低密度奇偶校验码LDPC（六）——SPA和积译码算法

往期博文

低密度奇偶校验码LDPC（一）——概述_什么是gallager构造-CSDN博客

低密度奇偶校验码LDPC（二）——LDPC编码方法-CSDN博客

低密度奇偶校验码LDPC（三）——QC-LDPC码概述-CSDN博客

低密度奇偶校验码LDPC（四）——双对角线结构的QC-LDPC编码-CSDN博客

低密度奇偶校验码LDPC（五）——译码算法概述-CSDN博客

低密度奇偶校验码LDPC（六）——SPA和积译码算法-CSDN博客

QC-LDPC的FPGA实现

基于QC-LDPC编码的循环移位网络的FPGA实现_5g ldpc编码 桶形移位寄存器-CSDN博客

一、和积算法基本原理

 二、概率域上的和积算法

传递消息定义

 概率域的迭代更新公式

 三、对数域上的和积算法

预备知识：对数似然比

预备知识：Gallager引理

 VN消息初始化

 对数域的迭代更新公式

 对数域SPA算法总结

四、SPA算法的调度策略

 泛洪调度策略

分层调度策略

 仿真对比 调度策略总结

 泛洪SPA译码算法

 分层SPA译码算法

五、SPA算法的Matlab实现

泛洪调度(Flooding Schedule Scheme)

function [x_hat, iter_this_time] = Flooding_BP_decoder(llr, H_row_one_absolute_index, H_comlumn_one_relative_index, N, M, vn_degree, cn_degree, max_iter)
VN_array = zeros(max(vn_degree), N);
CN_tanh_tmp = zeros(max(cn_degree), 1);
iter_this_time = max_iter;
for t = 1 : max_iter
    %% L(q_ij) Update
    sum_VN = sum(VN_array);
    for v = 1 : N
        for v_neighbor = 1 : vn_degree(v)
            VN_array(v_neighbor, v) = sum_VN(v) - VN_array(v_neighbor, v) + llr(v);
        end
    end
    
    %% L(sigma_ji) Update
    for c = 1 : M
        product = 1;
        for c_neighbor = 1 : cn_degree(c)
            CN_tanh_tmp(c_neighbor) = 1 - 2/(1 + exp(VN_array(H_comlumn_one_relative_index(c, c_neighbor), H_row_one_absolute_index(c, c_neighbor))));
            product = product * CN_tanh_tmp(c_neighbor);
        end
        for c_neighbor = 1 : cn_degree(c)
            x = product/CN_tanh_tmp(c_neighbor);
            x = sign(x) * min(abs(x), 1 - 1e-15);% Numerical Stability.
            VN_array(H_comlumn_one_relative_index(c, c_neighbor), H_row_one_absolute_index(c, c_neighbor)) = log((1 + x)/(1 - x));
        end
    end
    
    %% L(Qi) Update
    x_hat = (sum(VN_array)' + llr) < 0;% Belief propagation Decision.
    parity_check = zeros(M, 1);
    for m = 1 : M
        for k = 1 : 1 : cn_degree(m)
            parity_check(m) = parity_check(m) + x_hat(H_row_one_absolute_index(m, k));
        end
    end
    if ~sum(mod(parity_check, 2))% early stop, to see whether Hx = 0.
        iter_this_time = t;
        break;
    end
end

分层迭代(Layered Schedule Scheme)

function [x_hat, iter_this_time] = Layered_BP_decoder(llr, H_row_one_absolute_index, H_comlumn_one_relative_index, N, M, vn_degree, cn_degree, max_iter)
VN_array = zeros(max(vn_degree), N);
CN_tanh_tmp = zeros(max(cn_degree), 1);
iter_this_time = max_iter;
for t = 1 : max_iter
    for c = 1 : M
        product = 1;
        for c_neighbor = 1 : cn_degree(c)
            Lji = sum(VN_array(:, H_row_one_absolute_index(c, c_neighbor))) + llr(H_row_one_absolute_index(c, c_neighbor)) - VN_array(H_comlumn_one_relative_index(c, c_neighbor), H_row_one_absolute_index(c, c_neighbor));%VN update. However, this simple MATLAB sentence consumes a lot of time.
            CN_tanh_tmp(c_neighbor) = 1 - 2/(1 + exp(Lji));
            product = product * CN_tanh_tmp(c_neighbor);
        end
        for c_neighbor = 1 : cn_degree(c)
            Lij = product/CN_tanh_tmp(c_neighbor);
            Lij = sign(Lij) * min(abs(Lij), 1 - 1e-15);
            VN_array(H_comlumn_one_relative_index(c, c_neighbor), H_row_one_absolute_index(c, c_neighbor)) = log((1 + Lij)/(1 - Lij));
        end
    end
    
    x_hat = (sum(VN_array)' + llr) < 0;
    parity_check = zeros(M, 1);
    for m = 1 : M
        for k = 1 : 1 : cn_degree(m)
            parity_check(m) = parity_check(m) + x_hat(H_row_one_absolute_index(m, k));
        end
    end
    if ~sum(mod(parity_check, 2))
        iter_this_time = t;
        break;
    end
end

 参考文献

[1] 白宝明 孙韶辉 王加庆. 5G 移动通信中的信道编码[M]. 北京: 电子工业出版社, 2018.

[2] William E. Ryan, Shu Lin. Channel Codes - Classical and Modern[M]. Cambridge University Press, 2009.

[3] Gallager R. Low-density parity-check codes[J]. IRE Transactions on information theory, 1962, 8(1): 21-28.