pouët.net

OpenCL 1.0 specification released.

category: general [glöplog]
 
Quote:
Khronos Group has today announced the ratification of the OpenCL 1.0 specification! The 1.0 specification of the Open Computing Language is backed by Apple, AMD, NVIDIA, Intel, and other industry leaders as a new open standard to exploit graphics processors for general-purpose computational needs.

What OpenCL 1.0 defines is a C99 programming language with extensions geared for parallel programming, an API for coordinating data and task-based parallel computation across a wide range of heterogeneous processors, numeric requirements based on the IEEE 754 standard, and efficient interoperability with OpenGL, OpenGL ES, and other graphics APIs.

Yay! I hope THIS TIME this is gonna mean that there will be a well established standard in gpu computing. And yes, I'm a bit realistic too... there will be massive incompatibilies for sure but at least one will not have to write (completely) different code for various cards.
added on the 2008-12-09 12:42:27 by masterm masterm
so that's what they were doing instead of a proper opengl spec!
added on the 2008-12-09 12:45:36 by Gargaj Gargaj
Microsoft are doing their own thing for directx, as usual. It didn't sound anywhere near as useful, but I only took a cursory glance at it.

Hopefully this will fulfill the promise that cuda and all the other half-assed and incompatible standards are showing, and replace them all.
added on the 2008-12-09 12:46:07 by psonice psonice
Lookie, mom! SpookySys' name is in the "Acknowledgements"-section! Let's blame him for all the issues!
added on the 2008-12-09 12:53:05 by kusma kusma
i wonder how much the framework packs for 4ks.. maybe it is more useful than CUDA.
added on the 2008-12-09 13:07:32 by nystep nystep
Quote:
Microsoft are doing their own thing for directx, as usual. It didn't sound anywhere near as useful, but I only took a cursory glance at it.


depends if you call useful == "will actually be likely to be supported properly by all the consumer gfx hardware on the market in a couple of years". i do. :)
added on the 2008-12-09 16:56:24 by smash smash
Now I finally have the chance to learn coding in something that is really new!
added on the 2008-12-09 17:55:43 by Adok Adok
oh jolly! *rubs hands*
nystep, my cuda mandelbrot fractasl is 1k. I will try opencl and see.
added on the 2008-12-09 19:24:26 by iq iq
It would be nice to see some eye-popping raytraced demo on it. :)
added on the 2008-12-09 19:45:31 by masterm masterm
since i liked cuda (means: doing stuff with cuda, NOT the hassle getting it to work), this might be worth a look. btw, cuda failed for me because it was not able to provide double precision at the time i was using it on a 8800gtx.
added on the 2008-12-09 19:55:13 by jco jco
AFAIK CUDA has double precision but it's utterly slow compared to single precision.
added on the 2008-12-09 20:12:40 by masterm masterm
The performance of double versus single precision depends on what the hardware is designed for. Since most of the sales of highly parallel computing devices are driven by graphics cards, and double precision usually does not give much benefit in graphics, there is at the moment not much incentive for hardware vendors to spend significant chip space on it. But the fact that it supports double precision at all signifies a shift towards computing which probably will only get stronger.

As an example, to quote Nvidia's own specifications for their Tesla chip (which is primarily targeted at computing, though it is of course a spinoff of their GTX 200 series GPUs): "933 GFlops Single Precision and 78 GFlops Double Precision".

78 GFlops is still pretty heavy compared to what you get from a CPU of course. :)
added on the 2008-12-09 21:54:57 by Blueberry Blueberry
Quote:
depends if you call useful == "will actually be likely to be supported properly by all the consumer gfx hardware on the market in a couple of years". i do. :)


Totally.

Also, OpenCL with shitloads of extensions versus Microsoft, that reminds me of an old thing somehow. Let's see if the Khronos people learn anything at all from their mistakes, especially the disappointing OpenGL 3.0
added on the 2008-12-15 17:41:57 by keops keops
If you think OpenGL 3.0 was a downer you haven't had a look at the OpenVG spec. It's an API kind off like OpenGL, but aimed for vector graphics (think flash or svg-rendering). Aimed for embedded systems... Ought to be a simple thing, right?

However, the guys who wrote the spec (Khronos that is) made sure that the spec is so much over the top that it's nearly useless. Want some examples?

* The spec has no fixed-point interface and internally floating point precision is required. Great if you work on a ARM <= ARMv6 or on a DSP.

* You have to support each and every possible pixel-format. They even be so bold to *invented* a former unseen order to pack A,R,G and B into a dword. Not that we already have enough variations around.

* The guy who wrote the spec was so amazed about linear color-space that all color-conversions happen in linear color space. And they did it to the extreme: The conversion rules are speced to bo be piecewise (e.g. not a simple pow (a,x) table. And they specced it in a way that you have to do it channel by channel. Well - this would offer the opportunity for tone-mapping and all, right? The bit-richest color-format is ARGB8888 though. Totally useless for linear color space usage.

Don't get me wrong. OpenVG fills a gap and it would be damn cool to have OpenVG as a standard API on PC's, but what these guys did was over the top. And when it comes to features it seems like the guy who wrote the spec was carried away quite a bit.


Khoronos == Stupid Ivory Tower Idiots.

added on the 2008-12-15 18:00:02 by torus torus
How likely are extensions though? For graphics, I can see the need.. there's always new techniques, formats, etc. that need supporting. For a compute language I can't see it being anywhere near so much of an issue.

Also, nvidia + ati are both supporting this.. from what I've read, nvidia are saying opencl will just be an extra layer on top of cuda, so if they + ati are putting cuda/stream into their drivers adding openCL support shouldn't take long.

As a side note, I know that apple were demoing opencl based stuff back in september, and they'd need driver support to do that so I guess that the vendors are pretty far along with supporting it..
added on the 2008-12-15 18:01:28 by psonice psonice
[quote] so that's what they were doing instead of a proper opengl spec! [\quote]
what gargaj said
added on the 2008-12-15 19:27:07 by auld auld
fail!
added on the 2008-12-15 19:27:39 by auld auld
this could be a way to get some pretty sweet (and pretty) particle effects, like snow blowing around in a blizzard.

login