Shadow Volumes on GPU

[Alexander K]

Did somebody Shadow volumes full_on_GPU? I think, that its good idea for underground FPS. Like Prey )
Need help: shaders, links, similar topics...

Sorry my english))

[voxel]

Did somebody Shadow volumes full_on_GPU? I think, that its good idea for underground FPS. Like Prey )
Need help: shaders, links, similar topics...

Sorry my english))

I was under the impression that shadow volumes were almost solely on the GPU (after the models have degenerate quads on the edges). The extrusion is done in the vertex shader and the rendering is stencil buffer / pixel shaders. Sorry, it's been 5-6 years since I've deal with them. What do people do on the CPU with shadow volumes these days?

[sinbad]

Calculating the light facing in order to calculate the silhouette.

There are ways to put everything on the GPU but it requires a hell of a lot of extra geometry and just isn't worth it. Texture shadows are the future for GPU driven shadows.

[voxel]

Calculating the light facing in order to calculate the silhouette.

Crap forgot about "finding which edges" to extrude.

There are ways to put everything on the GPU but it requires a hell of a lot of extra geometry and just isn't worth it. Texture shadows are the future for GPU driven shadows.

Are you talking about shadow maps? Up until I read about Variance Shadow Maps I found SM fraught with aliasing (too dependant on resolution and light distance) and precision issues (z and perspective projection). I want something like Deep Shadow Maps implemented on the GPU - i.e each pixel is a "linear function." I also had great hopes for Pre-computed radiance transfer until I actually tried it - you need dense meshes and a long pre-computation time.

I kinda like the sharp / discontinuous shadows that shadow volumes gives

[PolyVox]

I also had great hopes for Pre-computed radiance transfer until I actually tried it - you need dense meshes and a long pre-computation time.

What is Pre-computed radiance transfer? Are they dynamic? ('precomputed' implies not - but you mention them alongside shadow maps/volumes...). And I have dense mehes so they sound interesting.

[voxel]

I also had great hopes for Pre-computed radiance transfer until I actually tried it - you need dense meshes and a long pre-computation time.

What is Pre-computed radiance transfer? Are they dynamic? ('precomputed' implies not - but you mention them alongside shadow maps/volumes...). And I have dense mehes so they sound interesting.

PRT is an extension of the "spherical harmonics" rendering research.

In SH - lighting (i.e a light dome) is converted into 3D function kinda like how Fast Fourier Transform converts a wave signal into co-efficients for 2D sin + cosine. This allows you to do diffuse lighting using HDR domes (better than directional lighting) without needing to an environment/image map to sample. You don't even need a model in this process. The HDR light dome gets converted to a 3D function (3 x 9 co-efficients) through a offline process which then can be plugged into an vertex shader. You passe a vertex normal into this 3D function / Vertex shader and get a colour back.

PRT uses multiple SH representations - one for the light dome, one for shadowing / self-shadowing, and another for glossy / interreflections (i.e red shirt bleeds onto white face). To get the best results you need dense meshes because PRT/SH is vertex-based and PRT builds a virtual light-dome at every vertex (so 3D function exists at every vertex) whereas SH is model-independent (one 3D function for the entire world).
So if you have 100k vertices you'll have 100k 3D functions also (27 co-efficients). The data size is brutal.

PRT is light-independent, but if the model deforms everything breaks. You can get some beautiful cars / spaceships rendered using this method. Changing the lighting on the fly costs zero as all self-shading + glossy reflections are "pre-computed" (the light paths are, but not the values).

It doesn't have much use in games to be honest. I think it makes for great rendering in main menus / UI selection screens or in single room games. Normal SH rendering, I believe, is pretty common - it's super cheap and easy. I'm sure I've missed out on the advances in the past three years so my knowledge might be out of date.

[PolyVox]

Well thanks for the info It sounds interesting but if it breaks when the model is deformed then I can't use it in my project. Plus I'm gonna be tight on memory as it is... It always good to know about new techniques though!