function [f1_saved,f2_saved,info]=designHBF(fp,delta,debug)
%function [f1,f2,info]=designHBF(fp=0.2,delta=1e-5,debug=0)
%Design a half-band filter which can be realized without general multipliers.
%The filter is a composition of a prototype and sub- filter.
%Input
% fp The normalized cutoff frequency of the filter. Due to the
% symmetry imposed by a HBF, the stopband begins at 0.5-fp.
% delta The absolute value of the deviation of the frequency response from
% the ideal values of 1 in the passband and 0 in the stopband.
%
%Output
% f1,f2 The coefficients of the prototype and sub-filters
% and their canonical-signed digit (csd) representation.
% info A vector containing the following data (only set when debug=1):
% complexity The number of additions per output sample.
% n1,n2 The length of the f1 and f2 vectors.
% sbr The achieved stob-band attenuation (dB).
% phi The scaling factor for the F2 filter.
% To Do: Clean up the code a bit more, esp. wrt the use of the struct. arrays.
% Use the phi variable to cut down on the number of adders in F2.
% Apply a simulated annealing/genetic optimization alg instead
% of the ad hoc one I have now.
%Handle the input arguments
parameters = ['fp ';'delta';'debug'];
defaults = [ 0.2 1e-5 0];
for i=1:length(defaults)
if i>nargin
eval([parameters(i,:) '=defaults(i);'])
elseif eval(['any(isnan(' parameters(i,:) ')) | isempty(' parameters(i,:) ')'])
eval([parameters(i,:) '=defaults(i);'])
end
end
%Try several different values for the fp1 parameter.
%The best values are usually around .04
%Surrender if 3 successive attempts yield progressively greater complexity.
lowest_complexity = Inf; prev_complexity = Inf;
for fp1 = [.03 .035 .025 .040 .020 .045 .015 .05]
failed = 0;
[f1 zetap phi] = designF1( delta, fp1 );
if zetap == 1 % designF1 failed
failed = 1;
if debug
fprintf(2,'designF1 failed at fp1=%f\n',fp1);
end
end
if ~failed
f2 = designF2( fp, zetap, phi );
n1 = length(f1); n2 = length(f2);
if n2 == 0 % designF2 failed
failed = 1;
if debug
fprintf(2,'designF2 failed when zetap=%f, phi=%f\n',zetap,phi);
end
end
end
if ~failed
% complexity(+ performance) = the number of two-input adders (+ sbr)
complexity = size([f1.csd],2) + (2*n1-1)*(n2+size([f2.csd],2)-1);
if debug
msg = sprintf('%d adders: n1=%d, n2=%d, (fp1=%.2f, zetap=%.3f, phi=%4.2f)', ...
complexity, n1, n2, fp1, zetap, phi );
else
msg = '';
end
[fresp pbr sbr] = frespHBF([], f1, f2, phi, fp, msg);
if pbr <= delta & sbr <= delta
complexity = complexity + sbr;
if complexity < prev_complexity
worse = 0;
if complexity < lowest_complexity
lowest_complexity = complexity;
f1_saved = f1; f2_saved = f2;
phi_saved = phi;
if debug
fprintf( 1, '%s\n', msg )
end
end
else
worse = worse + 1;
if worse > 2
break;
end
end
prev_complexity = complexity;
end % if pbr <= delta
end
end % for fp1
if isinf(lowest_complexity)
fprintf(1,'%s: Unable to meet the design requirements.\n', mfilename);
elseif debug
complexity = floor(lowest_complexity);
msg = sprintf( 'Final Design: %d adders', complexity);
[junk pbr sbr] = frespHBF([], f1_saved, f2_saved, phi_saved, fp, msg);
n1 = length(f1_saved); n2 = length(f2_saved);
fprintf(1,'%s (%d,%d,%.0fdB)\n', msg,n1,n2,dbv(sbr));
info = [ complexity n1 n2 dbv(sbr) phi_saved ];
end
return
function [f1_saved,zetap,phi] = designF1(delta, fp1)
% [f1 zetap phi] = designF1(delta, fp1) Design the F1 sub-filter
% of a Saramaki halfband filter. This function is called by designHBF.m.
%
% f1 A structure array containing the F1 filter coefficents and
% Their CSD representation.
% phi The scaling factor for the F2 filter (imbedded in the f1 coeffs.)
passband = exp(4*pi*j*linspace(0,fp1));
ok = 0;
for n1 = 1:2:7 % Odd values only
if n1 == 1
h = [0.5 0.5];
else
h = firpm(2*n1-1,[0 4*fp1 1 1],[1 1 0 0]);
if ~(abs(sum(h)-1) < 1e-3 ) % remez bug! Use firls instead
h = firls(2*n1-1,[0 4*fp1 1-1e-6 1],[1 1 0 0]);
end
end
fresp = abs( polyval(h,passband) );
if max( abs(fresp-1) ) <= delta
ok = 1;
break
end
end
if ~ok
zetap = 1; % Use this as an indication that the function failed.
return
end
% Transform h(n) to a chebyshev polynomial f1(n)
% Sum(f1(i)*cos(w)^n)|i=1:n1 + Sum(h(n1+i))*cos(n*w))|i=1:n1, n = 2*i-1;
w = pi*rand(1,n1);
cos_w = cos(w);
A = zeros(n1,length(w));
B = zeros(1,n1);
for i = 1:n1
n = 2*i-1;
A(i,:) = cos_w .^ n;
B = B + h(n1+i)* cos(n*w);
end
f1 = B/A;
% Matlab Ver. 5 change:
phivecb = [];
% Optimize the quantized version of f1 to maximize the stopband width
% ( = acos(zetap) )
zetap = 1;
testPoints = [0 logspace(-2,0,128)] - 1;
for nsd = 3:8
f1a = f1'; f1b = f1'; % First try the unperturbed filter.
for phia = 1 ./ [1 f1]
phia = phia / 2^nextpow2(phia); % keep phi in (0.5,1]
% Try a bunch of coefficients in the current neighborhood,
% shrinking the neighborhood once 10 successive trial values show no
% improvement. If 2 successive shrinkages do no good, try a higher nsd.
count = 0;
nohelp = 0;
neighborhood = .05;
while neighborhood > 1e-5
phivec = phia .^ [1:2:2*n1-1]';
% Matlab Ver. 5 change:
if isempty(phivecb); phivecb = phivec; end
f1q = bquantize( f1a.*phivec, nsd );
F1 = evalF1( [f1q.val], testPoints, phia );
fi = find( abs(F1) > delta );
zeta = -testPoints( max( fi(1)-1, 1 ) );
%fprintf(2,'nsd=%d, nbhd= %f, count=%d, zeta = %f, phia=%f\n', ...
% nsd, neighborhood, count, zeta, phia );
if zeta < zetap
count = 0;
nohelp = 0;
zetap = zeta;
f1b = [f1q.val]';
f1_saved = f1q;
phi = phia;
phivecb = phivec;
else
count = count + 1;
end
if count > 10
count = 0;
neighborhood = neighborhood/2;
nohelp = nohelp +1;
if nohelp > 2
break;
end
end
f1a = f1b./phivecb + neighborhood*(rand(size(f1b))-0.5);
phia = phia + neighborhood*(rand(1,1)-0.5);
end
if zetap < 1 % Found a filter with adequate attn.
break;
end
end % for phia ...
if zetap < 1 % Found a filter with adequate attn.
break;
end
end
return
function f2 = designF2(fp,zetap,phi)
% f2 = designF2(fp,zetap,phi) Design the F2 sub-filter
% of a Saramaki halfband filter. This function is called by designHBF.m.
% subfilter design:
% 1 - delta2' < |F2/phi| < 1 for f in [0 fp];
% -1 < |F2/phi| < -1 + delta2' for f in [0.5-fp, 0.5];
% 1-delta2' = (1-delta2)/(1+delta2)
delta2 = (1-zetap)/(1+zetap);
%delta2p = 1 - (1-delta2)/(1+delta2);
% determine the minimum order required by the filter
passband = exp(j*linspace(0,4*pi*fp));
for nsub = 3:2:17
h2 = firpm(nsub,[0 4*fp 1 1], [1 1 0 0]);
mag = abs( polyval(h2,passband) );
if max(abs(mag-1)) < delta2;
break;
end
end
n2min = (nsub+1)/2;
% Search all n2,nsd pairs, in order of the product n2*(nsd+1)
% allowing fp to be a variable?
success = 0;
nsdmin = 3; nsdmax = 6;
for product = (nsdmin+1)*n2min:(nsdmax+1)*n2min
for nsd = nsdmin:nsdmax
n2 = product/(nsd+1);
if floor(n2) ~= n2 % Only take integer n2,nsd pairs
break
end
nsub = 2*n2-1;
% Could try a bunch of fp values
%fprintf(2,'designF2: Trying (n2,nsd2,fp)=(%2d,%2d,%6.4f)\n',n2,nsd,fp);
h2 = firpm(nsub,[0 4*fp 1 1], [1 1 0 0]);
h2 = h2/(phi*(1+delta2)); % Adjust the coefficients.
f2 = bquantize( h2(n2+1:nsub+1), nsd );
h2 = (1+delta2)*phi*[f2(n2:-1:1).val f2.val];
mag = abs( polyval(h2,passband) );
if max(abs(mag-1)) < delta2;
success =1;
break;
end
end
if success
break;
end
end
if ~success
f2 = [];
q2 = [];
end
return