Re: RFC: Generic Fixed-Point IIR Filter

[Brian Drummond <brian@shapes.demon.co.uk>]

On Thu, 26 Mar 2015 09:25:25 -0700, Patrick Noffke wrote:

> I would greatly appreciate some feedback/help on implementing a generic
> fixed-point IIR filter package.
...
> 
>  - If I try to use Fixed_Type for X_Array and Y_Array, then I do in fact
>  get a constraint error (due to range overflow for an intermediate
>  calculation).  Is there a better way to do a calculation for a
>  fixed-point type when you expect the result to be within the specified
>  range of Fixed_Type, but intermediate calculations may not?

I wrote (adapted) a generic SVD (matrix operation) to instantiate with 
either fixed or floating point. The original was floating point : the 
intent was to investigate if different fixed point formats could be made 
stable, before porting to VHDL and implementing in hardware. It quickly 
answered the question : for matrix inversion, no way!

Intermediate addition operations seemed OK : intermediate operations 
involving multiplication (or division) had to be rewritten using 
intermediate variables. 

A moment's thought showed this made sense : you may want the intermediate 
results stored at a different resolution : this is also usually the case 
with IIR filters.

In other words I had to break up 
   f := a * b * c;
into something like
   temp := b * c; 
   f := a * temp;
where temp may have been a higher resolution type, which allowed (i.e. 
forced!) me to explicitly control the numerical properties. This is the 
point Randy made, 4.5.5(19.1/2).

>  - is there another way to
>  define the constants that will "convert" them to a multiple of small?

I would probably use a conversion function on real (float) constants.

>  - I realize my range and precision for Filter_Value_Type are not
>  necessarily ideal (the range depends on the filter coefficients, the
>  filter architecture, as well as the filter order).  Do you know of any
>  references or have suggestions for how to analyze a filter to determine
>  the ideal range and precision?

Signal processing texts aimed at hardware engineers should cover this : 
"Synthesis and optimisation of DSP Algorithms" (Constantinides,Cheung,Luk) 
comes to mind. But with a generic, you can instantiate it with different 
fixed (and float) types, measure the quality of each, and choose.

-- Brian