Would you like to be notified by email when mnentwig publishes a new blog?
A floating point model for a CIC decimator, including the frequency response.
A CIC filter relies on a peculiarity of its fixed-point implementation: Normal operation involves repeated internal overflows that have no effect to the output signal, as they cancel in the following stage.
One way to put it intuitively is that only the speed (and rate of change) of every little "wheel" in the clockworks carries information, but its absolute position is arbitrary.
Modeling a CIC filter without use of bit-accurate numbers is not completely straightforward. Here, I'll show some "instant" solution for the decimating variant. Just add water and stir.
The good. The bad. The ugly.
Let's start with "ugly".
The "time domain" implementation (complete code below) is what I get when I close my eyes to the numerical overflow problem and simply write the CIC difference equations with floats.
Now double precision arithmetics are quite forgiving, and it works well enough.
It should do fine for a classroom demo (and please don't use it in the next Mars mission).
Then, the "frequency domain" implementation:
This is only "bad", as taking the FFT of the whole signal all at once might be inconvenient.
Finally, what I'd consider a "good" solution:
The approach is accurate, as despite its internal recursion, the impulse response of a CIC filter has a finite length (it is of FIR type, not IIR).
To apply the impulse resonse b to an input sequence x, use y = filter(b, 1, x);
The plot shows that all three methods give exactly the same impulse response (accurate to ~10-15). Decimating discards three samples out of four, leaving only red-trace samples that coincide with the other two traces.
The frequency response is evaluated on the higher (input) rate. Therefore, it is straightforward to investigate the rejection on input frequencies that will cause aliasing, once decimated.
The example shows how to model a CIC decimator in floating point as a conventional FIR filter by sampling the finite-length impulse response.
% CIC decimator example
% - time domain implementation
% - frequency domain model via z-domain transfer function
% R = rate change factor
% N = nStages
% M = differential delay
% reference:  http://www.altera.com/literature/an/an455.pdf
testvec = zeros(1, 43);
R = 4;
testvec(1) = 1;
N = 3;
M = 2;
a = CICdec_timeDomainModel(testvec, N, M, R);
[b, H] = CICdec_freqDomainModel(testvec, N, M, R, false);
[c, H] = CICdec_freqDomainModel(testvec, N, M, R, true);
figure(1); grid on; hold on;
plot(1:R:R*numel(a), a, 'bx');
plot(1:R:R*numel(c), c, 'k+');
title('CIC decimator impulse response');
legend('time domain', 'freq. domain', 'freq. domain decimated');
H = fft([ifft(H), zeros(1, 1000)]); % zero-padding
figure(2); clf(); grid on; hold on;
plot(linspace(-0.5, 0.5, numel(H)), fftshift(20*log10(abs(H) + 1e-15)), 'b');
function [vec, H] = CICdec_freqDomainModel(vec, N, M, R, doDecim)
flag = isreal(vec);
% evaluate frequency response (z-domain transfer function)
n = numel(vec);
zInv = exp(-2i*pi*(0:(n-1))/n);
b = ones(1, R * M);
H = polyval(b, zInv) .^ N;
H = H / H(1);
% apply frequency response
vec = ifft(fft(vec) .* H);
vec = vec(1:R:end);
% don't let FFT roundoff error turn real signal into complex
vec = real(vec);
function vec = CICdec_timeDomainModel(vec, N, M, R)
nLeadIn = M * N * R;
nLeadOut = nLeadIn / R;
ix = mod(-nLeadIn:-1, numel(vec)) + 1;
% prepend end of cyclic signal
vec = [vec(ix) vec];
for ix = 1:N
vec = cumsum(vec);
vec = vec(1:R:end);
for ix = 1:N
vec = vec - circshift(vec, [0, M]);
% remove the added length
vec = vec(nLeadOut+1:end);
% scale gain
gain = (R * M) .^ N;
vec = vec / gain;
Add a Comment