OpenCL

[pHySiQuE]

I'm just curious what the OpenCL stuff in the Newton SDK does? Is this usable ir just experimental?

[Julio Jerez]

I found out that must game are in fact bound by graphics and videos, and in academia no one had the need to do gpu physics with physic engines.
for the most part GPU physic are check box used by Video car makers with the purpose of selling hardware, no nothing more.

Basically physics engines are lost leaders for Video card companies, they spend 20 to 30 millions at year in high paid staff and marketing so that the go to convince game developers to support GPU physics as an add on.
They actually have no interest on making the GPU physic a solid core technology, they want it to be special.
for a person like me even if I have the best GPU implementation ever, it will not be accepted by main stream developer, no because the engine quality but because of politics and publicity.
Newton is already far superior to Havok and PhysX, but that plays not role at all on the perception from developers.
Their only inters is to have the company listed as "Uses GPU physics" because that stimulate the sales of high end GPU among hard core games. It is a mutual symbiosis between Card makers, Game developers and Game publishes.

The way I see it, it is only a matter of a year or two until 8 cores CPU became the main stream and possible entry level and by that time there will be no need to do GPU physics anymore.
I mean PlayStation 8 and xBox next are both 8 core CPUs wih 8 G of ram and they are coming out to the market this week.
That will force Intel to come up with 16 or maybe even more cores mainstream CPUs, they already have 32 an 50 cores, so is I not even a technological challenge at all, it is market.

so I guess the answer is, is still experimental.
I am not going to continue doing the Playing monkey game. "Do this or else I will no use it". I do what I can and If people do not want to use, then more power to them, there are planty oteh engine outthere

[pHySiQuE]

No problem, I just saw it and was just curious. Paying attention to the core features is definitely more important.

[Julio Jerez]

I looks like Intel release a new version fo Open CL
http://semiaccurate.com/2013/04/29/a-lo ... rformance/

thing do chneg from season to season, a year back, I thought that maybe OpenCL is a waste fo time,
however the trends in CPU are making me chnage my mind.

I beleive that at least in teh Mid Range and low end CPU teh future of Descrete GPUs its not very safe.

I do beleive that tehre will alwy be High GPU, but I beleive that most system will have integerated CPU and GPU
I you look at the new consoles they do no have GPUS anymore. is all a CPU/GPU with integrated memory.
The change a lot of things. GPUS on CPU are more flexible that Discrete CPU wi th exclusive Memory.

on the article say that Intell has actually done a Good Job with OpenCL 1.2, it is actaully beter that AMD even when AMD
has being trumpeting the APUs.
Maybe it is time to give anotherr crack at NewtonCL

[Neill3d]

Hi Julio,

interesting news. But what about compute shaders (they are avaliable now in OpenGL 4.0+) and CUDA. Why dont you look at these tech ? There are not enough memory, shared memory for working with physics ?

[Julio Jerez]

Oh wow, I did not know that, I always thought that compute shader and direct compute was a D3D thing.
But you are right OpenGL support Compute shader and even better it is use with GLSL

I now read some and it appear that Compute shader is also independent of the graphics, but their real strength is the access to graphics buffers.
It sound really good.
If I can do this with having to install any external SDK, I will actually try Computer Shaders instead.

to me the only attractiveness of OpenCL was that Intel was making for Intel CPUs which can use the AVX stuff.
But after trying I found that there are lots of limitation simulate to the CUDA, D3d limitations with Vendors and platforms.

Using OpenCL is really cumbersome it requires installing a separate driver and separate libraries lot of setup (queries, platforms, devices, queues, kernels, etc.).
and worse many of those libraries can not coexist, OpenCL is not as wondefull as it is depited bu the Fans, and the people hwo write article which usually simple run a simple test
and know nothing about the details.

If ConputeShader is just there an only not installation other than the graphics drive that most be in the system, hell yea! I will go with it
I really like the ComputeShder Idea, I think I will give it a try.

[Neill3d]

Yes, you should try, they become really powerfull and independent (means multiplatform) There is a nice resource with good samples to check - https://github.com/progschj/OpenGL-Examples
There are samples which covered basic n-body simulation and one more sample about computer shaders particle system - https://github.com/StanEpp/OpenGL_ParticleSystem

[Julio Jerez]

yes specially now that I am adding the cloth and soft bodies, the compute shader seems a good candidate for accelerating that.

[Neill3d]

I'm currently learning TressFX and dreaming to implement it on OpenGL with compute shaders - http://developer.amd.com/tools-and-sdks ... adeon-sdk/
It's a particle system with 2point poly chains and simple dynamic and simple collisions, but quite interesting tech for real-time characters.

[d.l.i.w]

In an ideal world compute shader really would be the a good option to do physics calculation.
And in contrast to general GPGPU (aka OpenCL and CUDA) they are very closely connected with graphics.

But they have two big problems:
There are different compute shaders for OpenGL and DirectX, which means maintaining two sets of independent shaders and they require quite recent hardware capable of DX11 (for DirectX compute shader) or OpenGL 4.3 (for OpenGL compute shader).

I don't want to say, OpenCL would be better in any case - it's far from perfect, but it works on older (and wider range) of hardware and has full support for both, OpenGL and DirectX since OpenCL 1.2.
https://anteru.net/2012/10/30/1998/ gives a short overview (although it's mostly outdated now)