Ditto on what AndyTX said. I've been playing around with deferred rendering for the past month or two, and I have yet to come across an effective way of implementing MSAA (MRT or not). Having just got my 8800 recently, I haven't played around with retrieving the pre-resolved textures but I'd suspect the performance would be closer to SSAA than it would be to MSAA.

Quote:Original post by Cypher19
Besides, the driver/GPU can probably handle the AA'd buffer in a smarter manner, in terms of performance, than just falling over and having to deal with a really large buffer.

Yes of course, your comments are correct. However, Andy's point was about the interaction between user supplied pixel shaders and the AA buffer. Since there seem to be no (easy ?) way of knowing whether some property within an AA sample cell is constant, you cannot optimize your shaders the same way MSAA optimizes over SSAA. In other words, you have to execute your shader once for every subpixel in the AA buffer, which would amount to SSAA from the performance penalty of shader invocation.

Quote:Original post by AndyTX

Quote:Original post by Yann L
Oh, and could ATI *PLEASE* fix the two trillion depth buffer related bugs in their current OpenGL driver already ? Working around their bugs is a real PITA.

Haha, you ran into this too?

Oh God, yes... Please don't remind me... I just ranted about this over in the OpenGL forum.

Quote:Original post by AndyTX
Sounds really nice - any chance we can see some screenshots or get some info on the techniques that you're using?

Oh, it's not that complex actually.

It's basically a form of advanced ambient occlusion field technique mixed with PRT. In ShaderX4, there is an article by Kontkanen/Laine about cubemap compressed ambient occlusion fields for contact shadows. Our system uses a similar principle, but encodes the data in a different way.

For each reference object, we're using one set of 3D textures and one cubemap set. The 3D textures spans the objects bounding box, and the cubemaps close the box like an exterior skin. Each 3D voxel contains not the spherical cap size and direction (as in the Shader X4 approach), but an entire SH coefficient set. This way, we can describe the transfer function at each inner point of the object towards the outside, from every possible direction, taking into account all forms of self shadowing and self reflections.

The outer cubemap encodes compressed polynomials, much like Kontkanen/Laine did, except we're storing outwards going SH coefficients again instead of a simple spherical cap.

When rendering, we project the 3D texture/cubemap set from every object onto the environment, compute the light transfer function at each pixel. Since we store an entire SH set at the inner and outer regions of each object, we can easily extend the ambient shadows to include direct and indirect lighting effects. Means that we don't only get near surface contact shadows through the 3D texture, but full soft shadows through the PRT polynomials on the cubemap set. See it as a kind of "inverted shadowmap" approach, but instead of storing depth values, we store light transfer polynomials.

The main idea is to get realistic shadows from many large dynamic objects in a complex GI environment (with area lights, sun and skylight), used in photorealistic architectural visualization.