CPU is i7-3770 and software is Mint 17 Live CD/USB, kernel 3.13.0-24-generic and Beignet 0.9.2 (64-bit) and 1.0.2 (32-bit, for Seti). Beignet 0.9.2 was compiled with/against llvm 3.4, clang 3.4, gcc 4.8.2 and libdrm 2.4.56.
I had to try several Beignet versions before I found one that works for BRP4. 1.0.1 - 1.0.3 failed with the Illegal pointer error, 0.9.3 and 1.0.0 did run but produced incorrect results. 0.9.2 worked so I didn't bother trying versions 0.9.1 and earlier.
So far the valid task count is about 350. There has been two inconclusives (not counting the ones where wingman is running bad Windows drivers and such). One of them turned in my favor and the other is still waiting for a third opinion.
As for performance, I believe it's on the same level as with Windows drivers, perhaps a bit slower. Comparing with other hosts is not too easy because I don't know how the CPU is loaded in them. For anyone wishing to compare, I have run this test without loading the CPU.
I also tried to run BRP4 on an i7-4790 but that didn't work. If I recall correctly, the outcome was, depending on the version of Beignet, either the Illegal pointer error or segmentation fault. I didn't update or patch the kernel for the i7-4790 test but perhaps I should have.
In summary, based on one week of testing, Beignet looks good and is, in my opinion, ready for testing on more hosts. I've been borrowing these hosts so I'll be leaving further testing to those who own suitable Intel hardware. My interest in this test has been to see if Beignet works on an Intel only machine. The other people who have tested Beignet (here and elsewhere) without success seem to have jumped straight to mixed Intel+AMD/NVIDIA setups.
Any help with that xml file? I'd really like to try it and force the amd binary to my intel gpu, but I don't know how to set the syntax....
First, a quick primer for those who are less familiar with anonymous platform mechanism.
*app_info.xml files go in to project's directory, in this case projects/einstein.phys.uwm.edu/ .
*You only get work for the common subset of the applications listed in app_info.xml and applications you have selected in E@H preferences.
*When you use app_info.xml BOINC will not download any applications from server, not updated versions of applications listed in the file or any other. That means you are responsible for updating the applications.
*BOINC will delete all work you have on board for any application that is not listed in app_info.xml, even if that work is already started or completed and ready to report.
I'm sure I forgot something but don't worry, BOINC will tell you while it trashes all your work :P
Since not all versions of Beignet work for Einstein, I prepared a small test package. This way nobody needs to trash a ton of work in order to find a working version.
Download brp4-igpu-test, extract it somewhere and open terminal there. Run the runtest script (type ./runtest) and wait for the test to finish.
Once the test run has finished there will be stderr.txt and results.cand0 files. Verify that the stderr.txt file looks similar to the ones on result pages on this web site (ignore the lines about init data file and user/host details). The package contains a reference result called results-1.06-rpi that was produced on Raspberry Pi (and has been validated). Compare results.cand0 and results-1.06-rpi using some tool that can show differences between files, preferably one that shows files side-by-side and highlights the differences within the lines. I usually use program called meld. Minor differences in last decimal places should be ok. If the files look very similar, go ahead, start live testing.
If the result files are different, remove all the files the test has created, install another Beignet version and start again.
The package also contains another result file, results-1.52-intel_gpu. You can compare that file with the reference result to see what you can expect from your own test result.
And finally, here's app_info.xml for BRP4 only. If you run any other applications you need to add them to the file. You can download the program file from here. It's also contained in the test package. Remember to set the executable permission.
Yes. There is a bug in the newer driver versions. This is why you will not get tasks unless you install an older driver. You can find links in the posts above.
Yes. There is a bug in the newer driver versions. This is why you will not get tasks unless you install an older driver. You can find links in the posts above.
Thanks for the hint. With an old driver i got plenty of work. (And the results are valid)
After this i upgraded the box to "Windows 10". Win10 has its own driver (10.18.15.4124) and this driver seems to be o.k. The results get validated....and this driver is about 15% faster than the old one!
But now the box gets no new work, so in a few hours it will run dry. Is there any possibility to get new work for Win10? app_info.xml?
Can you help me a little with
)
Can you help me a little with that app_info.xml?
I found this example, but I'm unsure about the values. It's for Cuda only, not for OpenCL.
So they're working on a
)
So they're working on a patch, should be finished by next week. I'm curious if it will work with e@h.
Could a forum moderator move beignet related postings to a separate thread?
Thank you.
Any help with that xml file?
)
Any help with that xml file? I'd really like to try it and force the amd binary to my intel gpu, but I don't know how to set the syntax....
I have just noticed after
)
I have just noticed after getting the AMD drivers updated this in the boinc logs, and clinfo reports the Intel GPU.
[pre]
26-Apr-2015 19:02:57 [---] OpenCL CPU: Intel(R) Core(TM) i5-4690K CPU @ 3.50GHz (OpenCL driver vendor: Advanced Micro Devices, Inc., driver version 1642.5 (sse2,avx), device version OpenCL 1.2 AMD-APP (1642.5))
[/pre]
So am I right in saying AMD have just released a OpenCL driver for Intel GPU for linux?
RE: I have just noticed
)
No, that says the Cental Processing Unit, ie the CPU has OpenCL support via the AMD drivers, no mention of an Intel GPU there.
Claggy
I've got Beignet working on
)
I've got Beignet working on this host.
CPU is i7-3770 and software is Mint 17 Live CD/USB, kernel 3.13.0-24-generic and Beignet 0.9.2 (64-bit) and 1.0.2 (32-bit, for Seti). Beignet 0.9.2 was compiled with/against llvm 3.4, clang 3.4, gcc 4.8.2 and libdrm 2.4.56.
I had to try several Beignet versions before I found one that works for BRP4. 1.0.1 - 1.0.3 failed with the Illegal pointer error, 0.9.3 and 1.0.0 did run but produced incorrect results. 0.9.2 worked so I didn't bother trying versions 0.9.1 and earlier.
So far the valid task count is about 350. There has been two inconclusives (not counting the ones where wingman is running bad Windows drivers and such). One of them turned in my favor and the other is still waiting for a third opinion.
As for performance, I believe it's on the same level as with Windows drivers, perhaps a bit slower. Comparing with other hosts is not too easy because I don't know how the CPU is loaded in them. For anyone wishing to compare, I have run this test without loading the CPU.
I also tried to run BRP4 on an i7-4790 but that didn't work. If I recall correctly, the outcome was, depending on the version of Beignet, either the Illegal pointer error or segmentation fault. I didn't update or patch the kernel for the i7-4790 test but perhaps I should have.
In summary, based on one week of testing, Beignet looks good and is, in my opinion, ready for testing on more hosts. I've been borrowing these hosts so I'll be leaving further testing to those who own suitable Intel hardware. My interest in this test has been to see if Beignet works on an Intel only machine. The other people who have tested Beignet (here and elsewhere) without success seem to have jumped straight to mixed Intel+AMD/NVIDIA setups.
(Same story for Seti is here.)
RE: Any help with that xml
)
First, a quick primer for those who are less familiar with anonymous platform mechanism.
*You only get work for the common subset of the applications listed in app_info.xml and applications you have selected in E@H preferences.
*When you use app_info.xml BOINC will not download any applications from server, not updated versions of applications listed in the file or any other. That means you are responsible for updating the applications.
*BOINC will delete all work you have on board for any application that is not listed in app_info.xml, even if that work is already started or completed and ready to report.
I'm sure I forgot something but don't worry, BOINC will tell you while it trashes all your work :P
Since not all versions of Beignet work for Einstein, I prepared a small test package. This way nobody needs to trash a ton of work in order to find a working version.
Download brp4-igpu-test, extract it somewhere and open terminal there. Run the runtest script (type ./runtest) and wait for the test to finish.
Once the test run has finished there will be stderr.txt and results.cand0 files. Verify that the stderr.txt file looks similar to the ones on result pages on this web site (ignore the lines about init data file and user/host details). The package contains a reference result called results-1.06-rpi that was produced on Raspberry Pi (and has been validated). Compare results.cand0 and results-1.06-rpi using some tool that can show differences between files, preferably one that shows files side-by-side and highlights the differences within the lines. I usually use program called meld. Minor differences in last decimal places should be ok. If the files look very similar, go ahead, start live testing.
If the result files are different, remove all the files the test has created, install another Beignet version and start again.
The package also contains another result file, results-1.52-intel_gpu. You can compare that file with the reference result to see what you can expect from your own test result.
And finally, here's app_info.xml for BRP4 only. If you run any other applications you need to add them to the file. You can download the program file from here. It's also contained in the test package. Remember to set the executable permission.
[pre]
einsteinbinary_BRP4
einsteinbinary_BRP6_1.52_x86_64-pc-linux-gnu__BRP6-Beta-opencl-intel_gpu
einsteinbinary_BRP4
152
7.5.0
intel_gpu
1
opencl-intel_gpu
einsteinbinary_BRP6_1.52_x86_64-pc-linux-gnu__BRP6-Beta-opencl-intel_gpu
[/pre]
I am getting this on a box
)
I am getting this on a box Win7 64bit,Intel GPU 4600:
Does this mean i have to install an older driver?
RE: I am getting this on a
)
Yes. There is a bug in the newer driver versions. This is why you will not get tasks unless you install an older driver. You can find links in the posts above.
RE: Yes. There is a bug in
)
Thanks for the hint. With an old driver i got plenty of work. (And the results are valid)
After this i upgraded the box to "Windows 10". Win10 has its own driver (10.18.15.4124) and this driver seems to be o.k. The results get validated....and this driver is about 15% faster than the old one!
But now the box gets no new work, so in a few hours it will run dry. Is there any possibility to get new work for Win10? app_info.xml?