Hey guys,
I'm currently trying to implement Variance Shadow Maps and I just can't figure out how to filter the cubemaps for the omnidirectional lights. For the spotlights I just apply a gaussian blur, but I can' find a way to do this with cubemaps. How do other games tend to solve this problem?
A cubemap is just 6 2D textures, one for each face of the cube, so if you are able to filter a regular 2D texture then filtering a cubemap would just be an extension of that. The simplest approach would be
1. Create additional cubemap to hold result.
2. Bind each face of the cubemap created in 1, sample the sample face in the shadow cubemap and do your filtering.
3. Write result into the output texture ( current face of the cubemap created in 1. )
4. Repeat for the other 5 faces.
Thank you for your answer!
But wouldn't this approche have potential artifacts at the edges of each face? I mean when a shadow is spread across two faces, shouldn't the blur take some texels from both faces?
Yes, you will potentially get artifacts along the seams of the cube map if you blur them the way cgrant suggests. I think a lot of games solve this problem by just not having shadowmap support for point lights (or if they do, they might use dual paraboloid mapping or something to minimize both the number of seams that have to be hidden and the memory footprint of the shadowmap).
If you really want to use cube maps and filter them correctly you'll have to filter one face at a time, and for each texel you will have to reconstruct the normal that corresponds to that texel, then perform the gaussian blur by sampling the cubemap in the neighborhood around that normal on the unit sphere.
They use a single 2D texture with all 6 faces on it.I think a lot of games solve this problem by just not having shadowmap support for point lights
Filtering a cubemap texture is extremely slow.I just can't figure out how to filter the cubemaps for the omnidirectional lights.
On all modern hardware a cubemap is just a texture array with NumSlices = 6, so there's not really much difference (from a hardware POV) between a cubemap and an 'atlased' texture where 6 faces are packed into one long horizontal strip. The main difference is that if you want to you have access to an easy direction->UV conversion in shader, or alternatively if you just want to treat it as an array you can easily address each face using an array index. In D3D11 which behavior you get is dependent on how you set up your shader resource view, and what type is declared in your shader. Personally I would probably go with a cubemap to get the nice cubemap behavior when sampling the filtered shadow map, especially since in D3D11 hardware filtering works correctly across cubemap seams.
