Krypt0n

probably filtering from outside the texture area of the triabgle. you need to create a border around the texutre areas, some kind of dilation filter

so you're doing world space SSAO, but your normals are view space?

 

try to set qp here

    float dp = max(dot(normal, normalize(r - position)), 0.0f);

to 1.f and although it's ignoring the normal, it should look somehow stable, then you know, you have to either use worldspace normals, or to transform the world space vector between the position of the surface and the 'hit' position on the ssao sample into view space.

it doesn't really make sense to render "hundreds of millions" on one to 20 million pixel. before you start optimizing the brute force part, you might evaluate using a better culling. 

1. PVS

2. occlusion culling

3. precalculate face culling (most distance object can be made out of 1 or 2 face (2 or 4 triangles).

4. you are most likely fillrate bound I'd think -> you could precalculate per building at what distance you see how much of it, the lower parts of most buildings wont be visible, this can be easily adjusted in shader.

5. try to generate an impostor cubemap of distance objects every 10frames or something, like a skybox, it will even reduce pixel-aliasing.

I could imagin the drawing is done in linear space, while yours and gimps blending works in srgb space.

 

another point might be the internal precision, especially when you blend a lot of layers on top, the quality loss due to rounding can be quite noticeable, having 16bit internal precision can be quite beneficial.

 

(newer CPUs have instructions to convert float16 to float32  aka half to float, this could speed your rendering up).

output the depth into the alpha channel, it's not very accurate, but enough for your filter.

 

as an optimization for your current 3 phases, swap 1 and 2. it's beneficial, especially on mobile devices, to render one screen in one go, instead of switching from it to a temporary and back, as your GPU needs to store and restore the framebuffer into internal caches. doing it in one go means 67% traffic saving (just one write, instead of write, read, write).