ATI Stream SDK V2.0 | OpenCL Full CPU & GPU Support | Introduction

Printable View

10-13-2009, 11:09 AM
64NOMIS

ATI Stream SDK V2.0 | OpenCL Full CPU & GPU Support | Introduction

The ATI Stream Team is proud to make available the fourth beta release of ATI Stream SDK v2.0 that provides the first complete OpenCL™ development platform. This release is certified fully compliant with OpenCL 1.0 by the Khronos Group and is supported on a wide range of AMD GPUs as well as any x86 multi-core CPU supporting SSE3. AMD offers the market both high-performance CPU and GPU technology, and as such we are delivering on this unique ability to provide an OpenCL platform that enables developers to create applications that run the way they were meant to be run, on all the available processors in the system! The beta is available for immediate download as part of our ATI Stream SDK beta program and we encourage you to take a look. For an introduction, see Simon Solotko on the impact of open, paralell computing.
Dealing With Reality | The Introduction | ATI Stream Technology and OpenCL | Part 1 by Simon Solotko

And for anyone who has writen in C or C++, I thought there was a gap between some of the deep dive technical material and the glossy material so I interviewed AMD's Ben Sander and we explain how OpenCL works and some of the basic development methodologies.
Dealing With Reality | The Introduction | ATI Stream Technology and OpenCL | Part 2 by Simon Solotko

Edit - Good discussion point that explains how some of this works:

Quote:

Originally Posted by Solus Corvus

After looking at the documentation it seems that I was correct. You shouldn't have to recompile for other architectures because OpenCL enumerates and utilizes resources (GPU/CPU/Cell/Etc) at runtime. A compiled OpenCL Kernel should be able to run on either company's hardware. There is a way to include what they call Native Kernels in your program. They are hardware specific - but they also aren't standard and being able to execute native kernels apparently isn't even a requirement in the standard.

Brace yourselves. The Kernel is not compiled until just prior to execution (hardware-specific native Kernels seem unlikely in practice since the whole idea is to be hardware independent but the HPC guys might do this) when it is directed to be built/compiled by the application. The compilation step is hardware dependent; if the application is executing on a platform with an x86 processor and an ATI GPU the Kernel will have either path available for compilation, potentially even determined at runtime! So if the GPU is busy, your applications could complile the Kernel for the CPU, execute it, and at a later time, compile it for the GPU, and execute it! You could hold an iner-application race, compiling separate instances for both, one for x86 and one for the GPU, and running them concurrently! I describe some of this in the interview: http://links.amd.com/openinterview
10-13-2009, 11:24 AM
netkas

So, opencl support only for radeon gpus with compute shaders, its 4 series and higher.
10-13-2009, 11:40 AM
zalbard

But something made in OpenCL should still work with Geforce, right?
10-13-2009, 12:01 PM
flippin_waffles

You know, I just now realised the huge advantage AMD has here. When AMD says they are the only ones with both CPU and GPU, it's more than just marketing talk. Once DirectX11 and OpenCL release, AMD's platform releases with both, this should be the platform of choice for developers. Now that we have this announcement, all that is left is the Win7 launch I guess.
10-13-2009, 12:03 PM
Final8ty

Quote:

"ATI Radeon HD 5700 Series" = ati2mtag_Evergreen, PCI\VEN_1002&DEV_68B8
"ATI Radeon HD 5700 Series " = ati2mtag_Evergreen, PCI\VEN_1002&DEV_68B9
"ATI Radeon HD 5700 Series " = ati2mtag_Evergreen, PCI\VEN_1002&DEV_68BE
"ATI Radeon HD 5800 Series" = ati2mtag_Evergreen, PCI\VEN_1002&DEV_6898
"ATI Radeon HD 5800 Series " = ati2mtag_Evergreen, PCI\VEN_1002&DEV_6899
"ATI Radeon HD 5900 Series" = ati2mtag_Evergreen, PCI\VEN_1002&DEV_689C
"ATI Radeon HD 5900 Series " = ati2mtag_Evergreen, PCI\VEN_1002&DEV_689D

inf's
10-13-2009, 12:22 PM
NBF

Its very nice to be able to test CPU and GPU performance and check which one runs better for the specific code. :up:
10-13-2009, 12:50 PM
Manicdan

Quote:

"ATI Radeon HD 5900 Series" = ati2mtag_Evergreen, PCI\VEN_1002&DEV_689C
"ATI Radeon HD 5900 Series " = ati2mtag_Evergreen, PCI\VEN_1002&DEV_689D

is there a 5900 being planned?
10-13-2009, 12:51 PM
64NOMIS

Here is the link for the Stream SDK V2.0 - lots of info and the download.

http://developer.amd.com/GPU/ATISTRE...s/default.aspx
10-13-2009, 12:53 PM
SKYMTL

Quote:

Originally Posted by zalbard

But something made in OpenCL should still work with Geforce, right?

Yes. That's the whole point of having an open platform.
10-13-2009, 12:55 PM
zalbard

Quote:

Originally Posted by SKYMTL

Yes. That's the whole point of having an open platform.

Just making sure that it is the case for this SDK in particular, you never know lol.
10-13-2009, 12:59 PM
Chumbucket843

Quote:

Originally Posted by zalbard

But something made in OpenCL should still work with Geforce, right?

you still have to port it. its not like you compile it twice and you are done. the question is "will it run well on both gpu's". opencl is a lot like cuda but a bit more abstract so it can be run on more architectures.
10-13-2009, 01:10 PM
Solus Corvus

About time ATI. Now let's see some OpenCL benchmark programs made.

Quote:

Originally Posted by Chumbucket843

you still have to port it. its not like you compile it twice and you are done. the question is "will it run well on both gpu's". opencl is a lot like cuda but a bit more abstract so it can be run on more architectures.

Are you sure about this? If so it's the worst news I have heard in a while. I thought the whole point of OpenCL was that you would get a single binary compatible with multiple GPU architectures. If you have to port the code to each GPU and compile a new binary for each that would pretty much ruin it, imo. Now I'm going to have to go do some research to make sure that you are wrong, lol.
10-13-2009, 01:11 PM
netkas

Developers still can make opencl code nvidia-only or ati-only, for example, in SnowLeopard, some of demos from Apple and Nvidia runs on Radeons, some doesnt, because of errors.

or can be blocked by vendor-id, just like in batman aa.
10-13-2009, 01:21 PM
Solus Corvus

I don't doubt that it has that capability. Any robust language/API should let you make exclusive cases like that, if you choose to do so.

My problem would be if there isn't the ability for binary compatibility across multiple GPU architectures/vendors. If it HAS to be like: ATI exe, NV exe, larrabee exe, CPU exe, etc. Then half the point of OpenCL just failed before it even started.
10-13-2009, 01:35 PM
netkas

if opencl kernel code exists in source form in app- then it should be almost compatible, but if it exists in binary form for some specific target device, then, you know the answer.
10-13-2009, 01:52 PM
Solus Corvus

After looking at the documentation it seems that I was correct. You shouldn't have to recompile for other architectures because OpenCL enumerates and utilizes resources (GPU/CPU/Cell/Etc) at runtime. A compiled OpenCL Kernel should be able to run on either company's hardware. There is a way to include what they call Native Kernels in your program. They are hardware specific - but they also aren't standard and being able to execute native kernels apparently isn't even a requirement in the standard.

Some information:
http://www.khronos.org/registry/cl/s...ncl-1.0.43.pdf
http://arstechnica.com/apple/reviews...-x-10-6.ars/15
10-13-2009, 02:05 PM
Eson

Sigh... I just don't seem to have any talent in parallel/stream programing. Ive looked at Cuda, CTM, OpenCL and C-based hardware description languages such as Verilog, Mitrion C and Impulse C. Many claim they use ANSI C. But i just don't have any clue in any of their 'easiest' examples.

Why aren't there never any 'Hello World' examples? Is it because i don't know C++ that i never understand anything?
I do know a lot of C and Java.
10-13-2009, 02:07 PM
Solus Corvus

"Hello World" isn't very parallel. :p:
10-13-2009, 02:08 PM
RaZz!

Quote:

Originally Posted by Manicdan

is there a 5900 being planned?

i'm wondering as well, as there wasn't a x900 before with the 3000 and 4000 series, only a 4890 :confused:
10-13-2009, 02:14 PM
Serra

Quote:

Originally Posted by Eson

Sigh... I just don't seem to have any talent in parallel/stream programing. Ive looked at Cuda, CTM, OpenCL and C-based hardware description languages such as Verilog, Mitrion C and Impulse C. Many claim they use ANSI C. But i just don't have any clue in any of their 'easiest' examples.

Why aren't there never any 'Hello World' examples? Is it because i don't know C++ that i never understand anything?
I do know a lot of C and Java.

Well... here's a hello world program for CUDA. Only issue is it's got to be 100+ lines long... not exactly what I would call intuitive.

I can get around basically in C++, but I don't think I understand more than a few lines of that. Wish I did though, I always thought it would be fun to know how to use CUDA.
10-13-2009, 02:18 PM
Eson

OK now im crying. Did they make a multithreaded 'Hello World' or is all that just to get a 'Hello World' printed out on the screen?
10-13-2009, 02:42 PM
Solus Corvus

That's not too bad. Most of that's just setup, the actual kernel isn't very big.
10-13-2009, 03:56 PM
EniGmA1987

Quote:

Originally Posted by Solus Corvus

"Hello World" isn't very parallel. :p:

Ah but wouldnt it be fun to see just how fast you could process that command with 8 cores or more :D
10-13-2009, 04:10 PM
Pontos

No HD3000 series support? :confused:

Why is that? they support Compute Shaders, just like the HD4000 series cards...
10-13-2009, 04:26 PM
Solus Corvus

Also from the links provided so far: http://developer.amd.com/gpu/ATIStre...ialOpenCL.aspx
10-13-2009, 07:40 PM
64NOMIS

Quote:

Originally Posted by Solus Corvus

After looking at the documentation it seems that I was correct. You shouldn't have to recompile for other architectures because OpenCL enumerates and utilizes resources (GPU/CPU/Cell/Etc) at runtime. A compiled OpenCL Kernel should be able to run on either company's hardware. There is a way to include what they call Native Kernels in your program. They are hardware specific - but they also aren't standard and being able to execute native kernels apparently isn't even a requirement in the standard.

Brace yourselves. The Kernel is not compiled until just prior to execution (hardware-specific native Kernels seem unlikely in practice since the whole idea is to be hardware independent but the HPC guys might do this) when it is directed to be built/compiled by the application. The compilation step is hardware dependent; if the application is executing on a platform with an x86 processor and an ATI GPU the Kernel will have either path available for compilation, potentially even determined at runtime! So if the GPU is busy, your applications could complile the Kernel for the CPU, execute it, and at a later time, compile it for the GPU, and execute it! You could hold an iner-application race, compiling separate instances for both, one for x86 and one for the GPU, and running them concurrently! I describe some of this in the interview: http://links.amd.com/openinterview
10-14-2009, 12:28 AM
fellix_bg

Quote:

Originally Posted by Pontos

No HD3000 series support? :confused:

Why is that? they support Compute Shaders, just like the HD4000 series cards...

R600 architecture does not support LDS memory.
10-14-2009, 04:59 AM
Marios

I can't see why a developer might use opencl to accelerate a program for the CPU and not program directly in a multi core multi thread way.
This way there is no need to use opencl runtime environment to generate binaries on the fly. I suppose this is a time consuming task.
There is no word from Intel about opencl support for their cpus. Why?
Is it because Larrabee as a "cpu", can be programed in a straightforward way to support GPGPU (physx etc) and doesn't really need opencl?
Intel owns Havok, Nvidia owns cuda Physx and promotes it with game developers, AMD is coming out with opencl.
Nice standards we have for our PCs.
10-14-2009, 11:40 AM
To(V)bo Co(V)bo

Quote:

Originally Posted by 64NOMIS

Brace yourselves. The Kernel is not compiled until just prior to execution (hardware-specific native Kernels seem unlikely in practice since the whole idea is to be hardware independent but the HPC guys might do this) when it is directed to be built/compiled by the application. The compilation step is hardware dependent; if the application is executing on a platform with an x86 processor and an ATI GPU the Kernel will have either path available for compilation, potentially even determined at runtime! So if the GPU is busy, your applications could complile the Kernel for the CPU, execute it, and at a later time, compile it for the GPU, and execute it! You could hold an iner-application race, compiling separate instances for both, one for x86 and one for the GPU, and running them concurrently! I describe some of this in the interview: http://links.amd.com/openinterview

Think if they wrote AI "artificial intelligence" in parallel. Where you could have a program answer its own questions. If you think about life in general it is all reasoning. If you were able to compute enough possible outcomes to a given problem (or moment) and choose the best answer. You would have a form of intelligence. The main function of life is self-preservetion, apply this together and you got yourself a terminator. I dont think its to far off if your using the compute power these days. If you think about it, it isn't to far fetched. You got to think, modern inventions were all just imagination at one time or another.
10-14-2009, 01:44 PM
Hell Hound

Quote:

Originally Posted by To(V)bo Co(V)bo

Think if they wrote AI "artificial intelligence" in parallel. Where you could have a program answer its own questions.You would have a form of intelligence. The main function of life is self-preservetion, apply this together and you got yourself a terminator.

:rofl::ROTF::up:
10-14-2009, 04:45 PM
To(V)bo Co(V)bo

I bet you skynet can run on opencl, using parallel instructions. A terminator could run on as little as 30 tflops. It could crunch the physics of gun trajectory, distances, and time. Finally a real use for physics.
10-14-2009, 06:54 PM
EniGmA1987

Well hopefully if/when "skynet" takes over, it will use x86 architecture. Because then either it will make x86 into the best computer architecture ever, or it will have so much junk we have a good change of defeating it :)
10-15-2009, 10:47 PM
To(V)bo Co(V)bo

No way possible, a terminator would definately need way more than 4gb of system memory to effectively kill people. I bet they also need Gddr5 bandwith to render complex infra-red video, and not forget last, ray tracing.
10-16-2009, 12:31 AM
Xoulz

Quote:

Originally Posted by To(V)bo Co(V)bo

No way possible, a terminator would definately need way more than 4gb of memory to effectively kill people. I bet they also will need Gddr5 bandwith to render complex infra-red video, and not forget last, ray tracing.

:eek:

Dude, it's widely known that the early terminator models used RAMBUS.

(:hehe:)
10-16-2009, 01:15 PM
To(V)bo Co(V)bo

Your right forgot about those ones with rambus, they didnt use gddr5 till after the patent infringement lawsuit.

Skynet became self-aware after a new malicious folding at home client was released for the gt300 architecture. The new client was able to give skynet access to thousands of computers, and allowed for true real time physics processing without taking a fps hit.
10-17-2009, 05:00 AM
Lightman

Quote:

Originally Posted by To(V)bo Co(V)bo

Your right forgot about those ones with rambus, they didnt use gddr5 till after the patent infringement lawsuit.

Skynet became self-aware after a new malicious folding at home client was released for the gt300 architecture. The new client was able to give skynet access to thousands of computers, and allowed for true real time physics processing without taking a fps hit.

:rofl:

So now we know why 12/2012 will be end of the world ....
10-17-2009, 06:09 AM
Angry

Quote:

Originally Posted by Lightman

:rofl:

So now we know why 12/2012 will be end of the world ....

Oh come on, we all know the world will end when Fermi is released...
12-22-2009, 02:33 PM
zalbard

AMD's ATI Stream SDK 2.0 no longer beta!

Quote:

AMD has decided to give developers an early Christmas gift by releasing the final version of its ATI Stream Software Development Kit 2.0 which brings OpenCL 1.0 support as well as the hardware acceleration for applications.

The final version of the ATI Stream SDK 2.0 features:

* First production release of ATI Stream SDK with OpenCL™ 1.0 support.
* New: Support for OpenCL™ ICD (Installable Client Driver).
* New: Support for atomic functions for 32-bit integers.
* New: Microsoft® Visual Studio® 2008-integrated ATI Stream Profiler performance analysis tool.
* Preview: Support for OpenCL™ / OpenGL® interoperability.
* Preview: Support for OpenCL™ / Microsoft® DirectX® 10 interoperability.
* Preview: Support for double-precision floating point basic arithmetic in OpenCL™ C kernels.
* Updated OpenCL™ runtime to conditionally load ATI CAL runtime libraries to allow execution on compatible CPUs without ATI Catalyst™ installed.
* Updated OpenCL™ runtime to allow simultaneous use of OpenCL™ and ATI CAL APIs in a single user application.
* Updated cl.hpp from the Khronos OpenCL working group release.
* Various OpenCL™ compiler and runtime fixes and enhancements

It supports Windows XP, Vista, 7 as well as openSuse 11.0 and Ubuntu 9.04 operating systems and supports all the graphics cards ranging from Radeon HD 4000, HD 5000, Mobility HD 4000 as well as ATI FireStream 9270 and 9250 and FirePro cards.

You can find it here.

Hm, time to see if there's a performance improvement for benches. :yepp:
12-22-2009, 02:38 PM
zanzabar

Quote:

Originally Posted by zalbard

AMD's ATI Stream SDK 2.0 no longer beta!

Hm, time to see if there's a performance improvement for benches. :yepp:

it benches the dc benchmark 10-15% faster
12-22-2009, 03:38 PM
Boogerlad

still doesn't work with dolphin. *sigh*
12-22-2009, 05:08 PM
Calmatory

... NOW I wish I had Radeon GPU on my laptop. :(

Is this compatible with mobile HD3470?
12-22-2009, 05:13 PM
Nedjo

Quote:

Originally Posted by Calmatory

... NOW I wish I had Radeon GPU on my laptop. :(

Is this compatible with mobile HD3470?

supports all the graphics cards ranging from Radeon HD 4000, HD 5000, Mobility HD 4000 as well as ATI FireStream 9270 and 9250 and FirePro cards
;)
12-22-2009, 05:56 PM
Calmatory

Ah, damn. Thanks. :up:

Guess I'll build a dedicated box for OpenCL stuff then. Easier and cheaper than replacing the current laptop + OS.