Collision detection help

[NecroDOME]

You should use shaders. No idea how that works, but the got it working with havok. But GPU is only for simple physics. You can calculate +10.000, but al in a simple way. It you really want to have a large scale I suggest to take a look at agiea physics engine. They also make the hardware. And if ist only server-side, the only one needing a physics card is the server. (they cost around $250 I think)

[Setharian]

I'll look at both the shader approach (if it is actually useable) and also at Ageia PhysX engine....but more tips are welcome still

[NecroDOME]

You could also "share" physics, let multiple computers do the calculations...

however, here's a demo movie from havok using nVidia card for physics:
http://video.google.com/videoplay?do...arch&plindex=1

[Mirage]

Setharian, before you start investigating a cluster-based/additional hardware approaches, I propose you to make sure that a single server can't perform the task. I think it can.
To speed-up the calculations you can use an early sphere-sphere test, as well as a kind of space partition (I suggest BSP for static, non-heightmap-based landscape/level).

[NecroDOME]

200.000 moving objects can be calculated on the CPU, but it would be VERY slow. Newton as already trouble with 200-300 moving and colliding objects. And then I'm talking about 1 room. If you want to test where the object is, what it collides with etc, you can't preform that on today's default desktop pc's. even with a quad core you still have not enough power. Thats why Havok uses the video card and Agiea uses its own hardware.

However you could give it a try, but I'm sure its slow.

[Setharian]

maybe cluster-based approach won't be needed....the great advantage is that one doesn't have to consider collision of 2 moving objects in detail (simplified bounding volumes like spheres/elipsoids are used for this purpose), only moving object <--> static object checks are needed to be more precise because usually the decoration (and terrain) cannot be encapsulated in a simple bounding volume for a reliable collision detection....object management works on an octree principle to quickly perform enumeration/addition/removal and I was thinking to extend it to contain static geoemtry as well....every individual "geometry object" would contain an AABB and an internal BSP tree of all the polys contained within it....also models used by the server are simplifications of those used by the client (usually have only 10-15% of polys of the original model)....because all of the geometry is static, AABBs and BSP trees can be precomputed....do you guys think there is any "hidden" problem in this approach?

[JernejL]

New newton release is slowly closing release, and it runs off GPU with support for multiple cpu cores, so it runs pretty fast:

http://www.newtondynamics.com/forum/...r=asc&start=75

see videos posted there on bottom of that page.

[savage]

maybe cluster-based approach won't be needed....the great advantage is that one doesn't have to consider collision of 2 moving objects in detail (simplified bounding volumes like spheres/elipsoids are used for this purpose), only moving object <--> static object checks are needed to be more precise because usually the decoration (and terrain) cannot be encapsulated in a simple bounding volume for a reliable collision detection....object management works on an octree principle to quickly perform enumeration/addition/removal and I was thinking to extend it to contain static geoemtry as well....every individual "geometry object" would contain an AABB and an internal BSP tree of all the polys contained within it....also models used by the server are simplifications of those used by the client (usually have only 10-15% of polys of the original model)....because all of the geometry is static, AABBs and BSP trees can be precomputed....do you guys think there is any "hidden" problem in this approach?

wow that's hard to read! Made me realise how much I take white space for granted.

[Setharian]

New newton release is slowly closing release, and it runs off GPU with support for multiple cpu cores, so it runs pretty fast:

http://www.newtondynamics.com/forum/...r=asc&start=75

see videos posted there on bottom of that page.

Newton is great overall but I fear it is "too precise" for our needs, I think it takes into consideration a lot of factors which is kinda of an overhead when they are constant. I'll do some further research on this.

wow that's hard to read! Made me realise how much I take white space for granted.

I will try to add white space next time I'll write a long post.