Real-Time Physics Simulation Forum

GJK-EPA on GPU Robust in latest Bullet version?

2020-03-13T18:14:49+00:00

Hi,

I'd like to use the GJK-EPA algorithm to computing the penetration depth between pairs of colliding bodies (convex polytopes), specifically for robotics applications as my master thesis. To accelerate the collision detection process, the plan was to do implement this using GPUs and CUDA. I see that Bullet already has an OpenCL version of the GJK-EPA algorithm and I was thinking or reusing this as is or as a reference to implement a CUDA variant. Does anybody who has used this particular algorithm on Bullet for computing penetration depths on GPUs know if it runs into numerical stability issues? Is it safe to use as is? Thanks a lot for your inputs and time in advance!

Warm regards,
Abhiraj

Statistics: Posted by abhirajbishnoi — Fri Mar 13, 2020 6:14 pm

Re: Bullet on GPU

2016-02-29T07:50:23+00:00

For one, opengl compute support is even more fragile, having only been added in 4.3 and not really designed for this type of application. So far the only real competitor is CUDA, but that of course requires Nvidia hardware, and the DirectCompute/OpenGL Compute systems are not quite ready to tackle non-trivial physics (but they excel at particles and even particle based fluids and solids)

I find compute shaders much more interesting than opencl. It extends an already stable API, supports natively GLSL types, can be run between different graphic stages and has access to all the opengl input infrastructure.

It is really a huge advantage for me, given one uses bullet in combination with graphic 95% of time.

Statistics: Posted by elect — Mon Feb 29, 2016 7:50 am

Re: Bullet on GPU

2016-02-22T05:38:20+00:00

Quick question, if opencl support is so fragile, why don't we use opengl compute?

For one, opengl compute support is even more fragile, having only been added in 4.3 and not really designed for this type of application. So far the only real competitor is CUDA, but that of course requires Nvidia hardware, and the DirectCompute/OpenGL Compute systems are not quite ready to tackle non-trivial physics (but they excel at particles and even particle based fluids and solids)

Statistics: Posted by Basroil — Mon Feb 22, 2016 5:38 am

Re: Bullet on GPU

2016-02-12T15:41:23+00:00

Quick question, if opencl support is so fragile, why don't we use opengl compute?

Statistics: Posted by elect — Fri Feb 12, 2016 3:41 pm

Re: Bullet on GPU

2014-10-23T13:00:42+00:00

I am trying to write a physics engine in WebGL here : http://www.borbitsoft.com/
Do all the heavy lifting on the gpu using shaders.
The best way to run the demos is to use Chrome or Opera on a pc with a recent graphics card and up to date drivers.
(Although IE 11 displays true, its WebGL implementation is not mature enough to run the tech demo's, have to fix that)

One of the hardest things I found was debugging the shaders.
I ended up by writing lots of tests and I build a system where a shader was run multiple times
to emit certain key variables selected by a debug id.
One of the issues with that was that due to the extra complexity of the added ifs the shader would
not compiler or emit bogus data.
Also the count of bugs encountered compared to working in C/C++ surprised me.
Now with the DX11 backend for WebGL in Chrome and the extra massaging the angle team (https://code.google.com/p/angleproject/) applies to the shaders this has improved a lot.

How do you debug the gpu code in bullet ?

Statistics: Posted by TiborDenOuden — Thu Oct 23, 2014 1:00 pm

Re: Bullet on GPU

2013-07-15T17:39:49+00:00

I'm wondering if there will be a double precision compatible bullet 3 at some point

Statistics: Posted by Granyte — Mon Jul 15, 2013 5:39 pm

Re: Bullet on GPU

2013-07-06T21:44:54+00:00

Hi

I'd like to know if Bullet 3.x (aka Bullet on GPU) is stable enough for production use. I see that it has moved from your "experiments" github directory, and I think you've been saying at GDC 2013 that it is almost ready for release.

I would like to use it with Flash through FlasCC (CrossBridge).

Bullet 3.x may not be officially released yet, but can it be used this way safely or should I wait. If so, do you have a rough idea of the official release date ?

Thanks !

Statistics: Posted by Satori — Sat Jul 06, 2013 9:44 pm

Re: Bullet on GPU

2012-05-26T18:00:29+00:00

The GPU work on Bullet is making progress:

there is OpenCL and DirectCompute cloth:
Bullet/Demos/OpenCLClothDemo and Bullet/Demos/DX11ClothDemo

as well as a full OpenCL rigid body pipeline. The work is part of the current Bullet source code release:
Bullet/Extras/RigidBodyGpuPipeline and videos http://youtube.com/erwincoumans

Statistics: Posted by Erwin Coumans — Sat May 26, 2012 6:00 pm

Re: Bullet on GPU

2012-01-13T16:55:13+00:00

It doesn't look like this thread has been updated since 2007, possibly 2006 regarding the status.
If this isn't live, and it isn't functioning, please remove the sticky status of this.

Statistics: Posted by kloplop321 — Fri Jan 13, 2012 4:55 pm

Re: Bullet on GPU

2011-03-06T12:18:49+00:00

what is the current state of GPU support in bullet ? are there any acceleration available for physic for Blender 2.5 ? or will it be ?

Statistics: Posted by delphi — Sun Mar 06, 2011 12:18 pm