WARNING: This website is obsolete! Please follow this link to get to the new Albert@Home website!
App for Intel HD graphics? |
Message boards :
Wish List :
App for Intel HD graphics?
Message board moderation
Previous · 1 · 2 · 3 · 4 · Next
Author | Message |
---|---|
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
The task has now been completed and no errors occurred, so looks good and well done! =) Total runtime was 2h21m59s and CPU time ~2m23s. I've zipped the result files and stderr.txt and can e-mail it if your interested. Tomorrow I'll try with the system fully loaded on all 8 threads and the GTX660Ti. |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
Hi! Thanks very much, see your PM box for an address to send the result files to. I think for an integrated GPU the runtime isn't shabby at all. Thanks again! HB |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Tomorrow I'll try with the system fully loaded on all 8 threads and the GTX660Ti. So I just did a short test with the computer otherwise fully loaded with 8 Einstein tasks on the CPU and 2 BRP5 tasks on the GPU, the result was as I suspected from experience with other OpenCL apps. The GPU load reported by GPU-Z dropped to ~20-25% and as soon as I suspended tasks in Boinc to free up one thread the GPU load returned to about 80%, so the conclusion is that this app needs one thread free to run at optimal speed. From comments on other projects' message boards this seems to be a driver issue when one runs OpenCL apps both on Intel and Nvidia GPUs. |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
Hi! Finally I got the BRP app running even under BOINC control on my PC. Because the server software doesn't yet support Intel GPUs (and it's not trivial to add the support in our customized server software), it is a combination of app_info.xml and cc_config.xml file hack. I used this cc_config.xml file (goes into the BOINC data directory):
Where The app_info.xml file (goes into the projects\albert.phys.uwm.edu sub directory):
This assumes you have renamed the downloaded executable to einsteinbinary_BRP4U_1.35_windows_x86_64__opencl-intel_gpu.exe . This will process the unbundled BRP4U workunits to allow speedy tests. The tasks should automatically reserve a full CPU core. On my PC, I now get close to 100% GPU usage this way, according to GPU-Z. So running many of these units in parallel on my HD 2500 GPU would not help. I used the currently recommended BOINC client version for this. Cheers HB |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Nice to see you got it running under Boinc, will definitely make it easier to test and validate! I've just completed the first BRP4U task and it took 18m30s on the HD 4000, the average GPU load according to GPU-Z was 76% and it varies between 75-85%. Link to BRP4U tasks on host #2267 As a note to other testers the <coproc_debug> log flag in cc_config.xml is not needed for this to work, what it does is tell you which coproc Boinc assigns a task to and then at regular intervals it confirms the assignment. As I was composing this the second task completed (18m21s) and validated against the Neon-app! =) Oh and by the way did my first standalone test "validate"? |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
Oh yes, just checked it by diff and eyeball and spotted only insignificant differences. Cheers HB |
Alex Send message Joined: 1 Mar 05 Posts: 88 Credit: 398,734 RAC: 0 |
I have seen that the intel open-cl app is for linux only; are there plans to port that to windows too? Otherwise I will try to install a dual boot system over the weekend. |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
No, the app is for Windows, I'm running it on Win7 x64. Download the app from the post about standalone test but then follow the instructions about using a app_info.xml and cc_config.xml to get it running. Don't forget to rename the .exe or make the appropriate changes to the app_info. |
Bernd Machenschalk Volunteer moderator Project administrator Project developer Send message Joined: 15 Oct 04 Posts: 1956 Credit: 6,218,130 RAC: 0 |
AFAIK Intel doesn't even provide an OpenCL-capable driver for Linux. BM |
Alex Send message Joined: 1 Mar 05 Posts: 88 Credit: 398,734 RAC: 0 |
So this entry on the application page is a fake? Linux running on an AMD x86_64 or Intel EM64T CPU 1.34 (opencl-intel_gpu) 20 Mar 2013 | 10:19:43 UTC Anyway, I get the pc over the weekend and will try that. Thx for the reply! |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
That entry is caused by a deprecated app, it was never used in the project. The only OpenCL driver for Linux that Intel offers is supporting the CPU cores itself (so for multi-threading CPU programs), and the Xeon Phi accelerator card, not for the build-in Intel HD GPUs :-( Cheers HB |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Nice! I've been running this for since yesterday and so far I have 52 valid and 45 pending task. Only one is inconclusive and that's because my wingman got a validate error. So all looks good! Just to test I suspended tasks in Boinc so that only 6 CPU task where running and check GPU-Z and the load on the GPU went up to ~95% compared to 7 CPU tasks and ~85% load. Seems to me that the CPU doesn't respond quick enough to keep the GPU fully loaded. Next test will be to run 2 at a time to see what happens. |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Next test will be to run 2 at a time to see what happens. Did a test run with the help of a app_config.xml and set both the CPU and GPU usage to 0.5 to run x2 and still only reserve one CPU for support. GPU load was ~93-95% and run time went from just over 17 min running x1 to ~28 min running x2 so there was some gain. Note that I only ran the 2 task like this and then reverted to x1 again. I'd say lets get this running properly first and then look at how to run it optimally. |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
Cool! The plan on our side here is to test Intel OpenCL <-> AMD OpenCL driver compatibility, as there were reports about problems earlier. Hopefully this is fixed by now. We as a project would be reluctant to advertise Intel GPU OpenCL processing on E@H if it is likely to result in problems for AMD GPU users. I guess I'll try this weekend to put a HD 7790 in my i3-3220 system, install the drivers and see what happens. If that goes well, we will discuss the matter at E@H and see if the effort for the necessary server side changes is worth it, but my personal view on this is rather optimistic now. Currently I have no idea how much computing power can be gained by it (I guess the target group would be all Windows Systems running on third generation or later Intel Core processors with a BOINC client version of 7.0.29 or later). Holmis, I have a question for you: your system also has an NVIDIA GFX card (which, I assume, is used for the primary display?). This page https://boinc.berkeley.edu/wiki/GPU_computing states that you need to connect a monitor or VGA-dummy-plug for the Intel GPU to be detected (or activated...). Can you confirm this? Cheers HB |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Holmis, I have a question for you: your system also has an NVIDIA GFX card (which, I assume, is used for the primary display?). I have a TV connected to the Nvidia card and as you thought it's used as the primary display. To get the HD 4000 detected by Boinc I have to connect it to a monitor and choose to stretch/clone the image to that monitor plus the monitor has to be turned on when starting Boinc. I don't have a dummy-plug so can't test that. I think the reason is power saving and the HD 4000 gets turned off if nothing is using it. Once the HD 4000 is detected by Boinc I can turn the monitor off and all works as it should. |
Bikeman (Heinz-Bernd Eggenstein) Volunteer moderator Project administrator Project developer Send message Joined: 28 Aug 06 Posts: 1483 Credit: 1,864,017 RAC: 0 |
Thanks, I got my HD iGPU working together with a Radeon HD 7790. But as you suggested, it took some configuring in the BIOS settings and connecting a second monitor. I'm afraid this is a big caveat wrt. to the usefulness of Intel HD GPUs for Einstein@Home (and other BOINC projects). Maybe I'm wrong but I'm afraid most of the Ivy Bridge hosts crunching for Einstein@Home will have iGPUs not visible to BOINC either because of the iGPU being diabled by BIOS default settings in connection with a dedicated card, or because it is deactivated when not connected to a monitor :-(. Bernd is currently taking care of the server code on Einstein to make it "Intel GPU aware", I guess we will have a clearer picture about the number of Intel GPU OpenCL enabled hosts on E@H. Cheers HB |
Holmis Send message Joined: 4 Jan 05 Posts: 104 Credit: 2,104,736 RAC: 0 |
Happy to hear that you got it working! =) I agree that we're probably going to have to answer a lot of questions on the message boards about how to get it running if released on Einstein but at the same time I think it would be the first project to release a Intel GPU app that don't require the use of anonymous platform. There is some potential outside the group of more knowledgeable users that may benefit from the app as at least here in Sweden there is a market for cheaper laptops without a discrete graphics card, they only have Intel HD Graphics. Can't say anything about how big or small this group might be or if they are willing to use the GPU though... Now on to some more test results. I've been running my machine fully loaded with 8 CPU task, 2 BRP5 tasks on the GTX660Ti and 1 GPU task on the HD 4000 for the whole weekend, all seems to be working fine though maybe not at peek performance. I noticed that the BRP5 tasks where a bit slow and the GPU load on the 660Ti was about 10% lower than before I started running on the HD 4000. So I decided to reserve another core by setting my local prefs to only use 87.5% of the processors. This seems to have helped as the GPU load on the 660Ti have gone up to almost the same as before, maybe 2-3% lower than normal and on the same time the load on the HD 4000 is now on average 93% over a full task. I'm going to run like this for a few days to see how it goes. By the way anyone else that's tried this? Want to try but not sure how? Just ask and I'll try to help! =) |
Alex Send message Joined: 1 Mar 05 Posts: 88 Credit: 398,734 RAC: 0 |
I've tried it right now. With a dummy plug the device manager in windows showed a vvideo device without driver. The drivver could be installed online. Windows devvice manager shows now correct 3 graphic cards + 1 video hook driver. But the messages in BM are a bit strange now: 24.06.2013 22:50:09 | | VirtualBox version: 4.2.12 24.06.2013 22:50:09 | | CAL: ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (CAL version 1.4.1741, 2048MB, 2008MB available, 7580 GFLOPS peak) 24.06.2013 22:50:09 | | CAL: ATI GPU 1: ATI Radeon HD 5800/5900 series (Cypress/Hemlock) (CAL version 1.4.1741, 1024MB, 991MB available, 4422 GFLOPS peak) 24.06.2013 22:50:09 | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (driver version 1214.3 (VM), device version OpenCL 1.2 AMD-APP (1214.3), 2048MB, 2008MB available, 7580 GFLOPS peak) 24.06.2013 22:50:09 | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (driver version 1214.3 (VM), device version OpenCL 1.2 AMD-APP (1214.3), 2048MB, 2008MB available, 7580 GFLOPS peak) 24.06.2013 22:50:09 | | No NVIDIA library found The OpenCL lines show twice GPU 0, both times the same card. And the Intel GPU is not recognized. https://dl.dropboxusercontent.com/u/50246791/Screenshot%20graphic%20cards.PNG I guess this should be good enough to be detected by boinc. Using BM 7.1.17, Win7 64, i3 3220 |
Alex Send message Joined: 1 Mar 05 Posts: 88 Credit: 398,734 RAC: 0 |
Tried it anyway; got one wu that failed immediate. 24.06.2013 23:08:37 | Albert@Home | Scheduler request completed: got 1 new tasks 24.06.2013 23:08:37 | Albert@Home | [error] Missing coprocessor for task p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2473_1; aborting 24.06.2013 23:08:39 | Albert@Home | Started download of p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2473.bin4 24.06.2013 23:08:40 | Albert@Home | Finished download of p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2473.bin4 24.06.2013 23:09:37 | Albert@Home | Sending scheduler request: To fetch work. 24.06.2013 23:09:37 | Albert@Home | Reporting 1 completed tasks 24.06.2013 23:09:37 | Albert@Home | Requesting new tasks for CPU 24.06.2013 23:09:40 | Albert@Home | Scheduler request completed: got 1 new tasks 24.06.2013 23:09:40 | Albert@Home | [error] Missing coprocessor for task p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2474_1; aborting 24.06.2013 23:09:42 | Albert@Home | Started download of p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2474.bin4 24.06.2013 23:09:43 | Albert@Home | Finished download of p2030.20120226.G194.26-02.01.S.b6s0g0.00000_2474.bin4 24.06.2013 23:10:40 | Albert@Home | Sending scheduler request: To fetch work. Edit: And another thing is strange: also they are very different GPU's, the progress is exact the same of the FGRP wu's: https://dl.dropboxusercontent.com/u/50246791/Screenshot%20crunching%20albert.PNG |
Alex Send message Joined: 1 Mar 05 Posts: 88 Credit: 398,734 RAC: 0 |
I did a little more investigation. Intel drivers have been in wrong directories, most likely caused by automatic installation from windows. https://dl.dropboxusercontent.com/u/50246791/intel%20opencl%202.PNG https://dl.dropboxusercontent.com/u/50246791/intel%20opencl%201.PNG So I thought Intel might be a good source for getting the correct drivers. https://downloadcenter.intel.com/Default.aspx?ProductFamily=Graphics&ProductLine=Desktop+graphics+drivers&ProductProduct=3rd+Generation+Intel%C2%AE+Core%E2%84%A2+Processors+with+Intel%C2%AE+HD+Graphics+4000%2f2500&ZeroResult=True&lang=eng&ProdId=3498 This is a good source for the correct drivers. The messages now show 25.06.2013 13:30:22 | | Local time is UTC +2 hours 25.06.2013 13:30:22 | | VirtualBox version: 4.2.12 25.06.2013 13:30:22 | | CAL: ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (CAL version 1.4.1741, 2048MB, 2008MB available, 7488 GFLOPS peak) 25.06.2013 13:30:22 | | CAL: ATI GPU 1: ATI Radeon HD 5800/5900 series (Cypress/Hemlock) (CAL version 1.4.1741, 1024MB, 991MB available, 4368 GFLOPS peak) 25.06.2013 13:30:22 | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (driver version 1214.3 (VM), device version OpenCL 1.2 AMD-APP (1214.3), 2048MB, 2008MB available, 7488 GFLOPS peak) 25.06.2013 13:30:22 | | OpenCL: AMD/ATI GPU 0: AMD Radeon HD 7870/7950/7970 series (Tahiti) (driver version 1214.3 (VM), device version OpenCL 1.2 AMD-APP (1214.3), 2048MB, 2008MB available, 7488 GFLOPS peak) 25.06.2013 13:30:22 | | OpenCL: Intel GPU 0: Intel(R) HD Graphics 2500 (driver version 9.18.10.3165, device version OpenCL 1.2, 1624MB, 1624MB available, 17 GFLOPS peak) 25.06.2013 13:30:22 | | No NVIDIA library found This looks better now. Bios settings: primary graphic card pcie, gpu detection auto, dummy plug on vga-connector of internal gpu, monitor connected to first pcie graphic card. @ HBE: What you posted some time ago regarding app_info and cc_config , is this still required or was this for testing only? Edit: might be a little bit tricky because all GPU's have nr. 0 now ... |