> So I would like to implement a running average algorithm, which gives the
> average of the last eight values. It should add up the last eight values
> and divide then by 8. but my values can be everything between 256 and 0.
> If I divide every value before adding them up I get very arbitrary
> averages from low input numbers due to rounding errors. Is there a
> possibility to do 16 bit integer arithmetric with two data registers.

My personal preferred approach to this problem is to average pairs,
then average pairs of averages, etc.  If I had eight numbers num0 to
num7, then I'd do something like this:

Ave2:   ds      1
Ave4:   ds      1
Ave8:   ds      1
Temp:   ds      1
...
AveEmAll:
        movf    Num0,w
        addwf   Num1,w
        movwf   Ave2
        rrf     Ave2

        movf    Num2,w
        addwf   Num3,w
        movwf   Temp

        rrf     Temp,w
        addwf   Ave2,w
        movwf   Ave4
        rrf     Ave4

        movf    Num4,w
        addwf   Num5,w
        movwf   Ave2
        rrf     Ave2

        movf    Num6,w
        addwf   Num7,w
        movwf   Temp

        rrf     Temp,w
        addwf   Ave4
        movwf   Ave8
        rrf     Ave8

        ; 22 instructions total

Not quite as precise as doing full-precision adds, but a little faster if
the numbers may be in the range 128-255.  If the numbers are known to be
smaller (e.g. 128 or less) then it may make sense to add pairs of then and
then average those pairs.