[email protected] wrote:

> Hello all,

>

> Can anyone point me to a good general purpose paper about selecting

> appropriate bit-widths for a fixed point implementation of a signal

> processing algorithm? I've looked around and haven't found anything

> that describes a general methodology to use. I have implemented a

> design and it isn't performing as well as I would like. I suspect it

> has to do with some poor selection of bit slicing when it comes to

> multipiers and accumulators.
I don't know any papers. Personally, I do an implementation of an

algorithm in Matlab, first with floating point numbers, and then with

integers. I can then make quick changes and graph the two

implementations to compare the results. That allows me to determine how

many bits I need for various portions of the algorithm to get acceptable

results.

By the way, I tried the Matlab fixed point package, but found that the

performance was too poor to use on processing even a modest sized data

set, so I stick with using ordinary integers. Typically I stick a

comment at the end of most lines indicating the scaling of that

particular parameter.