OpenCL In Gallium3D By Summer

[Wyatt]

Ahh, LLVM again. Every inch it gains makes me happier, and I can't wait until Clang is further along in its support for C and C++.

[Pfanne]

Current thinking is that the open source drivers will use Gallium3D for 3xx and up. The R300 work MostAwesomeDude is doing should cover 3xx-5xx, and we plan to switch our 3D work on 6xx/7xx to Gallium3D once we have basic 3D up and running on those chips using the "classic" Mesa hw driver model.

cant tell you how BADLY i want this

[rogerdv]

Well, this is a bit offtopic but, what about propietary ATI/Nvidia drivers?

[Pfanne]

Well, this is a bit offtopic but, what about propietary ATI/Nvidia drivers?

they might be faster with the opencl support but they certainly wont support gallium3d.
but thinking about proprietary gallium3d drivers is interessting xD

btw:
talking about gallium and stuff:
would gallium actually be usefull for multi gpu or on the fly gpu switching since it abstracts the hardware.
i dont know to much about gpu drivers but is it possible to switch the underlying gallium driver?

[drag]

i dont know to much about gpu drivers but is it possible to switch the underlying gallium driver?

I think that is the point behind Gallium.

With the older-style Mesa-based DRI drivers there was a hell of a lot of code that was made dependent on one card or another.. a monolythic approach at doing drivers.

With Gallium it's designed in a more modular manner were the device-specific code is kept as concentrated as possible. I don't know what it is any more, but there was a 'Winsys' driver concept were Gallium would just need that portion rewrote for each GPU while the rest of the driver remains relatively unchanged device to device.

This way your able to abstract the differences and support all sorts of different setups... Keep in mind that GPUs are designed for highly parrallel operations, each modern GPU is made up of dozens of smaller 'cores' or pipelines or whatever that can be used simultaniously. So that supporting multiple GPUs or architectures like Cell shouldn't be terrifically difficult.

Which makes sense then because OpenCL is designed to work across lots of different archs.

[_txf_]

who is willing to bet that nvidia comes out with fully working OpenCL support in linux before everybody else(they are only a step away with cuda)?

Not that I like proprietary drivers, but it seems to me that open source stuff (particularly graphics) seldom arrives anywhere near "on time". I'm not being disparaging but just saying that this is what usually happens for better or worse.

[Pfanne]

who is willing to bet that nvidia comes out with fully working OpenCL support in linux before everybody else(they are only a step away with cuda)?

Not that I like proprietary drivers, but it seems to me that open source stuff (particularly graphics) seldom arrives anywhere near "on time". I'm not being disparaging but just saying that this is what usually happens for better or worse.

well ati has stream
but i think nvidia will be first, too

but having early opencl support can only be usefull...
who would write software that isnt supported by any drivers?
so by the time gallium3d or ati adopts opencl we have a nice set of opencl software already written
well at least thats the theory

[MetalheadGautham]

Wait a sec, I read somewhere that OpenCL (about which I just have very little abstract knowledge) is owned by Apple ? And its OpenSource ???

[BlackStar]

Wait a sec, I read somewhere that OpenCL (about which I just have very little abstract knowledge) is owned by Apple ? And its OpenSource ???

AFAIK, Apple owns the trademark. OpenCL is an "open" specification, which means that anyone can implement it. For an implementation to use the "OpenCL" name, it will have to pass a certification process for a nominal fee, but that's it.

OpenCL is not open source, because open *source* does not make sense for a spec. OpenCL can have open source implementations, however - pretty much like Mesa3D is a non-certified, open source implementation of the OpenGL spec.

[89c51]

http://fireuser.com/blog/amd_opencl_parallel_computing_demo_from_siggraph_asia_2008/

it seems that ATI has something