Skip to content

Improved the Matlab version #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
qizhu8 wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Improved the Matlab version #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d863b32
largely improved the EnDec alg's performance
Jan 5, 2017
9523704
changed the README.md
Jan 5, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
227 changes: 129 additions & 98 deletions LdpcM/LDPCCode.m
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
classdef LDPCCode < handle
% Implements LDPC functionality
%
%

properties
H;
Expand All @@ -14,7 +14,7 @@
col_mat;
row_mat;

Z;
Z;
end

methods
Expand Down Expand Up @@ -42,8 +42,8 @@ function efficient_pcm(obj)
obj.row_weight_vec = sum(obj.H, 2);
obj.max_col_weight = max(obj.col_weight_vec);
obj.max_row_weight = max(obj.row_weight_vec);
obj.col_mat = zeros(obj.N, obj.max_col_weight);
obj.row_mat = zeros(obj.M, obj.max_row_weight);
obj.col_mat = zeros(obj.N, obj.max_col_weight);
obj.row_mat = zeros(obj.M, obj.max_row_weight);
for i_row = 1 : obj.M
index = 1;
for i_col = 1 : obj.N
Expand All @@ -67,114 +67,145 @@ function efficient_pcm(obj)
end

function [codeword] = encode_bits(obj, info_bits)
% Does encoding by back substitution
% Assumes a very specific structure on the partiy check matrix
codeword = zeros(obj.N, 1);
codeword(1:obj.K) = info_bits;

parity = zeros(obj.M, 1);
for i_row = 1 : obj.M
for i_col = 1 : obj.max_row_weight
if (obj.row_mat(i_row, i_col) > 0) && (obj.row_mat(i_row, i_col) <= obj.K)
parity(i_row) = parity(i_row) + codeword(obj.row_mat(i_row, i_col));
end
end
end

parity = mod(parity, 2);

for i = 1 : obj.Z
codeword(obj.K + i) = mod(sum(parity(i:obj.Z:end)), 2);
end

for i_row = 1 : obj.M
for i_col = 1 : obj.max_row_weight
if (obj.row_mat(i_row, i_col) > obj.K) && (obj.row_mat(i_row, i_col) <= obj.K + obj.Z)
parity(i_row) = mod(parity(i_row) + codeword(obj.row_mat(i_row, i_col)), 2);
end
end
end

for i_col = obj.K + obj.Z + 1 : obj.Z: obj.N
% Does encoding by back substitution
% Assumes a very specific structure on the partiy check matrix
codeword = zeros(obj.N, 1);
codeword(1:obj.K) = info_bits;


parity = obj.H * codeword;
% parity = mod(parity, 2);

for i = 1 : obj.Z
codeword(obj.K + i) = mod(sum(parity(i:obj.Z:end)), 2);
end

H_part2 = obj.H(:, obj.K+1:obj.K+obj.Z);
codeword_part2 = codeword(obj.K+1:obj.K+obj.Z);
parity = mod(parity + H_part2 * codeword_part2(:), 2);

for i_col = obj.K + obj.Z + 1 : obj.Z: obj.N
codeword(i_col: i_col + obj.Z - 1) = parity(i_col - obj.K - obj.Z : i_col - obj.K - 1);
parity(i_col - obj.K : i_col - obj.K + obj.Z - 1) = mod(parity(i_col - obj.K - obj.Z : i_col - obj.K - 1) + ...
parity(i_col - obj.K : i_col + obj.Z - obj.K - 1), 2);
end

end

% parity = zeros(obj.M, 1);
% for i_row = 1 : obj.M
% for i_col = 1 : obj.max_row_weight
% if (obj.row_mat(i_row, i_col) > 0) && (obj.row_mat(i_row, i_col) <= obj.K)
% parity(i_row) = parity(i_row) + codeword(obj.row_mat(i_row, i_col));
% end
% end
% end
%
% parity = mod(parity, 2);
%
% for i = 1 : obj.Z
% codeword(obj.K + i) = mod(sum(parity(i:obj.Z:end)), 2);
% end
%
% for i_row = 1 : obj.M
% for i_col = 1 : obj.max_row_weight
% if (obj.row_mat(i_row, i_col) > obj.K) && (obj.row_mat(i_row, i_col) <= obj.K + obj.Z)
% parity(i_row) = mod(parity(i_row) + codeword(obj.row_mat(i_row, i_col)), 2);
% end
% end
% end
%
% for i_col = obj.K + obj.Z + 1 : obj.Z: obj.N
% codeword(i_col: i_col + obj.Z - 1) = parity(i_col - obj.K - obj.Z : i_col - obj.K - 1);
% parity(i_col - obj.K : i_col - obj.K + obj.Z - 1) = mod(parity(i_col - obj.K - obj.Z : i_col - obj.K - 1) + ...
% parity(i_col - obj.K : i_col + obj.Z - obj.K - 1), 2);
% end

end

function [decoded_codeword, error_vec] = decode_llr(obj, input_llr_vec, max_iter, min_sum)
eta = zeros(obj.M, obj.N);
lasteta = zeros(obj.M, obj.N);
error_vec = 0;
updated_llr_vec = input_llr_vec;
error_vec = zeros(max_iter, 1);
for iter = 1 : max_iter

for i_m = 1 : obj.M
for i_n1 = 1 : obj.row_weight_vec(i_m)
n1 = obj.row_mat(i_m, i_n1);
if min_sum
if min_sum
eta = zeros(obj.M, obj.N);
lasteta = zeros(obj.M, obj.N);
for iter = 1 : max_iter
for i_m = 1 : obj.M
for i_n1 = 1 : obj.row_weight_vec(i_m)
n1 = obj.row_mat(i_m, i_n1);
pr = 100;
else
pr = 1;
end
for i_n2 = 1 : obj.row_weight_vec(i_m)
if i_n1 == i_n2
continue;
end
n2 = obj.row_mat(i_m, i_n2);
l1 = (updated_llr_vec(n2) - lasteta(i_m, n2));
l1 = min(l1, 20);
l1 = max(l1, -20);
if min_sum
for i_n2 = 1 : obj.row_weight_vec(i_m)
if i_n1 == i_n2
continue;
end
n2 = obj.row_mat(i_m, i_n2);
l1 = (updated_llr_vec(n2) - lasteta(i_m, n2));
l1 = min(l1, 20);
l1 = max(l1, -20);
pr = sign(pr) * sign(l1) * min(abs(l1), abs(pr));
else
pr = pr * tanh(l1/2);
end
eta(i_m, n1) = pr;
end
if min_sum
eta(i_m, n1) = pr;
else
eta(i_m, n1) = 2 * atanh(pr);
end

lasteta = eta;

for i_n = 1 : obj.N
updated_llr_vec(i_n) = input_llr_vec(i_n);
for i_m = 1 : obj.col_weight_vec(i_n)
m = obj.col_mat(i_n, i_m);
updated_llr_vec(i_n) = updated_llr_vec(i_n) + eta(m,i_n);
end

end
end

lasteta = eta;

for i_n = 1 : obj.N
updated_llr_vec(i_n) = input_llr_vec(i_n);
for i_m = 1 : obj.col_weight_vec(i_n)
m = obj.col_mat(i_n, i_m);
updated_llr_vec(i_n) = updated_llr_vec(i_n) + eta(m,i_n);
end
end

decoded_codeword = (updated_llr_vec < 0);
if obj.check_codeword(decoded_codeword)
return;
else
error_vec(iter) = 1;
end

decoded_codeword = (updated_llr_vec < 0);
if obj.check_codeword(decoded_codeword)
error_vec = 0;
return;
else
error_vec = 1;
end
end
else % my decoding codes
tH = sparse(obj.H);
lasteta_s = tH*0;
saturatedBound = 20;
n = obj.N;
m = obj.M;
for iter = 1 : max_iter
H_blf = tH * spdiags(updated_llr_vec, 0, n, n) - lasteta_s;
H_blf = min(max(H_blf, -saturatedBound), saturatedBound); % bound above and below
r = tanh(H_blf/2);
r_prod = full(real(exp(sum(spfun('log',r),2)))); % magic step, use log to compute product
pr = spdiags(r_prod, 0, m, m) * real(spfun('exp', -spfun('log',r))); % crying... I could not find a proper function to compute reciprocal
eta = 2 * atanh(pr);
lasteta_s = eta;
updated_llr_vec = input_llr_vec + sum(eta).';

decoded_codeword = (updated_llr_vec < 0);
if ~nnz(obj.H * decoded_codeword)
error_vec = 0;
return;
else
error_vec = 1;
end
end
end

end

function [b] = check_codeword(obj, x)
b = 1;
for i_check = 1 : obj.M
c = 0;
for i_n = 1 : obj.row_weight_vec(i_check)
c = c + x(obj.row_mat(i_check, i_n));
end
if mod(c, 2) == 1
b = 0;
break;
end
end
b = 1;
for i_check = 1 : obj.M
c = 0;
for i_n = 1 : obj.row_weight_vec(i_check)
c = c + x(obj.row_mat(i_check, i_n));
end
if mod(c, 2) == 1
b = 0;
break;
end
end
end


function lifted_ldpc(obj, baseH, Z)
circ_mat_array = {};
Expand Down Expand Up @@ -205,7 +236,7 @@ function lifted_ldpc(obj, baseH, Z)


function load_wifi_ldpc(obj, block_length, rate)

H_1296_1_2 = [40 -1 -1 -1 22 -1 49 23 43 -1 -1 -1 1 0 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1;
50 1 -1 -1 48 35 -1 -1 13 -1 30 -1 -1 0 0 -1 -1 -1 -1 -1 -1 -1 -1 -1;
39 50 -1 -1 4 -1 2 -1 -1 -1 -1 49 -1 -1 0 0 -1 -1 -1 -1 -1 -1 -1 -1;
Expand Down Expand Up @@ -272,7 +303,7 @@ function load_wifi_ldpc(obj, block_length, rate)


H_1944_5_6 = ...
[ 13 48 80 66 4 74 7 30 76 52 37 60 -1 49 73 31 74 73 23 -1 1 0 -1 -1;
[ 13 48 80 66 4 74 7 30 76 52 37 60 -1 49 73 31 74 73 23 -1 1 0 -1 -1;
69 63 74 56 64 77 57 65 6 16 51 -1 64 -1 68 9 48 62 54 27 -1 0 0 -1;
51 15 0 80 24 25 42 54 44 71 71 9 67 35 -1 58 -1 29 -1 53 0 -1 0 0;
16 29 36 41 44 56 59 37 50 24 -1 65 4 65 52 -1 4 -1 73 52 1 -1 -1 0];
Expand Down Expand Up @@ -338,9 +369,9 @@ function load_wifi_ldpc(obj, block_length, rate)
end

baseH = eval(H_var);

obj.lifted_ldpc(baseH, obj.Z);

end

end
Expand Down
2 changes: 1 addition & 1 deletion LdpcM/run_WiFi_LDPC.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
max_runs = 100;
max_decode_iterations = 20;
ldpc_code = LDPCCode(0, 0);
min_sum = 1;
min_sum = 0;
n_0 = 1/2;


Expand Down
16 changes: 15 additions & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ The results are using LdpcC code, and are based on 50K runs.
Runtime performance C and MATLAB
-----

The run time comparison is as follows (run on a single macbook pro 2015):
The run time comparison is as follows (run on a single macbook pro 2015)[This part may need changing]:
<table>
<caption> Comparison of number of runs per second </caption>
<tr align="center">
Expand Down Expand Up @@ -65,6 +65,20 @@ The run time comparison is as follows (run on a single macbook pro 2015):

The above numbers are based on 20 iterations of the BP decoder *without* early termination. So, typically, one should see much better performance.

For the improved matlab version EnDec BP algorithm(use matrix multiplication rather than for loop, there is a tremendous performance improvement)
<table>
<caption> Comparison of number of running time for 100 iterations </caption>
<tr align="center">
<th>Org Codes(sec) </th>
<th>New Codes(sec)</th>
<th>Speedup org/new</th>
</tr>
<tr align="center">
<td>236.7094</td>
<td>30.8783</td>
<td>x7.67</td>
</tr>
</table>
Code Interface
------

Expand Down