Open questions for Bullet Physcis

[Lonbjerg]

I look at this site but couldn't find an email to contact people from this site, so I figured I would write them here and hopyfully get some answers.

Hardware physics is a hot topic today.
It's debated across many sites/boards.
That is why I am seeking information about the benefits of Bullet OpenCL physics and a comparision to Bullet CPU physics.

1. Does anyone have any documentation or informed views regarding how much more powerfull Bullet 2.77 OpenCL compared to Bullet CPU physics?
Like how much more powerfull is it to run Bullet on a GPU compared to eg. dual, quad and hexacore CPU?.

2. Where is the current limitions for achiveable Bullet CPU's physics?
Can you run cloth on a top end CPU today?
Tearable cloth?
Interactive smoke?
Interactive fluids?
Destrauctable architechture?

3. How big is the difference on simple riggid bodies physics when comparing CPU's to GPU's?
Like how many object can be run on the CPU and how many object can be run on the GPU?
I am trying to quantify the difference

4. How many CPU cores would it take to match a lowend GPU?
And a midrange range GPU.
And a high end GPU?
Is there a interection point?

5. Is OpenCL done or is there more work required beore the framewrok is done?

I hope I can get some informed inforamtion fro mthis board...or get pointed to the right place for these question.

Tanks in advance

Lonbjerg.

[Erwin Coumans]

The GPU acceleration of cloth in Bullet 2.77 using OpenCL or DirectCompute is just a starting point. This is work-in-progress and Bullet 3.x will have more GPU acceleration.

You can perform the comparison of CPU versus GPU using the precompiled Windows demos:
GPU_Demos/AppOpenCLClothDemo_MiniCL_CPU.exe runs using multi-core CPU in 4 threads, and GPU_Demos/AppOpenCLClothDemo.exe runs on GPU.

Just check out the console and see the performance numbers, in particular solveSoftConstraints (see below)
Dependent on the algorithms, a GPU can be more than 10 times faster than CPU in particular if you keep the data on the GPU
(avoiding the copy from GPU memory to CPU memory and back).

We will try to have some more information including a Bullet 3.x preview by March 2011 at the Game Developers Conference in San Francisco.
Thanks,
Erwin

GPU NVIDIA Geforce 260:
Profiling: Root (total running time: 0.704 ms) ---
0 -- stepSimulation (99.57 %) :: 0.701 ms / frame (1 calls)
Unaccounted: (0.426 %) :: 0.003 ms
...----------------------------------
...Profiling: stepSimulation (total running time: 0.701 ms) ---
...0 -- synchronizeMotionStates (0.00 %) :: 0.000 ms / frame (1 calls)
...1 -- solveSoftConstraints (75.75 %) :: 0.531 ms / frame (1 calls)
...2 -- internalSingleStepSimulation (12.55 %) :: 0.088 ms / frame (1 calls)
...Unaccounted: (11.698 %) :: 0.082 ms


CPU Intel Quadcore Q6800 at 2.93 Ghz:
Profiling: Root (total running time: 12.511 ms) ---
0 -- stepSimulation (99.96 %) :: 12.506 ms / frame (1 calls)
Unaccounted: (0.040 %) :: 0.005 ms
...----------------------------------
...Profiling: stepSimulation (total running time: 12.506 ms) ---
...0 -- synchronizeMotionStates (0.01 %) :: 0.001 ms / frame (1 calls)
...1 -- solveSoftConstraints (68.69 %) :: 8.590 ms / frame (1 calls)
...2 -- internalSingleStepSimulation (15.21 %) :: 1.902 ms / frame (1 calls)

I look at this site but couldn't find an email to contact people from this site, so I figured I would write them here and hopyfully get some answers.

Hardware physics is a hot topic today.
It's debated across many sites/boards.
That is why I am seeking information about the benefits of Bullet OpenCL physics and a comparision to Bullet CPU physics.

[Lonbjerg]

Thank you very much for taking the time to answer my question, it is much appriciated.
I looke forward to hearing more about Bullet 3.x