Render to Vertex Buffer

I think what you might be looking for is Vertex Texture Fetch (VTF). It allows a vertex shader to do texture lookups. However, current ATI cards don't support this feature (which is why you'd use R2VB on ATI cards to simulate VTF).

Quote:

Original post by Sc4Freak I think what you might be looking for is Vertex Texture Fetch (VTF). It allows a vertex shader to do texture lookups. However, current ATI cards don't support this feature (which is why you'd use R2VB on ATI cards to simulate VTF).

Yeah I have been thinking about that. But in this case I need to do 3 Tex2Dlod for every instance, send these values to the pixel shader and also do several operations on the data to obtain a matrix which I transform the mesh with. With VTF I have to do this process once for every vertex instead of now when I only do it once per instance, am I right about this? So I don't think I gain from it.

I think there was a mention of NVIDIA drivers supporting R2BV. On Vista, maybe? I hope I didn't just imagine this. Maybe someone with a better memory can comment.

One thing that made me think that it was possible to do R2VB on nVidia was these 2 publications.

Publication 1

Publication 2

Maybe they are just talking about VTF?

[Edited by - 51mon on January 29, 2007 4:23:22 AM]

In the GeoClipMapping article in GPU Gems 2 there also was some mention of 'using Render-to-Vertex-Array'. It's an NVidia book and I also recall seeing that technique mentioned in other NVidia sources, so it might indicate they obscurely support it. On the other hand, this specific article was written by MS folks, so they just might be pointing out alternatives.

Some form of R2VB will be supported by all Vista (SM4) hardware though, at least that's what I've been told when I asked a similar question. D3D10/SM4 supports stream rerouting to and from various resources, so you can set up your shader pipeline to do something equivalent to both VTF and R2VB. I hope the R2VB-equivalent is feasible on SM4 hardware; from what I've heard VTF is quite slow, plus VTF is applied every frame while R2VB offers more control.

From what I've seen, I don't think there are any easy ways of doing this with D3D9, at least not without falling to device-dependent features...

You might want to look into the GPGPU side of things for this sort of job - they tend to "hack" general purpose computations on top of the D3D/OGL API's (until CUDA/CTM appeared) and might have some nifty tricks for uploading arbitrary data and pulling it back again...

Quote:

Original post by remigius In the GeoClipMapping article in GPU Gems 2 there also was some mention of 'using Render-to-Vertex-Array'.

I could be wrong, but I think the 'Render to vertex array' feature is the OpenGL equivalent of VTF/R2VB...

Quote:

Original post by remigius Some form of R2VB will be supported by all Vista (SM4) hardware though, at least that's what I've been told when I asked a similar question. D3D10/SM4 supports stream rerouting to and from various resources, so you can set up your shader pipeline to do something equivalent to both VTF and R2VB.

VTF is supported as a basic feature in D3D10 - simply because you can bind a texture as an input to any of the three programmable stages.

Strem re-routing is via the GS with SO but it has its limitations - iirc it only allows for 1024 4xFP32 values to be written out, which may not be enough for complex models. The GS can be used with RT arrays to implement a form of scatter-writing which could get you close to R2VB - but I've not tried that or R2VB (don't use ATI hw) so I could be wrong.

hth
Jack