Jos, sorry, didn't saw your post. They say in docs and presentations that linear depth is used, as they use it for bunch of othre effects including one that reconstuct world position of pixel from depth (you need linear depth to do it). Maybe specifically for SSAO it's rescaled.

Yes, that is true, they use it for multiple things. But, lets say you have a scene with SSAO. This is scene 1. Now, lets say you have an exact same scene that's twice as big in all directions, scene 2. Lets say scene 2 is twice as far away from the camera. It should look exactly the same as scene 1. But, the depth range is twice as big. So, to get the exact same SSAO result, the log of the depth should be used.

As and extra advantage, storing the log depth in a 16 bit buffer gives you all the range you'll need. I'm not sure what kind of buffer you use now, but with log depth you could first store the linear depth in a single 32 bit channel texture. Then you can sample down, create a small poison disk in RGBA, take the log and store in a 16 bit 4 channel texture.

I'm rescaling depth influence for different distance. Shadows fade away because sampling radius has a maximum boundary, this is needed to keep good performance. So, when you get close to an object small cavities become too big for sampling radius to capture them. I can rescale depth in such a way that shadows near camera won't be white, but they will turn into plain gray color (half-occlusion).

Of course, I forgot about that.