feComposite arithmetic is broken with neg values

[v1nce]

I try to make the following filter

|a - b|

so I use
max(a,b) <=> lighter(a,b) <= ok
min(a,b) <=> darker(a,b) <= ok

|a,b| should be equal to max - min

So i joined max and min using a composite|arithmetic with k2=1 and k3=-1
I used a black to white gradient over grey
I expected to get a grey/black/grey gradient and I only get a full grey

I tried some other combination to see if I get what I'd expected
(x >0)
x 0 0 0 = ok
0 x 0 0 = ok
0 0 x 0 = ok
0 0 0 x = ok

eveytime a negative value comes into the game then things go wrong
0 1 -1 0 = ko
0 0 -1 1 = ko
0 -1 0 1 = ko

Inkscape 0.48.3.1 r 9886
Ubuntu 12.04

[su_v (suv-lp)]

Reproduced with Inkscape 0.48.3.1 on OS X 10.7.4
Not fully reproduced with 0.48+devel r11643 (see attached bitmap export)

I have yet to find an SVG renderer which would render the filter effect ("Max(a,b) - Min(a,b)") as shown in the "What's expected" sample in the file:
- Firefox 14, Chromium 23, Safari 6 fail completely to render any of the used filter effects.
- Opera 12 does better, but also fails to render the bottom one ("Max(a,b) - Min(a,b)").
- Squiggle (Batik 1.7, 1.8pre) does better, but differs from the claimed expected result, and from what Inkscape trunk renders.

-> needs further investigation as to what the correct rendering of the custom filter effect in SVG would look like.

[su_v (suv-lp)]

SVG rasterized with Batik 1.8pre+r1374667

[v1nce (vincent-pennec)]

I explain what I'd expected

for left value:
 max(0,0.5) = 0.5 (ok)
 min(0,0.5) = 0 (ok)
 k2* max + k3 * min = 1 * max + (-1) * min = max - min = 0.5 - 0 = 0.5 = grey

for middle value:
 max(0.5 , 0.5) = 0.5 (ok)
 min(0.5 , 0.5) = 0.5 (ok)
 k2* max + k3 * min = 1 * max + (-1) * min = max - min = 0.5 - 0.5 = 0 = black

for right value:
 max(1 , 0.5) = 1 (ok)
 min(1 , 0.5) = 0.5 (ok)
 k2* max + k3 * min = 1 * max + (-1) * min = max - min = 1 - 0.5 = 0.5 = grey

If I didn't miss something about premul or clamping, I think this is right

[Jaspervdg (jaspervdg)]

Both images have the same alpha channel. So max-min = 0 for the alpha channel, making the result completely transparent. I recommend using feComponentTransfer to invert the "min" image (which is effectively 1-min), as it can be applied per-channel, and then using something like k2=1, k3=1, k4=-1 (max + (1-min) -1). The alpha channel then remains 1 (1+1-1).

[v1nce (vincent-pennec)]

This worked with adding the following FeColorMatrix before feComposite
-1 0 0 0 1
0 -1 0 0 1
0 0 -1 0 1
0 0 0 0 1

what I don't understand is why this doesn't work when I keep the feComposite and add the following FeColorMatrix after
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 0 1