OpenCL SDKs not cross platform compatibles

[Julio Jerez]

So in the end just like everything that is controlled by the Khronos group control, OpenCL does not work.
I just bought a laptop with an AMD apu. I could buy a better CPU like and Icore 5 or 7 but I chose the lesser AMD for compatibility testing.
Guess what, Intell OpenCL SDK does not work on AMD hardware.
This OpenCL BS was suppose to be easy for the developer not a territorial war between Intel, AMD, NVidia and Mac, and the developer is the collateral damage.

The entity to Blame is the * Khronos group, there are suppose to enforce the standard, instead the let everyone vendor to get away with a Moronic Extension.

I am abandoning ApenCL all together I am sick and tired of the war between the the bullies where SDK A do not work on hardware B. The purpose of OpenCL to be cross compatible a the hardware and software level. the is ridiculous.

I will go with Microsoft AMP,
http://msdn.microsoft.com/en-us/library/hh265136.aspx

Miscrosoft is a software company their interest is to sopport all platform not just one.
I hope GCC adopt AMP, because I am no doing CUDA, Direct COMPUT, OpenCL pr any of these not Cross platform* BS.

[pHySiQuE]

.

[PJani]

I think, base engine should be finished first. Like vehicles, joints and other stuff which makes engine useful. And later you can continue with parallel processing.

[Julio Jerez]

Yes I agree too, I only want to build the code on an amd laptop that I bought.

I am moving the openCL to its own library, because VS can not open the physics project.
I am simple re factory the project so that the Physics project does nos uses any OpenCL stuff.
what I am working now is the continue collision

[JoeJ]

I'm not surprised that SDKs are not compatible.
Would it be an option to use multiple SDKs for multiple hardware / platforms? Would the same CODE work for all?
However, i agree it is the best to wait for something that's easy to use, as long as there are other things to do.
I do so for for years, the day will come.......

[Julio Jerez]

basically what I am doing is making two separate projects, rather than separate configurations in the same project
there will be a folder project_vs_2008 which will be the basic library with no openCL at all
and I also made projectOpenCL_2008 which has the openCL layer

all the openCL: code is move to its own library, and that library include the IntelSDK

I am no planning to make install any AMD, or Nvidia Cuda SDK but it should be very easy for anyone to install those SDKs and, copy and paste the Intel project to make a new one for a particular SDK.

I am still hoping that at the drive lever the binaries are still compatible, But knowing how the Khronos group works, I am not holding my breath. Everything those Morons take control over, they ruin it.
The destroyed Collada, they try to destroyed OpenGL an dtehy are still at it, and now they want destroy OpenCL. why Apple pass control of OpenCL to those idito is beyond comprehensions to my

[Julio Jerez]

O check in the new propject and solutiuons, for VS2008 and vs 2010 please see if they work without opencl sdk

[JoeJ]

Built newton.vcproj and demosSandbox_VSExpress.sln both without problems on VC 2008 Espress.

[Julio Jerez]

could you try please test this solution demosSandbox.sln in VS express
I think the express solution is not need it anymore.

note:
Oh there was still a bug, please sync and build again.
solution demosSandbox.sln spoudl buil with VS express.
also demosSandboxOpenCL.sln should build too if you has intel openCL installed

[JoeJ]

Oh, that was an old file.

demosSandbox.sln works too, but there are the usual warnings: Solution folders not supported.
Compiling works, but sandbox is terrible slow both in release and debug.
Maybe thats an GUI issue, because the timings show usual values for physics and graphics.

I tried the built dll in my project and get an assert in void CalculateClosestPoints (), Line 840.
Seems like a big bug. I need to disable pretty everything to get around it.
I can run with box ground and dynamic sphere + ragdoll, but it happens again when i smash the ragdoll on the ground. (but performance is ok in my app)
Want serialized scene?
If so, to what email?

[Julio Jerez]

The slow down is because I have a sleep(40) in teh main loop so that I can see teh contacts as they are generated.
I check it by mistake.

and I forget the VS express do not support solution forders. Maybe I should remove them, but there are so cool to work with.

any way the assert in closset point will be checked in with next update.

[JoeJ]

Maybe the new OpenGL Compute Shader is usable for alternative GPU acceleration? Requires 4.3.
http://www.opengl.org/wiki/Compute_Shader

[carli2]

Maybe the new OpenGL Compute Shader is usable for alternative GPU acceleration? Requires 4.3.
http://www.opengl.org/wiki/Compute_Shader

Please don't tell such Ideas to Julio. I'm running newton on a headless server and I don't want to install all the opengl libraries on that server including it won't be hw accellerated and newton may fail because the software renderer does not implement OpenGL 4.3.

IMO GPGPU will die with the upcoming of cloud services and 16/32-core CPUs.

[JoeJ]

With 'alternative' i mean the user should be able to choose that option.
However, i'm no GPU expert and don't know if Compute Shaders can compete against OCL or Cuda (it seems so!),
but OGL is everywhere, and you can use Compute Shaders without graphics context.
So it should be fine to mix it with DirectX graphics or anything... maybe.
Even a 32 Core CPU may be outperformed by a single NextGenGPU in lots of cases. They may share the same chip, but don't die.
But i agree that physics on GPU have low priority, as long as it's about rigid stuff.