Looks like this is already a solved problem, but here's another thread on it:

Codesampler thread;

That one shows how you can do a limited amount of additive blends as well. I'm still stuck on filtering though.

for your information, I am using the code given in that CodeSampler thread:

encode (fDist is the normalized distance):

	const float4 bitSh	= float4(   256*256*256, 256*256,   256,         1);	const float4 bitMsk = float4(   0,      1.0/256.0,    1.0/256.0,    1.0/256.0);	float4 comp;	comp	= fDist * bitSh;	comp	= frac(comp);	comp	-= comp.xxyz * bitMsk;	return comp;

decode (vec is the rgba encoded value):

	const float4 bitShifts = float4(1.0/(256.0*256.0*256.0), 1.0/(256.0*256.0), 1.0/256.0, 1);	return dot(vec.xyzw , bitShifts);

The code given by Ysaneya introduces some artefacts, certainly cause by the missing "comp -= comp.xxyz * bitMsk;" (which is actually the difference between Ysaneya's code and the one given in the CodeSampler thread).

Right now the result is very good, it allows me to store the scene camera depth texture using a ARGB8 rendertarget, which allows using the same Z-buffer as the main surface (and thus writing the Z buffer during my camera-depth texture generation, allowing early Z-cull during the main scene rendering pass)!

Well done guys, thanks a lot !

thanks for the feedback.

It is not a problem in my case, just have to divide my distance by (zfar + 1).

I guess in most of the cases a similar solution exists.