OpenCL (Open Compute Library)

So can anybody here tell us if nv's ocl drivers support double extension or not?

 

no.

 

You can't say they have double or you know that they don't have double?

 

They don't have it.

 

Just curious... which functions are so challenging? The ulp limits on the trig functions for example seem quite high (aside: it's also weird that the precision requirements for the *pi functions are less stringent than for the standard trig functions; I'd have thought range reduction easier to perform accurately for those).

I'm having a bit of a hard time buying that it's challenging for CPU based implementations to meet those specs, especially considering e.g. crlibm.

 

Does this mean that Adobe CS4 might finally offer some GPU acceleration for folk on WinXP 64bit?

I hear from developers that the WinXP 64 bit OpenGL support for CS4 by Nvidia (not sure of ATI) was so bad folk gave up and simply turned it off. Forcing it to be enabled - whilst possible - just generally throw garbage on the screen from my experience.

Anyone got clear intelligence on this please?

 

Has AMD achieved OpenCL conformance for GPUs? If you check this page, AMD's "confidential gpu" is mentioned as conformant on windows xp,vista, 7 as well as linux. So I guess its just a matter of time before we see a driver and sdk?

http://www.khronos.org/adopters/conformant-products/#topencl

 

Thanks for the update. I knew that it should be coming soon, because there was a press release a few weeks ago:
http://www.amd.com/us/press-releases/Pages/commitment-to-opencl-2009sep21.aspx

Lets hope that the drivers come out soon. Catalyst 9.10?

 

Although this is an OpenCL thread but I really do hope that 9.10 provides CS 4.0/4.1 support on RV770 >.<

 

pcchen: Seems unlikely from what demirug wrote yesterday

 

If AMD is not going to support CS 4.0 then it'd be very unfortunate because many of CS 4.0's limits are actually RV770's limits, not G8X's limits.

 

Bear in mind that RV770 is CS 4.1!

The D3D's team's priority over the past 6 months have obviously been to get DX11 (with DirectCompute 11) up to speed (well, past WHQL and then up to speed!). DirectCompute 10.1 will be available in the future.

 

I understand that. Great to hear that DirectCompute 10.x is still on the table.

 

(But will we see some usefull applications, that use DC 10.x?)

 

If it was supported on all DX10 cards, then it would be a no-brainer to use it for lots of stuff like post-processing and so forth rather than pixel shaders. However while it is supported on all of NVIDIA's DX10 parts, it's unclear whether this will be the case for AMD (or at least, I haven't heard a commitment either way on all of their DX10 generation parts... correct me if I'm wrong).

If you're forced to write a legacy pixel shader path, you're probably not going to get much more speed out of CS4.x honestly, so it's really only interesting if you can write that as the single path, in which case it's just much easier to do cool stuff.

 

Will it hurt much if AMD doesn't support 4.0 on R600 and just focuses on 4.1 (Rv670 and up)?

I can't imagine there's many R600 boards in the wild still. Heck, I tore down my R600 and recycled it once I replaced it.

Regards,
SB

 

Is it possible to support 4.0 on R600? I mean, I've heard conflicting reports on whether R600 has shared memory or not... If not, it's impossible to support CS 4.0. Although, as you said, there doesn't seem to be many R600 boards left in the wild though. It wouldn't hurt if R600 can support CS 4.0, of course.

As for CS4.x's practicality, yes I agree that if you don't use shared memory (which is fairly restricted but not completely useless) it's probably not very different from just using a pixel shader. However, there are a few applications that may find CS4.x better than pixel shaders: 1. those actually can use shared memory (duh), such as the n-body demo in the SDK, or the flock demo which was written for CUDA (should be able to port to CS4.x); 2. those non-3D related applications, using CS is beneficial as one doesn't have to handle 3D codes which is not needed in these applications.

 

I'd be amazed if AMD supports CS for GPUs older than RV770. Shared memory isn't part of the older architectures and would need to be emulated using video memory.

It's a bit weird that NVidia's support for CS is so patchy though.

Jawed

 

http://forums.amd.com/devforum/messageview.cfm?catid=390&threadid=119267&enterthread=y

OpenCL doesn't support R600&RV670 (it took long time to say this).
No compute shader support -> no DirectCompute support.

 

Ah, that's a shame as HD 3xxx actually sold relatively well.

Regards,
SB