pouët.net

I'm trying to learn it at the moment. Has anyone here done anything interesting with it here?

Not yet xernobyl, but going to soon... I'm also interested in this thread

I do. Using the beta drivers that come with Snow Leopard.

I'm just looking at the code samples in the ATI Stream 2.0 Beta SDK for now, using the ATI beta drivers.

as far as i understand in current state you have to compile your opencl programs for both nvidia and ati -_-.. and compilation is static.. which is pretty much a blocker for me :/ so i guess i just wait to upgrade to dx11 hardware to play with CS5.0. unless someone has a better option in mind... :)

I got my minimal dx11-computeshader5 setup up and running, and after looking very bad in the beginning it's now down to around 70 bytes more than the similar dx9-pixelshader3 intro.
I doubt opengl+opencl can compete with that (both need to be setup etc).
But ofcourse the opencl route supports winxp and maybe also have better dx10 level hw support than cs4.

and yes, there isn't the comon .dll solution for opencl on windows yet, which is a bit of a blocker for now.

I run linux 100% so I'm pretty much tied to opengl...

thanks for that vital piece of information that you run linux.

uhm... that doesn't sound very cool... do you know if this is goind to change in the future?

texel: I've recently downloaded GPU computing SDK from NVIDIA. In their example code OpenCL kernels are read from string/file and compiled *at run-time* during initialization (just like shaders).

masterm: the problem is the host program, not the opencl kernels. There is no such thing as a general "windows opencl driver" yet. On OSX, where apple is doing (the frontend of) all implementations that's not a problem ofcourse.

I tried NVIDA OpenCL when the drivers were first released, after 6th or 7th reboot while running the examples I decided that I'll wait a bit longer before I start coding OpenCL. Tested on Windows 7 RC, NVIDIA 8400GS.

I tried the sample examples on Macbook Pro 13", everything worked just fine...

I'm still figuring out if OpenCL gives me any advantage for gfx coding, compared to normal OpenGL+GLSL approach. I might start doing some stuff if i come up with something cool.

And no, i don't consider moving the load from cpu to gpu "cool" :)

If you have a use case where the actual processing has a stream-like structure, such as video decoding, animating particles, rigid body dynamics etc., the gpu will do it better. Stuff with lots of branching however does not fit well. Also keep in mind that GPGPU was originally invented for scientific computing, not for games.

I have a possible use for it (non-demo related.. doing some image recognition stuff, a lot of the machines that will run my app are the macbook pros with 2 GPUs, one could do the image recognition while the other does the image processing). But no hardware to run it on, as my ATI gpu lacks the hardware needed it seems :(

What's the situation going to be on windows? Won't a standard 'generic driver' be necessary? Separate builds for ATI/NVidia/Intel (well, one day maybe) would be horrible, and MS aren't likely to want to support it. Plus you want good CPU support, GPU based code isn't even the whole point...

Will it get to the stage where you can write generic openCL code and it'll run on any combination of CPU + GPU any time soon?

http://forum.beyond3d.com/showpost.php?p=1337667&postcount=276
kurli: the new buffers for writing (random acess, append) certainly open some options.

I do. I want those additional TFLOPs to do some nasty shit a CPU can only dream of. :)

These Khronos-standards usually needs some time to establish implementation guidelines in order to get (some) binary compatibility. They try not to interfere with the platform specifics too much early on, which IMO is a GoodThing(tm). At least that was the case for OpenGL ES.

In the mean-time this could be temporarily solved by dynamic loading of the libraries, or something like that. Yes, it's ugly, but it should work.

Is it possible to generate vertex buffers with it?
From my fast readings it looks like a possibility.