Intel Beignet Is Working Out Surprisingly Well For OpenCL On Linux

[opensource]

anyone tried e.g. matrix multiplication or so? you know, something more useful instead of that luxrender.

[liam]

AMD really pisses me off. They've got some amazing GPUs, but they do crap software that can't compile large kernels, continuing to only partially support opencl.

It seems that the best Opencl compiler right now is the Intel cpu one.

Cute, but you're missing the point, fanboy.
Nvidia actually has top flight devs, and in general the manpower, who could write such an implementation.
Nvidia is utterly blatent about just capturing users. That's what bothers me, and why I'm confused as to why they have their defenders in the linux community.

[89c51]

Is it possible to use this in parrallel with a radeon card? Ie have the radeon do the graphics and the intel 3770 gpu to handle the opencl stuff??

[kalrish]

Is it possible to use this in parrallel with a radeon card? Ie have the radeon do the graphics and the intel 3770 gpu to handle the opencl stuff??

Yes. It's a possible application for gaming: let the iGPU, which is less powerful to render but enough for other things, and does not require slow bus transfers, do general computation (e.g.: physics), which are nowadays not done using OpenCL because of such transfers, among other reasons, and the dGPU render. I believe AMD's HSA helps here, but I'm not sure how (avoids some transfers...?). Thanks to the ICD model, which, AFAIK, both Beignet and Mesa support, you could even switch the OpenCL providers.

[schmidtbag]

I was quite surprised, but when I ran SmallLuxGPU (the engine used in LuxMark) on my Intel i5 4440 with a HD4600 iGPU, the iGPU rendered faster than the CPU, and together it was (naturally) twice as fast as just using the CPU, which I was quite happy with.

If OpenCL support is perfected to work on more complex software, that basically turned my cpu from a quad to an 8 core

Imagine what you'd be missing out on using an APU. Keep in mind an A10-7850 has 2.4 million transistors - that is ONE MILLION more than the i7-4770K, and about half of the entire chip (including cache, CPU cores, memory controller, NB, etc) is just GPU. If AMD can get OpenCL fully functional on APUs, they're going to have something that seriously gives Intel processors a run for their money. But, until that happens, or, until we get more applications that'll actually take advantage of openCL, intel is still a better overall product.

[Sdar]

Cute, but you're missing the point, fanboy.
Nvidia actually has top flight devs, and in general the manpower, who could write such an implementation.
Nvidia is utterly blatent about just capturing users. That's what bothers me, and why I'm confused as to why they have their defenders in the linux community.

I'm an AMD user, but I'll probably go to the green side soon...

Anyway what's your point? AMD can't be blamed for not having any GPGPU implementation with true function calls support because they don't have "top flight devs"? that's your point?
Yeah nvidia Opencl is slower but it compiles big kernels flawlessly, Intel Opencl on CPU works great too but AMD compiler has problems on both GPUs and CPUs and it's a well known bug since 2010~2011 so there's no excuse.

Physx is just middleware that uses Cuda so it's only working with nvidia cards but it's just the same with MANTLE that don't works even with older amd cards nor with any other vendor.

For game developers and professional software it's usually recommended to use nvidia if you need a good Opengl 4.x experience so this is one of the main reasons of why they have their defenders in the linux community.

[zeealpal]

Imagine what you'd be missing out on using an APU. Keep in mind an A10-7850 has 2.4 million transistors - that is ONE MILLION more than the i7-4770K, and about half of the entire chip (including cache, CPU cores, memory controller, NB, etc) is just GPU. If AMD can get OpenCL fully functional on APUs, they're going to have something that seriously gives Intel processors a run for their money. But, until that happens, or, until we get more applications that'll actually take advantage of openCL, intel is still a better overall product.

Well I can also use my AMD HD 7970, which is faster again than both, but this is where I hit the problems of different OpenCL compilers. As I create more difficult/complex scenes (LuxRender only compiles OpenCL code for what a scene needs) it will fail to compile on various platforms in this order:

1- Intel iGPU (HD4600 in my case)
2- AMD GPU (HD 7970)
3- AMD OpenCL CPU (i5 4440)

And the Intel OpenCL CPU has not had any compile errors for me.

I can't comment on the nVidia OpenCL implementation, but I have heard good things about its stability, but not so much about its performance.


Also AMD's OpenCL CPU compiler fails to compile LuxRender when all features are enabled, which is really sad since it is compiling for CPU (and the C++ version compiles fine) and CPU's have a lot less hardware limitations than GPU's.
Intels OpenCL CPU compiler took 4.5 seconds, AMD's OpenCL CPU wasn't done after 15 mins
See here: http://www.luxrender.net/forum/viewt...p?f=34&t=11009

[89c51]

Yes. It's a possible application for gaming: let the iGPU, which is less powerful to render but enough for other things, and does not require slow bus transfers, do general computation (e.g.: physics), which are nowadays not done using OpenCL because of such transfers, among other reasons, and the dGPU render. I believe AMD's HSA helps here, but I'm not sure how (avoids some transfers...?). Thanks to the ICD model, which, AFAIK, both Beignet and Mesa support, you could even switch the OpenCL providers.

Is there documentation on how to set up something like this. And is there a way of explicitly setting each gpu for specific tasks. (ie graphics = use discrete OpenCl = use integrated)

[kalrish]

Is there documentation on how to set up something like this. And is there a way of explicitly setting each gpu for specific tasks. (ie graphics = use discrete OpenCl = use integrated)

The OpenCL API allows to retrieve the available compute devices at runtime and to choose one of them [for each context], so it's up to each program to provide a way for the user to set up its particular configuration. These options should be in the configuration section. However, you should first ensure that your particular program indeed supports OpenCL.

As for graphics, this configuration is outside application scope, because neither of the context-management APIs (EGL, GLX,...) provides methods to choose a provider (i.e.: there's no ICD model in OpenGL). It may be done with PRIME, with which the OpenGL provider may be chosen on a per-application basis (and set as default?). There's more information in the relevant page of the ArchWiki.

[89c51]

One more question. Beignet needs some X stuff. Are they planning to introduce support for wayland at some point.