Archives Discussions

[reillc01]

I just upgraded to Adobe CS6. The Mercury playback engine and most of the GPU accelerated effects are written for Cuda enabled cards (NVIDIA). I have been using a Radeon HD 6950 and have been happy with the performance and features, but GPU acceleration is a "must have feature" for me. Does anybody know if there are plans to implement the newly open-sourced CUDA code on the Radeon series cards? Any idea of a timeline? I'm starting to feel like buying a new Nvidia card might be the only option.

thanks!

[kcarney]

Hello,

I sent your question to some folks at AMD, and this is the reply that I received.

"This is currently true for Premiere Pro - we only have OpenCL for 15 inch MacBook Pro with Radeon 67{5,7}0 with 1 GB. We are working hard to change this in an upcoming release cycle.

But for Photoshop, you can use GPU accelerated effects with OpenCL enabled cards (AMD). Also AMD A-series APUs support GPU acceleration in Photoshop (we're working on getting E series Brazos as well, builds without GPU VRAM/bandwidth checks show 3-4X speed up on low end APUs)."

Cheers!

Kristen

[reillc01]

Thanks for the info. I (and zillions of other upgrading cs users) do hope

that AMD and Adobe can get premiere pro optimized for a larger variety of

graphics cards. I'll hold off buying a Nvidia card this week as long as a

fix is in the works. I'll stay posted for any information about the

expected release date.

[kcarney]

I really appreciate you taking the time to let us know about this issue. I'll make sure that your comments get to the right people.

Cheers!

Kristen

[yurtesen]

Here is a blog entry which you might find interesting:

http://blogs.adobe.com/premiereprotraining/2012/05/opencl-and-premiere-pro-cs6.html

[dongateley]

So, cutting through the doubletalk it looks like the answer is, "no."  That's a real shame because CUDA is being used for all kinds of computational work including DSP and audio apps and plugins.

[yurtesen]

dongateley wrote:

So, cutting through the doubletalk it looks like the answer is, "no."  That's a real shame because CUDA is being used for all kinds of computational work including DSP and audio apps and plugins.

OpenCL is also used. Actually, it is a shame that some programs use CUDA. Because if they used OpenCL, they could run on Nvidia, ATI, Intel(including new intel accelerators) hardware  and soon even on FPGA's and ARM devices. Now they are restricted to Nvidia GPUs only. But this is quickly changing, more and more programs are switching to OpenCL. In addition AMDs GCN GPUs in general perform better than latest Nvidia offerings in GPGPU loads, so programs using CUDA is not able to take advantage of this either...

[dongateley]

Thanks for the info. I'll pass that on to the developers of the audio plugin I want to use.

[yurtesen]

Thanks for taking time to report this issue to your plugin developers. The more people asking, the quicker it gets implemented (also if you have CUDA code, it is relatively easy to port it to OpenCL).

[dongateley]

Just to be sure, does OpenCL support all the heavy duty (vector) computational functions of CUDA that people are exploiting outside the domain of graphics?  The application I want to use exploits CUDA to perform FFT based real time audio stream convolution.  Will the developer find OpenCL equivalent to CUDA for that purpose?

If it is, the developers of FFTW might be interested in using it, when present, to make their FFT library even faster than it currently is.  If they aren't already.

[yurtesen]

The simple answer is that you can think OpenCL and CUDA are basically the same thing. The difference is that OpenCL is an open standart and supported by more than one company, while CUDA is proprietary framework from Nvidia and only work on Nvidia products. (While OpenCL also works equally good on Nvidia products)

I think the well-known fftw.org's library itself does not support GPU at all (I dont know if they want to do that?). However there are several open/closed source libraries which can accomplish FFT on GPUs. AMD provides an OpenCL FFT library freely as well:

http://developer.amd.com/tools/heterogeneous-computing/amd-accelerated-parallel-processing-math-libr...

Make a web search for 'opencl fft'

As a coincidence, if you open the wikipedia article for OpenCL the code examples are about doing FFTW calculations:

http://en.wikipedia.org/wiki/OpenCL

(and there is a nice list of products you can run your OpenCL code too ).

[dongateley]

Many thanks for the info.  From it I was able to put together a nice bookmark folder for future reference.  Now to have a dialog with my app vendor who is currently using CUDA.

[yurtesen]

reillc01 wrote:

Does anybody know if there are plans to implement the newly open-sourced CUDA code on the Radeon series cards? Any idea of a timeline? I'm starting to feel like buying a new Nvidia card might be the only option.

By the way, Nvidia just added CUDA backend support for open-source LLVM which allows it to build binary for Nvidia GPUs. This is simply Nvidia protecting the CUDA monopoly. Otherwise companies which use LLVM would end up using OpenCL which would benefit everybody equally.