay analysis of Fast-CASCL algorithm
[0068] (1) Fully parallel decoding assumption: In an ideal situation, assuming that the hardware can fully parallelize the calculations in the SC decoding structure, the decoding delay for Polar with a specific code length can be greatly reduced.
[0069] (2) Delay influencing factors: The decoding delay of the Fast-CASCL algorithm is mainly affected by the number of special nodes. Especially under low bit rate conditions, the proportion of certain low-latency type nodes increases, which further reduces the decoding delay.
[0070] (3) Practical hardware parallelism limitation: Although the parallelism is limited in the actual hardware implementation, the Fast-CASCL algorithm can still significantly reduce the latency through effective management and processing of special nodes, especially in various short code scenarios, achieving significant latency reduction compared with CASCL decoding.
[0071] S6. Build baseband software simulation and test environment, verify the rationality of Fast-CASCL, and test the performance indicators of Fast-CASCL hardware decoder
[0072] (1) By building a simulation environment on the Matlab platform, we simulated the performance of several common short and medium-sized Polar codes. By comparing the results, we confirmed that Fast-CASCL has significantly improved the performance compared with the SC and SCL algorithms while maintaining the performance level of CASCL.
[0073] (2) The Fast-CASCL hardware decoder was deployed on Xilinx's ZYNQ SoC XCZU9EG using the Vivado tool, and the hardware board verification was completed by building a prototype verification system. The system can simulate the complete communication process, including from signal source generation to BPSK modulation and demodulation, to Polar encoding and decoding, and finally performance testing.
[0074] The test results show that the Fast-CASCL hardware decoder works effectively at high clock frequencies, demonstrating the advantages of low resource consumption and high throughput, fully verifying its application potential in actual hardware implementation.
[0075] Example 2
[0076] This embodiment is a specific implementation of the method described in Example 1. The design and verification process of the efficient Polar code decoding method based on Fast-CASCL of the present invention is as follows: Figure 1 As shown, it mainly includes the following steps:
[0077] S1. Give the decoding tree structure of the SC decoding algorithm
[0078] Fast-CASCL is improved from CASCL with better decoding performance, and the basic calculation rules of CASCL come from SC. The decoding tree structure based on SC can easily summarize the fast decoding structure of the algorithm of the present invention. Figure 2 As an example, the SC unit factor decoding structure diagram with N=8 code length is shown in the figure. λ=0 is the decision layer, and the Decoding bits, channel layer λ = 3 receiving baseband soft demodulation LLR information The entire SC decoding process works from the channel layer to the decision layer, and the decoded bits are obtained serially. i=1,…,8. But in the original SC structure, Figure 3 The calculation method of the unit factor shown is too complicated and not suitable for hardware implementation. The present invention proposes a low-complexity calculation method optimized for hardware, including forward g, f operations and reverse hard bit propagation:
[0079]
[0080] The hard bit value B is a bit value of 0 or 1 determined by a formula according to the soft-demodulated LLR information.
[0081] Based on Figure 2 The dotted line division in the unit factor SC decoding structure diagram can be expressed as Figure 4 The SC decoding tree structure shown in Figure 1 is shown in Figure 2. Taking the node v at the λ layer as an example, v p is the parent node, v l With v r are left and right child nodes respectively, N v =2 λ The size of the node v contains 2 λ The LLR vector of length L v , back propagation 2 λ The long hard bit vector B v . Nodes v and v l 、v r The information exchange between Figure 3 The unit factor is equivalent to:
[0082]
[0083] Where sgn(x) is the sign function, the tree structure SC is Figure 3 The calculation process is the same as that of , which is just another equivalent form of expression. When the SC of the final decoding tree structure makes a decision on decoding, it also obtains the decoding bit according to the decision rule shown in the following formula:
[0084]
[0085] in is the transmitted information bit The subchannel index set of Indicates transmission freeze bit A set of subchannel indices, usually with the frozen bit set to 0.
[0086] Since the traditional SC decoding performance is very poor, some researchers later proposed a list decoding algorithm (SCL) for improvement. SCL considers using multiple paths to retain multiple decoding results, and selects the optimal path to output the decoding result according to the path metric (PM), thereby reducing the error propagation phenomenon of SC serial decoding. However, the actual decoding storage space is limited, and PM is required to perform path optimization, delete the poorer paths during the decoding process, and only retain the better paths. The traditional PM software calculation method based on LLR is too complicated and not conducive to hardware implementation. The present invention designs a PM update calculation method optimized for hardware:
[0087]
[0088] in The decoded bit The metric of the lth path of is the i-th decoded bit on the l-th path The corresponding LLR value is 1≤l≤List, 1≤i≤N, where List is the number of paths retained and N is the code length. The PM update calculation can be realized only by a simple addition rule, and the smaller the PM value, the better the decoding quality of the corresponding path.
[0089] However, the performance of the improved SC algorithm SCL is still far from the optimal decoding effect, and the Polar performance is still weaker than LDPC. Some researchers use CRC-assisted SCL for path selection (CASCL), which greatly enhances the decoding performance. The encoding and decoding structure is as follows Figure 5 In essence, CASCL is an algorithm that cascades the CRC and SCL processes. It only needs to generate CRC before encoding at the transmitting end, encode the check code and information bits at the same time, and the receiving end uses CRC check combined with PM to select the optimal path.
[0090] S2. Summarize four special fast decoding nodes and design decoding methods for each
[0091] The present invention takes the Polar code with N=8 code length and R=1 / 2 code rate as an example and proposes the following based on the SC decoding tree structure: Figure 6 The four special decoding structures shown form four decoding nodes in the tree: Rate-0 node, Rate-1 node, REP node and SPC node.
[0092] Rate-0 node: All leaf nodes of the subtree with this node as the root are frozen bits, such as Figure 6 The nodes are numbered {4,8,9,10}. v The decoding rule for the Rate-0 node v of size is that the decoding bits on the leaf nodes of the subtree with υ as the root node are All are 0, 0≤z≤log2λ. Feedback to the parent node v p Hard bit value B v The vector is all 0.
[0093] Rate-1 node: All leaf nodes of the subtree with this node as the root are all information bits, such as Figure 6 The nodes are numbered {7,11,13,14,15}. v The Rate-1 node υ of the same size is fed back to the parent node v p Hard bit value B v The judgment formula is as follows:
[0094]
[0095] in, is the soft information of node v at the i-th decoding bit, i = 1, ..., N v , N v is the code length of node v; then, find all the information bits corresponding to the Rate-1 node in The dimension of node v is N v This calculation method effectively replaces the (calculated laborious) "soft information" with the (calculated trivially) "hard information", saving considerable computational effort on the hardware.
[0096] SPC node, at any code rate (N-1) / N, the frozen bit is always u1, this code type is called single parity check code (SPC), such as Figure 6 {3,5,6} in the node, where {5,6} is also a REP node. Of all the leaf nodes of the subtree with this node as the root, only one is a frozen bit, and the rest are all information bits. For SPC nodes, extremely low-complexity maximum likelihood (ML) decoding can be performed: the hard decision estimate and parity check value of the SPC node decoder are calculated. If the parity check constraint is not satisfied, the estimate of the least reliable bit is reversed. For a size of N v The SPC node v is first input with LLR soft information L v Make hard decisions on the elements in:
[0097]
[0098] where i = 1, ..., N v According to the hard decision bit vector HD v Get the parity value: Then according to Find the least reliable index j, and according to the above conditions, get the final SPC node v and return it to the parent node v p The hard bit value B v for:
[0099]
[0100] After getting the B of SPC node v v After that, the Rate-1 node decoding formula can also be used Direct decoding.
[0101] In the leaf nodes of the subtree with REP node v, only one bit is an information bit, and the rest are frozen bits. Figure 6 {3,5,6} in the example are REP type, and the node {5,6} is also SPC type. The decoding rule of REP is: for all input soft information L v The value of is summed, and then the threshold is detected using the symbol function. The final detection result is copied as REP node v and returned to the parent node v p The bit decision value B v :
[0102]
[0103] where i, j = 1, ..., N v That is, when the sum of the soft information input to REP node v is greater than 0, Otherwise the returned hard bit value is a vector of all 1s. The leaf node of the subtree with REP node v has only one information bit, set and i=1,2,…,N v . And REP node v contains Code group The returned hard bit value B v It also complies with the Rate-1 node decoding calculation rules.
[0104] Based on the above four fast decoding nodes, the decoding process can be simplified, and there is no need to traverse the entire Figure 6 The decoding tree shown in Figure 1. Under ideal computing resources, the final code length N = 8 and the code rate is 1 / 2. Figure 6 The decoding tree shown can be pruned and optimized to Figure 7As shown in the decoding tree, the Polar decoding of this code type only requires two special operations of the REP node and the SPC node.
[0105] S3. Design a list decoding update rule for each fast decoding node to obtain the final Fast-CASCL algorithm
[0106] Since this algorithm is improved based on the CASCL algorithm, in order to ensure that Fast-CASCL continues the excellent decoding performance of CASCL while reducing the decoding delay, it is necessary to use the four fast decoding nodes (Rate-0, Rate-1, SPC and REP) proposed by S2 to design the path update rules for special nodes, that is, the calculation and path optimization of PM when splitting the path during special node list decoding. The PM calculation method for special nodes is designed based on the PM update calculation formula (4) suitable for hardware implementation in S1 list decoding, that is:
[0107]
[0108] This PM calculation method can ensure that the decoding complexity of the special node list in the actual decoder is very low.
[0109] For a N v For a Rate-0 node v with a size of 1,0 ... Give the parent node v p . Let the unreliability measure brought by Rate-0 node v on path l be Where l = 1, 2, ..., List. At the same time, the Rate-0 node on path l returns a vector with all elements being 0. The soft information L of node v input in the ideal decoding environment v [l] and the returned hard bit B v [l] should comply with the hard decision rule (3) for SC decoding in S1, since the B returned by the Rate-0 node v [l]={0}, then i=1,2,…,N v In actual decoding systems, unexpected errors may cause L v [l] and B v [l] does not conform to the judgment relationship, resulting in path unreliability The calculation process is as follows:
[0110]
[0111] The Rate-0 node v does not split the path, and the number of paths remains unchanged, but the path metric of the original path l changes:
[0112]
[0113] However, under the ideal hardware architecture, the Rate-0 node does not need to split the path and filter it, but only needs to update the path information, and the processing delay is T Rate-0 =1.
[0114] The path processing for REP node v is also relatively simple, because the decoding node in the node subtree has only one information bit, resulting in the hard bit B returned to the parent node v Only all 1 vectors and all 0 vector There are two situations. The REP decoding bits do not satisfy the decision relationship and the unreliability of path l is caused by the following two situations:
[0115]
[0116] This means that if the hard bits corresponding to path l are (all 1 vector), the path unreliability brought by node v is the input soft information greater than 0 The absolute value and; similarly, the hard bits on the path are The input soft information that brings unbelievable behavior to path l is less than 0 For the original List paths, after a split operation of REP node v, 2*List new paths are generated, and the following metric value changes are brought to the original paths, where l = 1, 2, ..., List:
[0117]
[0118] Obviously, under the ideal hardware architecture, the REP node only needs to perform one path split and one path information update operation, and the processing delay is T REP =2.
[0119] For a size of N v The Rate-1 node v, corresponding to N v All decoded bits are information bits, and each information bit The processing will split the two paths corresponding to the 0 and 1 decision bits. This type of node is similar to the path splitting rule in the SCL algorithm, except that it now needs to be fed back to the parent node v according to the decision criteria of the Rate-1 node. p The hard bit value B v [l] to perform the relevant path splitting. Then return to v pHard bit value The metric increase of path l is The unreliability of Then return The path l increases a certain degree of unreliability. Eventually, it will split Paths, and the paths need to be screened according to the PM size to retain the List of preferred paths. However, considering the one-time processing of N v The path splitting operation of each bit will increase the hardware complexity and resource consumption. The actual hardware decoder often uses time division multiplexing to process the path splitting of Rate-1 nodes in multiple shots. The processing process is as follows: Figure 8 Each split doubles the number of paths. Figure 8 The red dotted box in the middle represents the selected path, and the number of splits also corresponds to the number of times the hard bits are processed.
[0120] Let's assume that PM l is the metric value of the original path of List before the splitting operation of Rate-1 node υ. Before the path splitting, it is necessary to make a hard decision on the input soft information of each path l according to the following formula to obtain the hard bit sequence B v [l]:
[0121]
[0122] In order to minimize the increase in PM caused by the failure of the path to comply with the judgment result during the path splitting process, it is necessary to deal with the soft information in the node The smaller corresponding decoding bit i is processed first, thereby reducing the dynamic range of PM, reducing PM quantization requirements and saving hardware resources. v The elements in [l] are arranged in ascending order of absolute value: First, follow the principle of processing from small to large, and give priority to The corresponding bit Then process the hard bits one by one Corresponding hard bits There are two pre-decision results {0, 1}, and the corresponding path l splits into two new paths, where j = i1, i2, ..., N v .by Figure 8 Take the new path branched out each time as an example, the left branch is feedback According to the hard decision criterion, the corresponding path metric is updated, that is, the PM of the left branch will not change; while the right branch is feedback The path metric update is opposite to the decision criterion, that is, the PM of the right branch will increase The unreliability of path l. Then the metric of path l The update guidelines are as follows:
[0123]
[0124] in According to the calculation formula (5) of the hard bit value of Rate-1 node feedback in S2, each time a hard bit value After the corresponding path splitting operation, 2List new paths will be generated. These paths need to be screened to find the path with the smallest PM. Based on these paths, the index is processed in turn. The split path corresponding to the bit, until N v Hard bit value B v After all processing is completed and feedback is given, the filtered List path is finally obtained.
[0125] But in fact, Rate-1 nodes do not necessarily need to process all N v bits. When List≤N v When , if the root nodes of the finally screened List paths are the same, or the List paths are all split from a certain path l. Obviously, the path metric is The paths where a≥List will be deleted and discarded, and the metrics are The path with smaller PM will be retained. Therefore, in this case, only List-1 hard bits need to be split, and the remaining N v -(List-1) bits of hard bit value only need to use the soft information judgment result. If the root nodes of the final filtered List paths are different, the speed of filtering the paths will only be faster. v When N v The path splitting operation corresponding to the hard bit. In summary, the decoding time step of the Rate-1 node is T Rate-1 = min(List-1, N v ).
[0126] For N v -1 information bit N v The path splitting method of the SPC node v is similar to that of Rate-1, except that its hard bit B v The decision of [l] depends not only on the soft information value L v , is also affected by the parity check result, namely:
[0127]
[0128] Therefore, during the path splitting process, B v [l] The number of hard bit changes needs to satisfy the even number constraint, that is, the hard bit number in the subsequent path splitting operation It is necessary to invert an even number of bits at the same time so that the new path can ensure that the parity of the new path is consistent with the original path. The absolute value of the input soft information is arranged in ascending order to obtain Also, in order to reduce the dynamic range of PM, the bit with the smallest sum of the absolute values of an even number of LLRs is preferentially selected for hard bit inversion processing.
[0129] For the time complexity of SPC nodes, when List≤N v hour:
[0130] If the parity check value parity = 0, the input soft information is judged using formula (6) This is the hard bit value of SPC in At this time, the metric value of the split mother path is PM[l], l = 1, 2, ..., List. If the List paths processed and filtered by the SPC node v are all split from the same mother path k, it is obvious that the path metric is The paths with a≥List+1 are not in the preferred range, and the remaining metrics are PM l , The path is within the final optimization range, and the node processing time step is List-1. If the final filtered List paths are split from different parent paths, the speed of splitting and filtering the paths will only be faster.
[0131] If the parity check value is 1, according to equation (16), the hard bit corresponding to the minimum soft information (j=i1) has been reversed, so the paths in this list are all measured by The mother path is split from, where l = 1, 2, ..., List. In the corresponding path processing, both hard bit values need to be inverted, that is, inverted at the same time and in The final measurement is The path is within the preferred range, and List-1 needs to be processed similarly.
[0132] When List>N v When the node is similar to Rate-1, it also needs to perform N v The path splitting corresponding to the hard bit. Since SPC has an additional process of finding the least reliable bit compared to the path processing of Rate-1 nodes, the time step of SPC node path splitting is T SPC = min(List-1, N v)+1.
[0133] S4. Hardware architecture implementation based on Fast-CASCL algorithm
[0134] In the Polar hardware decoder based on SC operation, such as Figure 3 The unit factor shown is designed as a processing element (PE). The present invention is based on a semi-parallel implementation architecture, uses a reasonable number of PEs to implement semi-parallel operations, and defines a parallelism P. The unit factor calculation of each layer of decoding can reuse these P PEs to achieve a compromise between resources and efficiency. In addition, the upper limit of the parallelism P also limits the size of the special node v on the decoding tree, that is, N v ≤P. The hardware architecture design of the semi-parallel Polar decoder based on the Fast-CASCL algorithm is as follows Fig. 9 shown.
[0135] Fig. 9 In the figure, the blue module belongs to the storage unit, the gray module belongs to the logic calculation unit, the thick solid line is the data signal path, and the dotted line represents the control signal. The decoder receives the LLR soft information from the previous demodulation module according to the handshake signal and stores it in the channel layer LLR memory. Then the channel layer data will be sent to List unit factor calculation cores for related calculations, including forward LLR calculation, backpropagation hard bit value calculation and four types of fast node data preprocessing, and the unit factor calculation core is designed with a P parallel semi-parallel architecture. The generated LLR results and hard bit values B will be stored in the middle layer LLR memory and feedback bit memory respectively under the control signal. When the calculation core reaches the middle layer, it is necessary to continuously fetch data from these two memories according to the Figure 3Iterative operation. When the decoding process reaches the leaf node (decision layer) of the decoding tree or encounters four special fast nodes, the controller will enter the fast node algorithm process shown in S3 according to the special node index control process, that is, the node path splitting, path sorting, path optimization and path update, generate a list of paths and corresponding decoding results, and the metric information of these paths will be stored in the path metric memory. After the optimization, the path update information will be implemented using a special storage address mapping memory module to reduce hardware complexity. The node hard bit value corresponding to each path is stored in the feedback bit memory, and the decoding result is stored in the path memory. The calculation processing, path processing, and path update operations are continuously repeated until the decoder traverses the pruned decoding tree, which means that the N-length code block has been preliminarily decoded. Finally, with the assistance of CRC, the best decoding result is output from the List of paths in the path memory. Among them, the special node index in the decoding controller stores the fast node index under the P parallelism limit, which is equivalent to storing the topological structure of the pruned decoding tree. The entire decoding process is subject to the constraints of the decoding control, and its control state transition diagram is as follows Fig.10 As shown, it is basically similar to the above process.
[0136] Theoretical delay analysis of S5 and Fast-CASCL algorithm
[0137] When Arikan proposed Polar Code, he pointed out that if there is no bottleneck in the actual hardware, such as Figure 2 When all the f operations and g operations optimized for hardware as shown in equation (1) in each layer λ of the SC decoding structure shown in the figure can be completed simultaneously in one clock cycle (i.e., full parallel decoding), for a Polar code with a code length of N, the delay of its SC decoding is T SC =2N-2. At the same time, Arikan uses N=8 as an example to give the worst-case complexity of the SC decoding algorithm as O[N·logN]. Assuming full parallel decoding, Table 1 uses code length N=8 as an example to illustrate the operations performed in each clock cycle of SC, where λ represents the layer number in the decoding structure where the unit factor is working, and f and g indicate the operations performed in the unit factor.
[0138] Table 1 Decoding delay statistics of SC decoder with N=8
[0139]
[0140] For the list decoding algorithm, since CASCL only adds a CRC link on the basis of SCL, ideally CRC has no essential effect on the decoding delay. Similarly, if there is no bottleneck in the actual hardware, the hardware-optimized f and g operations shown in equation (1) can be completed within one time step, where each layer of the decoding structure is fully parallelized, and the path sorting and optimization can be completed within one clock step, then the list algorithm decoding delay is T cascl =T scl =2N+K-2. By utilizing the recursiveness of Polar decoding calculation and the special assignment mechanism during data update, list decoding can be implemented with a complexity of O[List·NlogN].
[0141] From the description of the Fast-CASCL algorithm in S3, we can see that when the actual hardware parallelism has no bottleneck, the processing delay of REP and Rate-0 nodes is not affected by the node size N. v , while SPC and Rate-1 nodes are affected by the number of overall lists and node size, i.e.
[0142]
[0143] It can be seen that the overall decoding delay of the Fast-CASCL algorithm is affected by the distribution of the number of four types of nodes. The lower the code rate, the higher the proportion of REP and Rate-0 nodes, and the lower the overall decoding delay of the algorithm. Take the three common short code types (N, K) as an example, the encoding CRC length is 4, and the hardware decoding parallelism is 64 (this limits the special node size N). v ≤64), the number of four special nodes contained in these code types is shown in Table 2:
[0144] Table 2 Statistics of node number and decoding delay
[0145]
[0146] Compared with CASCL decoding, Fast-CASCL algorithm has 69.6%, 82.1% and 72.6% latency reduction on these three code types respectively.
[0147] Example 3
[0148] This embodiment is a specific implementation of the method described in Embodiment 1, which builds a baseband software simulation and test environment, verifies the rationality of Fast-CASCL, and tests the performance indicators of the Fast-CASCL hardware decoder.
[0149] The decoding performance of Fast-CASCL was analyzed and the following Fig.11 The simulation environment shown in the figure. Four common short and medium code types (512, 356), (128, 64), (128, 32) and (64, 32) Polar codes in communication scenarios are simulated, and the BER of SC, SCL and CASCL algorithms are compared and analyzed to illustrate the rationality of Fast-CASCL decoding performance. The simulation environment uses BPSK baseband modulation and the corresponding soft demodulation algorithm, and the channel environment uses the common AWGN channel. Since Fast-CASCL is based on CRC-assisted decoding, a similar method is required before encoding at the transmitter. Figure 5 The CRC process is shown in Figure 1. Fig.12 The performance comparison test of Fast-CASCL and CASCL algorithms is shown in FIG. 1 , where the CRC length of both algorithms is 4 and the list size List is 8. Since the Fast-CASCL algorithm proposed in the present invention improves the decoding delay based on CASCL only, under the influence of fast nodes, it can be seen that the decoding performance of Fast-CASCL is still maintained at the level of CASCL and is not affected in any way under various short code structures. Compared with other decoding algorithms such as Fig.13 As shown, Fast-CASCL has a greater performance gain than traditional SC and SCL, has a maximum gain of 1.5dB compared to the traditional SC algorithm, and has an extremely high coding gain compared to ordinary BPSK communication, indicating that the algorithm proposed in the present invention still maintains the optimal level in decoding performance.
[0150] The present invention deploys Fast-CASCL hardware decoder on Xilinx's ZYNQ SoC XCZU9EG based on Vivado tool, realizes front-end simulation and back-end comprehensive implementation by using Vivado tool, and completes the board verification of decoder. Fig.14 The Polar codec prototype verification system shown is modeled after Fig.11 The software simulation system in is built, and the random source generator, BPSK modulation and soft demodulation, AWGN digital channel simulator and Polar encoder are implemented based on the Verilog language. The test IPs such as ILA and VIO are combined to control and capture the input and output data of the decoder.
[0151] The decoder supports three polar code types (120,60), (120,30), (60,30) which are rate matched from (128,64), (128,32), (64,32), and the decoder can dynamically adjust between these three code types in real time. Similarly, the polar encoder at the transmitting end also supports dynamic bit rate switching. In order to balance hardware resource consumption and decoding efficiency, the decoder designed by the present invention is set by the following parameters:
[0152] Parameters Numeric (N,K) (120,60)、(120,30)、(60,30) CRC length 4 Parallelism P 16 List Number 8 <![CDATA[Fast node size N v > <![CDATA[N v ≤16]]> LLR quantification Q6(1,2,3)
[0153] Where N is the code length, K is the information bit length, and the quantization index Q7 (1, 2, 4) represents 1 sign bit, 2 integer bits, and 3 decimal bits. The resource overhead of the decoder baseband loopback test system finally implemented on the ZYNQ chip is shown in the following table:
[0154] Resource Project Use quantity Available quantity Utilization % LUT 52925 274080 19.3 LUTRAM 896 144000 0.62 FF 11840 548160 2.34 BRAM 6.5 912 0.6 DSP 0 2520 0
[0155] It can be seen that the Polar decoder designed in the present invention has low resource consumption and has high application prospects on resource-constrained platforms.
[0156] The Fast-CASCL hardware decoder test pattern of the present invention is (120,60). b / N0=2.5dB AWGN channel environment, test 10 6 The encoded data of the frame, and the key waveform finally captured by ILA is as follows Fig.15 As shown in the waveform, you can see the comparison between the source data and the decoded output data, such as the red box and the white box. Obviously, the decoded data in the red box has errors, while the decoded data in the white box is completely correct. These error data are counted in the indicators of bit error and frame error, as shown in the lower half of the waveform. These data mainly count the number of bit errors in the current decoding (curr_frame_err_bits), the total number of error bits so far (total_num_err_bits), the total number of error frames so far (total_num_err_frames), the total number of frames received so far (total_num_rx_frames), and the total number of bits received so far (total_num_rx_bits).
[0157] Finally, the Fast-CASCL hardware decoder in the present invention can work at a main clock frequency of 200MHz. The three code types (120, 60), (120, 30), and (60, 30) are tested respectively. The waveform data is captured and analyzed in ILA, and the throughput and delay information are shown in the following table:
[0158] Code Type Average decoding delay (clock cycles) Throughput Mbps (120,60) 280 42.86 (120,30) 154 38.96 (60,30) 117 51.28
[0159] Due to the semi-parallel structure constraints and the delay of the intermediate control process of the Fast-CASCL decoder, the actual decoding delay is somewhat different from the theoretical analysis, but the throughput is within a reasonable range.
[0160] In ILA, set the total number of frames received (total_num_rx_frames) as the trigger condition to capture the error test data. Constantly change the signal-to-noise ratio of the test channel through VIO b / N0, output the error statistics under each signal-to-noise ratio to Matlab for further analysis, and obtain the following Fig.16 The bit error rate curve is shown. After comparison, it is found that the performance of the hardware Polar encoder and decoder based on FPGA test is basically consistent with the Q6 fixed-point simulation data, and the maximum difference from the floating-point performance is only 0.5dB, which verifies that the functions of the encoder and decoder are correct.
[0161] In summary, the Fast-CASCL algorithm proposed in the present invention can effectively reduce decoding delay while maintaining high decoding performance; in addition, the decoding calculation method proposed in the present invention for hardware optimization can reduce the complexity of the hardware decoder and make the architecture more flexible. By building a baseband hardware test environment on FPGA and combining software simulation analysis, the low resource consumption and high throughput performance of the decoder of the present invention are proved.
Claims
1. A flexible and efficient Polar code decoding method, characterized in that: The following steps are involved: S1: According to the structural characteristics of the SC decoding tree, four types of decoding nodes are summarized, namely, a.Rate-0 node: a subtree with all leaf nodes as frozen bits; b. Rate-1 node: a subtree whose leaf nodes are information bits; c. REP node: a subtree containing only one bit of information; d.SPC node: a subtree of a single parity check code, a subtree containing only one frozen bit; S2: Calculate the log-likelihood ratios (LLRs) of the Rate-0 node, the Rate-1 node, the SPC node, and the REP node respectively; S3: designing a list decoding method for the four decoding nodes respectively; S3.1: All the decoding bits corresponding to the leaf nodes of the subtree with Rate-0 as the root are frozen bits. According to the list decoding rule, there is no need to split the paths corresponding to these bits, and only hard bits are returned to the parent node; S3.2: For Rate-1 nodes, the corresponding decoded bits are all information bits. The processing of each information bit will be split into two paths corresponding to the decision bits 0 and 1, and the hard bit value will be fed back to the parent node to perform path splitting related to the information bit; S3.3: For all leaf nodes of the subtree with the SPC node as the root, only one is a frozen bit, and the rest are all information bits. First, a hard decision is made on the elements in the input soft information, and then the least reliable index is found. The hard bit value returned by the final SPC node to the parent node is obtained based on the hard decision and the least reliable index, and then the information bit-related paths are split based on the hard bit value and the least reliable index. S3.4: In the leaf nodes of the subtree with REP node as the subtree, only one bit is the information bit, and the rest are frozen bits. The values of all input soft information are summed, and the threshold detection is performed using the sign function. The final detection result is copied as the bit decision returned by the REP node to the parent node, splitting into two decoding paths; S4: Design path metric update rules for hardware optimization for the four decoding nodes, namely in, The decoded bit The metric of the lth path of is the i-th decoded bit on the l-th path The corresponding LLR soft information is 1≤l≤List, 1≤i≤N; List is the number of retained paths, and N is the code length; the path metric update rule is used to split the path at the node, the PM value of each path is calculated, the paths are sorted according to the PM value, the List of preferred paths are retained and the paths are updated; CRC check is performed on the retained List paths, and the path that passes the check and has the smallest PM value is the optimal path, the decoding bits corresponding to the optimal path are the decoding results of the Fast-CASCL algorithm, and the decoding results are output.
2. According to the flexible and efficient Polar code decoding method of claim 1, a theoretical delay analysis method for the Fast-CASCL algorithm is designed, characterized in that: Each layer of the decoding structure can achieve full parallel operation, path sorting and optimization can be completed within one clock step, and the delay rate of the Fast-CASCL decoding algorithm is statistically analyzed.
3. A flexible and efficient Polar code decoder, characterized in that: A flexible and efficient Polar code decoding method based on any one of claims 1-2, the decoder comprises: a channel layer LLR memory, an intermediate layer LLR memory, a selection module, a calculation core module, a feedback bit memory, a path splitting module, a PM sorting module, a path selection module, a path memory, a path metric memory, a fast decoding node index memory, a CRC check unit and a decoding result memory; The LLR data received from the channel layer is placed in the channel layer LLR memory. During decoding, the channel layer LLR memory sends the LLR data to the computing core module for operation. The core computing structure is calculated using the unit factor to calculate the LLR values corresponding to the four fast decoding nodes and other data required for node processing, and the LLR values are stored in the intermediate layer LLR memory. When the computing core works to the decision layer or encounters four fast decoding nodes, the four fast decoding nodes are split into paths, the PM value of each path is calculated, the paths are sorted according to the PM value, and a list of preferred paths are selected and retained. The address mapping memory is used to update the information index of the selected paths, and the generated hard bit values are stored in the feedback bit memory. The generated decoding results are stored in the path memory along with the paths. A CRC check is performed on the list of paths retained in the path memory. The path that passes the check and has the smallest PM value is the optimal path. The decoding bits corresponding to the optimal path are output as the decoding result of the Fast-CASCL algorithm, and the decoding result is stored in the decoding result memory.
4. A flexible and efficient Polar code decoder according to claim 3, characterized in that: In the computing core module, a semi-parallel computing core is used, and the upper limit parallelism of the number of computing units PE in the decoder is set to P, where P is the number of unit factor computing cores.
Citation Information
Cited By
5G PDCCH (Physical Downlink Control Channel) blind detection method
CN121792005A
