Hello guys,
for my thesis I have to optimally scale an IIR filter.
Changing the design of the filter is not my purpose.
I have read the Randy Yates papers about fixed-point arithmetic and
Fixed-point FIR implementations (thank you so much Randy), but I have a
question.
The filter has these coefficients:
b[0] = 1,0;
b[1] = 1,9059465;
b[2] = 0,9114024;
y[i] = b[0]x[i] + b[1]y[i-1] + b[2]y[i-2]
Input values comes from the output of a FIR filter and are stored on 32
bit variables (14 bit added precision). I want to store coefficients on 16
bit and the output and the ac***ulator of the filter on 32 bit.
I have to scale down FIR output before passing them into IIR filter , but
I have to choose a trade-off between coefficients and input dimension in
order to avoid overflow.
The more precision I give to IIR input, the less I will give to
coefficients.
Do you know how can I simply calculate this trade-off?
|
