Adobe and AMD Enable Brilliant Experiences

[iCyborg]

Originally Posted by ShintaiDK

Thats very misleading. Since the GPUs FP units can barely handle 5% of the code that the CPUs FP units can. Hence also why its so relatively easy to make fast. The bar is simply low enough.

This is also why GPGPU technologies like CUDA and OpenCL is a niche and will always be so.

I don't understand this. What is "code" here? What kind of FP computations can you do on a CPU, but cannot do via CUDA/OpenCL on a GPU, performance considerations aside?

I don't see how it's misleading: QuickSync doesn't try nor purport to do anything general, it's only for video transcoding. Unlike stream processors, even if they are often a very poor choice.

[Lonbjerg]

Originally Posted by ShintaiDK

Wierd isnt it. When Llano already supports it?

Must be a slow newsday....

It is...IVB launch is over.

BTW:
Jeg ved hvem du er
J..... ....

HLJ
FR


.oO(Ellers er sammentræffet ret griner...)

[ShintaiDK]

Originally Posted by Lonbjerg

It is...IVB launch is over.

BTW:
Jeg ved hvem du er
J..... ....

HLJ
FR


.oO(Ellers er sammentræffet ret griner...)

Haha Hej

[Lonbjerg]

Originally Posted by ShintaiDK

Haha Hej

Welcome to ATF

[ShintaiDK]

Originally Posted by iCyborg

I don't understand this. What is "code" here? What kind of FP computations can you do on a CPU, but cannot do via CUDA/OpenCL on a GPU, performance considerations aside?

I don't see how it's misleading: QuickSync doesn't try nor purport to do anything general, it's only for video transcoding. Unlike stream processors, even if they are often a very poor choice.

The CPU can do everything. This is basicly also what makes it "slow". Where the GPU is very limited in the abilities and what instructions it can execute. There is also another issue. 1 flop aint equal to 1 flop either. The GPU might need 4 flops to accomplish what the CPU handles in 1 flop.

Take a look on the OpenCL programming guide for nVidia:
http://www.nvidia.com/content/cudazo...mmingGuide.pdf

Then compare it to a regular one for x86/x64.

[Cerb]

Originally Posted by Phynaz

So any OpenCL capable GPU. Hardly an AMD coup.

Very good for AMD, but no, not nearly a coup. Adobe even said awhile back that they went with CUDA because there was nothing else, and would use OpenCL for future versions.

[iCyborg]

Originally Posted by ShintaiDK

The CPU can do everything. This is basicly also what makes it "slow". Where the GPU is very limited in the abilities and what instructions it can execute. There is also another issue. 1 flop aint equal to 1 flop either. The GPU might need 4 flops to accomplish what the CPU handles in 1 flop.

Take a look on the OpenCL programming guide for nVidia:
http://www.nvidia.com/content/cudazo...mmingGuide.pdf

Then compare it to a regular one for x86/x64.

Instruction set has nothing to do with this. Or does this mean Pentium 1 is not "general purpose" any more, because it lacks many instructions that IB/Bulldozer have? Or ARM? Also, GPUs have some instructions that not all CPUs have as of now, like FMA.

The point is that stream processors in GPUs are designed so they can also be used for many different purposes. In a Turing-complete sense, GPUs are equivalent to CPUs in terms of computability, i.e. they too can do everything. QuickSync was designed for one purpose and is not a step forward in GPGPU. It's a step forward in video transcoding.

[RavenSEAL]

Originally Posted by Olikan

it's just some cuda-related stuff... nothing really important here (exept for another bad news for nvidia)

Some people use "cuda-stuff", really expensive "cuda-stuff". Can't wait to see what my 6870s can in Premiere Pro once my company upgrades to CS6.

[nismotigerwvu]

Originally Posted by Lonbjerg

Well...then it's time to benchmark CUDA vs OpenCL in Adobe...and get the full picture.

Because that is what matters right?

Performance.

....and glide offered better performance than opengl and directx....As long as performance is "good enough" getting this out to the full market rather than just one vendor will be a big step forward.

[ShintaiDK]

Originally Posted by iCyborg

Instruction set has nothing to do with this. Or does this mean Pentium 1 is not "general purpose" any more, because it lacks many instructions that IB/Bulldozer have? Or ARM? Also, GPUs have some instructions that not all CPUs have as of now, like FMA.

The point is that stream processors in GPUs are designed so they can also be used for many different purposes. In a Turing-complete sense, GPUs are equivalent to CPUs in terms of computability, i.e. they too can do everything. QuickSync was designed for one purpose and is not a step forward in GPGPU. It's a step forward in video transcoding.

I think you should read the programming guide.

[iCyborg]

Originally Posted by ShintaiDK

I think you should read the programming guide.

That's your argument? Go read a 60-page document?
Here's a counterargument: I think you should read the x86/x64 guide:
http://www.intel.com/content/www/us/...r-manuals.html

[denev2004]

Originally Posted by ShintaiDK

The CPU can do everything. This is basicly also what makes it "slow". Where the GPU is very limited in the abilities and what instructions it can execute. There is also another issue. 1 flop aint equal to 1 flop either. The GPU might need 4 flops to accomplish what the CPU handles in 1 flop.

Take a look on the OpenCL programming guide for nVidia:
http://www.nvidia.com/content/cudazo...mmingGuide.pdf

Then compare it to a regular one for x86/x64.

This is merely an old programming guide...The newest should be version 4.0
BTW version 2.3 was..I guess for GT200.

[denev2004]

And...I do wonder only on the arithmetic level whether there is something GPU can't "do".
At least we get Taylor Series don't we guys?
Well, efficiency & speed must be a problem.
If I remember it correctly the INT shift instruction in GK104 runs about 1/6 of the FP SP speed

[Lonbjerg]

Originally Posted by nismotigerwvu

....and glide offered better performance than opengl and directx....As long as performance is "good enough" getting this out to the full market rather than just one vendor will be a big step forward.

Glide failed to to 3Dfx's hardware circus, that took out not only glide...but also VooDoo graphics chips...nice fail

[nismotigerwvu]

Actually glide started losing it's grip on the tail end of the Voodoo 2 era, which was unquestionably the fastest card still at the time. Nice attempted troll though.

[nismotigerwvu]

Go ahead and flip through http://www.anandtech.com/show/197 to refresh your memory

[AtenRa]

Heh, it only works on MAC OS at the moment.

http://icrontic.com/article/adobe-pr...port-conundrum

[Bruceee]

I found this thread, http://forums.adobe.com/thread/994892 on the Adobe site. Here is the reason OpenCL isn't implemented yet.


"David McGavran here, engineering manager for Premiere Pro. You can stop speculating. Porting MPE to OpenCL was hard, very hard! It took an entire cycle to get it to work. It isn't just about the engineering effort it is about the testing effort and making sure it will work. We worked very closely with partners to make this happen. We focused on the mac first because there aren't any NVidia cards in the apple store configurations for our customers here. We focused on the macbooks because many of our customers use these products. We developed the OpenCL code so that it can in the future possibly work on windows. We are very interested in making sure our product uses all the compute power that your computer has. No conspiracy theories just time and effort and balancing all the work we need to get done so that you can have the product you want to use to make great video.



Hope that helps



Dave"

I'll wait a while before I select a gpu. I would guess at cs6.5 there's an openCL for windows.

[Ferzerp]

10.3 years, 1 post.


nice.

[GlacierFreeze]

He's wins the Ultimate Lurker Award. lol

[Fox5]

Originally Posted by grimpr

AMDs APP SDK is seriously ahead of Intels software stack. It even supports AVX in compatible CPUs.

I thin the interesting thing here is that the Intel cpus are so competitive with basically everything except AMD's most recent cards, in a task that's extremely parallelizable.