Convolution Matrix Techniques / problem

I have been writing filter factory-type filters for dozens of years. I'm now looking to write a filterforge image convolution filter to do some simple smoothing, using a 3x3 kernel.

The convolution kernel I'm looking at is:
2 2 2
2 1 2
2 2 2
with the divisor = 17

I've looked at others' techniques for doing this. A common way is to use 9 offset components fed to 9 multiplier components fed to 9 add components.

I'm hoping my lua script would be simpler, faster, and more flexible. I think it works, in general. The script is:

Code

function prepare()    -- This function is called once,    -- before the rendering begins.    -- Use it for querying non-mapped    -- inputs and precalculation.    dx=1/OUTPUT_WIDTH    dy=1/OUTPUT_HEIGHT end; function get_sample(x, y)    -- Image generation code goes here.    local rul, gul, bul, aul = get_sample_map(x-dx, y-dy, SOURCE)       local ruc, guc, buc, auc = get_sample_map(x, y-dy, SOURCE)       local rur, gur, bur, aur = get_sample_map(x+dx, y-dy, SOURCE)       local rml, gml, bml, aml = get_sample_map(x-dx, y, SOURCE)       local rcent, gcent, bcent, acent = get_sample_map(x, y, SOURCE)       local rmr, gmr, bmr, amr = get_sample_map(x+dx, y, SOURCE)       local rll, gll, bll, all = get_sample_map(x-dx, y+dy, SOURCE)       local rlc, glc, blc, alc = get_sample_map(x, y+dy, SOURCE)       local rlr, glr, blr, alr = get_sample_map(x+dx, y+dy, SOURCE)       local rout = (rcent+2*(rul+ruc+rur+rml+rmr+rll+rlc+rlr))/17    local gout = (gcent+2*(gul+guc+gur+gml+gmr+gll+glc+glr))/17    local bout = (bcent+2*(bul+buc+bur+bml+bmr+bll+blc+blr))/17    local aout = 1    return rout, gout, bout, aout end;

But when I use this map filter, the resulting pixels do not exactly match the results from my Paint-Shop-Pro User Defined Filter, that I've been using forever.

I've turned anti-aliasing off and disabled automatic gamma-correction, but this had no effect.
Is there an HDR setting somewhere that is affecting things?

Thanks in advance!

I used the minimum of image width and height to calculate dx and dy. The filter performed *almost* perfectly.

The resulting pixel channels differ from the results from my Paint-Shop-Pro User Defined Filter by 0, 1, or -1.

What rounding scheme do you suggest to remedy this?

Here's a non-lua version hard coded with your values. You can use it for comparison, this one should be mathematically accurate. I'm going to refine it with nine inputs so you can use a different matrix and release it as a snippet. FF13 only.

3x3 Convolution Matrix 2.ffxml