Real-Time Physics Simulation Forum

Jack wrote:1) You Draw N polygons. Each polygon consists of N pixels. Right?

Yes.

2) Your render target texture contains only N pixels. Or N*N?

only N.

3) If N*N then I do not understand why you need second pass...

It's not N*N.

1) I see now. You execute it N times. Each execution may be repeated several times (bet it is 1 in most cases). So you make N GPU calls. I don't know what you mean by "GPU call". For 1024 objects, colliding against 1024 objects, each frame I do ONE call to glMultiDrawArrays which draws 10...

I took a look at your code. I am not sure, but I suspect that you send collision pairs to GPU one by one. glBeginQueryARB ( GL_SAMPLES_PASSED_ARB, query ) ; glMultiDrawArraysEXT ( GL_QUADS, (GLint*)& collstart, (GLint*)& colllength, 1 ) ; numPasses++ ; glEndQueryARB ( GL_SAMPLES_PASSED_ARB ...

So let's talk about simple shapes: Sphere/Sphere: My present code (in SVN - check it out!) is doing 'typical' sphere/sphere collision tests at the following speeds: 4 cubes: other = 0.25ms collisions = 0.75ms 16 cubes: other = 0.25ms collisions = 0.75ms 64 cubes: other = 0.25ms collisions = 0.84ms 2...

You should realise that the problem with using a cube-map is that there isn't a convenient way to pack lots of different shapes into a single cube-map. So my earlier concern about the number of textures you'd have to read in the shader would then be justified. This code actually works? I'm a little ...

I have this algorithm working in the SVN trunk. The results of a collision happening are not truly correct - but it looks OK at first sight! The system is doing this is as follows (and of course, runs entirely in the GPU): * For each object: *** Submit it to the collision detector and count how many...

1) 128x128x128 it is of course too much :shock: I played with 32x32x32 (64x64x64) At that resolution, pretty much everything is a 'blob'! Why bother? 2) No. The idea is not to replace simple shapes (spheres, cubes,...). ANY SHAPE. Then I don't think you'll have enough texture map memory for anythin...

Well, I guess the first issue is that you need a 3D texture with four components at whatever resolution you need your objects to be collided against. If you wanted (say) one part in 128 precision for your surface you'd need a 128x128x128x4x2 = 16Mbytes texture. You won't get many of those from a typ...

Maybe I am wrong, but I see you do not find COM (Center Of Mass) in local model coords. You assume that it is always at (0,0,0). Right? I believe that's true - it is the problem I've been having with some of my vehicle models which always come out top-heavy and hence roll over too easily. Being abl...

Can you find an easier way for us to download this? That site you've got it hosted on is a pain in the ass to use! It wants me to sign up for who-knows-what just to see your file!

If you'd like to email it to me at <sjbaker1@airmail.net> I'll put it up on my web site for people to get at.

(The GPU physics thread was getting kinda long...) I have a good way to do collisions worked out - it's not easy to understand - but it's going to work and it'll be fairly efficient. The idea is to go N-times around the 'main loop' where 'N' is the maximum number of collisions that any object underg...

We should probably do that conditionally so Linux version still works fast.

Incidentally, I got an email from Chris Seitz at nVidia offering help from nVidia on the GPU physics stuff. Initially, this will get me onto their developer forums - which may help me to understand why the heck this code doesn't work properly under Windows with vertex texturing enabled. (We know why...

I added some new code to the demo. * There is now a 'collisionMap' that stores up to four objects that have (nominally) collided with each object. * There is an extra shader that computes some sort of constraint force on every object from (up to) four objects it might have collided with. For the sak...

It is more about going parallel and how to deal with memory caching, rather then 'optimizing' the processor for physics. Well, there are some features that would make physics massively easier on the GPU. One of those would be the ability to write to multiple output textures instead of just one. Tha...