Newton on GPU (CUDA, OpenCL)

[aqnuep]

Sounds great!

Btw I'm afraid especially of soft bodies and destructible bodies, because these seem to be unfeasible for complex scenes without a GPU implementation, so it's good to hear that the new stuff will be OpenCL based.

Thanks anyway for your understanding and keep up the good work.

[Dave Gravel]

Just some new links about.
http://developer.apple.com/mac/library/documentation/Performance/Conceptual/OpenCL_MacProgGuide/Introduction/Introduction.html
http://developer.apple.com/mac/library/samplecode/OpenCL_Procedural_Grass_and_Terrain_Example/
http://developer.apple.com/mac/library/samplecode/OpenCL_NBody_Simulation_Example/index.html

Some more link and code exemple.
http://developer.apple.com/mac/library/navigation/index.html

Driver:
http://developer.nvidia.com/object/opencl-download.html#Windows

[Julio Jerez]

wow this does not wsoudn like very good news.
http://www.maximumpc.com/article/news/intel_cancels_larrabee_graphics_chip

I guess will will be stocked with dreaded GPUs for high performance computing.

[Dave Gravel]

If they do equivalente product like intel video card, maybe it is better for they to forget it.
I remember to have seen they talk about intel video in the pass for say it is good and next new version in coming is powerfull.
They talking in the pass about video card performing like nvidia or ati, but when the product have come out outch....
The intel video crash for nothing in 3d and is a hell to use in programmation.

Maybe intel need to buy nv like amd buy ati, but for now I don't think that is possible.
Buy a intel video card for play game is equivalent to buy a via cpu for play, you risk to just play to reboot your computer.

[Stucuk]

The problem with Intel Video cards is that they don't actualy support what they say they do. They say they support say an opengl extention called BlaBlabla yet in reality any game that uses it won't work as they fail to implement it properly or they don't implement all of it.

Basicaly Intel fail because they don't go all the way and make something that can compete, they just make stuff which works at the most basic level.

[Julio Jerez]

What I do no undertand is why they insist in making a video card out of so awesome CPU design,
They should just make Larrabee a general Proccesor and let the people use for whatever they want to use it.
Each Larrabee core is a general Pentium, why limited to graphics? what it can do Graphiscs and everythong.

Plsu must high en graphcis driver are in suftwere, stuf like Raytracers for graphics package lik mac, maya di rdiasity, and all teh stuff tehy do in software ray tracers it is not OpenGls.

In other word they just need to raplace the SSE/SSE2 witch are 128 bit wide with Larrabee vectos which are 512 bit wide.
I believe people will use it for making graphics drivers that will kick the ass of these GPU cacaminia car driver.
If they do that they will fill the reign of OpenGL and DirectX ass standard graphics API.

[Dave Gravel]

About video yes sure. inside!

[Ivan Mandić]

Excellent. I can't wait for NGD 3

[Nitrogenycs]

Are there any updates on the "Newton on GPU" topic?

I have a custom cloth simulations going on the CPU + GPU, but for collisions against rigid bodies it can get slow. Also I'd really love to see 3d fluid and soft bodies in my games The solution to a lot of these issues seems to be running at least parts of the physics on the gpu (especially all those particle based things like cloth + fluid).

So I am wondering if Newton is developing in this direction at all or if this was put on ice.

[JoeJ]

Maybe i've some inspiration for cloth simulation.
I was playing around with this some time ago, because of Softimage XSI cloth frustration.

The yellow rings on the green cloth rig are build from a pair of particles, which are constrained to stay on skeleton and to keep minimal distance from neighbouring particles.

The cloth itself is constrained to keep maximal distance to the rig, and more important not to cross the plane at the rigs attachement point.
This ensures stable cloth - it's impossible to cross a leg through a long skirt for example, no matter if character is kung fu fighting or teleports around.
The donwnside is that the cloth rig requires knowledge of the unique ragdoll, so this adds complexity to the content creation.
However, if the crossing thing becomes a problem - i think it's the "drag and jitter" in clothsim - it may be an option.

For coll. detection the cloth vertex traverses the ragdoll capsules as a tree to quickly find a set of skin vertices linked to a specific region on the capsule.
GPU Grid may be better as it also can handle self collisions in one go.

Performance is good, demo video done on single thread CPU, also skin catmull clark in realtime, anything unoptimised - no SIMD.
Planned to port to OpenCL, but 1 year ago Cuda was twice faster... has that changed meanwhile?

Video:
http://www.megaupload.com/?d=KOGVQBA6

[Julio Jerez]

Hey that look very good.
Much better than pasted cloth pain on a body. Very cool.

[JoeJ]

Yes, but it looks even better without the cloth

I first thought it's too specific for a common physics engine, but it would be possible if you give the user an option to specify a point and normal for each cloth vertex. So you can solve the plane and distance constraint, and the user decides if he uses skin point directly, or to implement some cloth rig like i did.

Maybe this is standart practice right now and you mean that with pasted cloth, just an idea.

[Julio Jerez]

I had being very reluctant to play with CUDA and OpenCL.
Not because it is difficult, I actually implemented the CUDA solver in Netwon 2.0 when I started.
The solver is fully functional, and I wanted to continue with teh collision part.
However the elitist of the people over the nvidea forum each time I ask a question alway replay with ironic statments or never reply at all.
and that made me abandone the project. I vowed never touch any free form nvidia or AMD tool even with a ten foot pool.
On the Collada forum, I remember one of the Guy there told me they do not want people making waves overthere. So I droped Colldada too.

But that was then, now It appears intel is making an OpenCL driver for thier new high performce cpu using AVX and simd.
Thsi soudn lie a good news, and can be use as a alternative to thes nonse sence liek CUDA, and Direct Compute.

I found this ove teh intel new stet: http://software.intel.com/en-us/article ... pencl-sdk/

So maybe it is time to look at OpenCL again for newton 3.0

[thedmd]

Quite good news. I doubt SSE 4.1 instruction was executed once on my machine, this is a chance to change that.

Julio, do you have any benchmark code for Newton?

[Julio Jerez]

what do you mean bench mark?
All I have is the limited profiler the measure the performance of the major systems.