Beignet 1.3 Released With OpenCL 2.0 Support

[phoronix]

Phoronix: Beignet 1.3 Released With OpenCL 2.0 Support

Intel developers today announced the release of Beignet 1.3 and it's by far their most significant release yet for this open-source OpenCL implementation for Intel graphics hardware...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-Beignet-1.3-Released

[darkbasic]

Ouch, my laptop is a Broadwell one

[ryszardzonk]

Making OpenCL great again!

Sorry, I could not resist

[Marc Driftmeyer]

This is an embarrassment to AMD for not reaching this milestone long before hand. I don't give a crap about Nvidia as they will push CUDA until the last drooling idiot goes bankrupt.

[eydee]

They should rename it. "Beginet" sounds like it has something to do with "net", internet for beginners. Some basic package of an ISP with low download speed and low cost.

[illwieckz]

This is an embarrassment to AMD for not reaching this milestone long before hand.

AMD proprietary OpenCL library works very well, is powerful, featureful (2.0 was reached a long time ago), solid and efficient. There is no need for proprietary kernel driver to run it (it just needs the vanilla amdgpu driver). It's a very viable option today. Buying AMD GPU for OpenCL on Linux is the best option today : you can use a vanilla kernel driver and the proprietary OpenCL library while waiting for this OpenCL library to be opened, and since it's that library that will be opened in the future, there is no big issue to rely it today and base our workflow on it. When this code will be opened, the only change will be the legal status. But it's already usable and powerful. I'm just waiting for the relicensing, but I use it happily and it works very well. I have no one issue with it. AMD OpenCL stack is just the better stack on Linux : you don't have to give money to people that have no plan to opening stuff, you don't have to mess with proprietary kernel drivers recompilation and versionning issue, you get powerful and featureful OpenCL stack, and in some time you will get openness! I'm using it with an R9 390X for darktable stuff and I'm very happy with it!

Regarding this beignet release, I just tested that beignet release with luxmark on Haswell, no one pixel was computed at all. Well… the proprietary intel OpenCL stack still works better currently, but unfortunately the open one seems to share nothing with the proprietary one, we have to wait a complete rewrite!

Regarding Nvidia, I have an optimus laptop with an Intel Haswell i7-4810MQ CPU/GPU and a Quadro K1100M discrete GPU, if I benchmark it with luxmark, these are the stacks sorted from the faster to the slower. I also added results from my workstation featuring an AMD FX-5950 CPU and an AMD R9 390X discrete GPU for comparison. I'm not comparing the performance between the laptop and the workstation, but I'm comparing each stack on each hardware when the hardware is the same.

- [x] multidevice AMD proprietary OpenCL (amdgpu-pro) on AMD FX-5950 CPU + AMD proprietary OpenCL (amdgpu-pro) on discrete R9 390X GPU [score: 19606]
- [x] AMD proprietary OpenCL (amdgpu-pro) on discrete R9 390X GPU [score: 18356]
- [x] C++ stack on AMD FX 9590 CPU [score: 2513]
- [x] multidevice Intel proprietary OpenCL on Intel i7-4810MQ CPU + Intel proprietary OpenCL on integrated Intel Haswell GPU + Nvidia proprietary OpenCL on discrete Quadro K1100M GPU [score: 2295]
- [x] C++ stack on Intel i7-4810MQ CPU [score: 1983]
- [x] Intel proprietary OpenCL on Intel i7-4810MQ CPU [score: 1615]
- [x] Intel proprietary OpenCL on integrated Intel Haswell GPU [score: 1608]
- [x] AMD proprietary OpenCL (amdgpu-pro) on discrete AMD FX 5950 CPU [score: 1446]
- [x] AMD proprietary OpenCL (amdgpu-pro) on Intel i7-4810MQ CPU [score: 1315]
- [x] Nvidia proprietary OpenCL on discrete Quadro K1100M GPU [score: 980]
- [_] AMD proprietary OpenCL (rocm) on discrete R9 390X GPU (hardware currently unsupported)
- [_] Intel all-open OpenCL (beignet) on integrated Haswell GPU (faulty, nothing is rendered at all)
- [_] Mesa all-open amdgcn libclc on discrete R9 390X GPU (kernel compilation error, missing features)
- [_] pocl all-open on AMD FX 5950 CPU (luxmark abort)
- [_] pocl all-open on Intel i7-4810MQ CPU (luxmark abort)

Yes, even on my laptop, even the AMD OpenCL stack running on an Intel CPU is 34% faster than the Nvidia stack running on the discrete GPU. The results above are for the LuxBall scene, for the hotel one, the nVidia stack is so slow the result can be validated and the benchmark is a failure. And to get that shitty result, I have to enter the proprietary driver hell and sell my soul to the evil. Did I say the only way to get a working external monitor on this laptop is to feed the display with the Intel driver? Nvidia on Linux is just pointless. Have you already seen those Nvidia CUDA stickers on Laptops as a selling argument? Even CPU is faster.

[illwieckz]

They should rename it. "Beginet" sounds like it has something to do with "net", internet for beginners. Some basic package of an ISP with low download speed and low cost.

It's not “Beginet” it’s “Beignet”, the “i” is before the “g”. See Wikipedia article: https://en.wikipedia.org/wiki/Beignet
Pronounce it “Benyey”, with “Be” like in “Betty”, “ny” like in “nyan cat” and “ey” like in “they”
It's a pastry, there is many kind of beignets: dougnuts are one kind of beignet for example.
 

[ariel]

Hoping that the improved opencl support in beignet will allow Darktable to finally enable support for it at least on skylake, and give it a performance boost.

[darkblu]

So SNB, IVB, HSW and BDW can go suck a lemon.

[Linuxhippy]

AMD proprietary OpenCL library works very well, is powerful, featureful (2.0 was reached a long time ago), solid and efficient.

Only Catalyst supported OpenCL-2, AMDGPU-Pro as well as the RockM stuff are at OpenCL-1.2 (well, except for the kernel language, but the really interesting OpenCL-2 changes are in the runtime - SVM, pipes and the like).

you can use a vanilla kernel driver and the proprietary OpenCL library while waiting for this OpenCL library to be opened, and since it's that library that will be opened in the future

I am usually a big fan of AMDs open-source efforts, but when it comes to OpenCL on Linux - there have been a lot of promises but no working code (same as with Vulkan).
Beignet 1.3 supports OpenCL-2 today, yes there may be bugs, but that will be fixed.

When this code will be opened, the only change will be the legal status.

From what I understand, AMD does not plan to open-source their Catalyst OpenCL-2 implementation - but instead the LLVM based RocM approach - which is a clean implementation, and shows completly different characteristics (just have a look at the phoronix benchmarks).


I really hope AMD will provide a useable and fully-fledged open-source OpenCL-2 implementation soon as promised ... I've been waiting for that to happen for years.