Message boards :
Number crunching :
App for Intel GPU on Linux. When? Or never?
Message board moderation
Author | Message |
---|---|
siu77 Send message Joined: 18 Apr 14 Posts: 7 Credit: 855,127 RAC: 4 |
Any thoughts? |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Any thoughts? When OpenCL runtime will work there. |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14649 Credit: 200,643,578 RAC: 874 |
|
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Juha has got it running at Einstein with beignet. I think on SETI forum asking for SETI app, not for Einstein's or ABCDEF one. Look corresponding thread on Lunatics where Urs precisely described current state of affairs. |
Juha Send message Joined: 7 Mar 04 Posts: 388 Credit: 1,857,738 RAC: 0 |
Multibeam does work but with a heavy dose of ymmv. I might (or might not) give Haswell another try this weekend with newer kernel and Beignet than previously. And it's not very nice to refer to a thread hidden in some dev only secret lair. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Richard has access. P.S. And if you have working build bother to share it with the rest, that would be nice. |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14649 Credit: 200,643,578 RAC: 874 |
But as you know, I'm not a developer myself, just a messenger between boards trying to cross-fertilise between green shoots in the different flower-beds. You point was "When OpenCL runtime will work 'there'" - I assumed you meant "When OpenCL runtime will work on intel GPUs". It seemed to me that beignet had been demonstrated to supply that runtime - though, as the post I linked made clear, there were ongoing problems in progressing from a runtime implementation to a complete and validating app. |
Juha Send message Joined: 7 Mar 04 Posts: 388 Credit: 1,857,738 RAC: 0 |
People keep posting their experiences with new Windows drivers here all the time. I see no reason why similar information for Linux couldn't be made public. I had only limited amount of time to do the test so I used the Windows version of Multibeam. No reason to make the testing harder by, say, compiling the app myself and using wrong #defines. I'm still going to leave real testing to those who own suitable hardware. |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14649 Credit: 200,643,578 RAC: 874 |
Look corresponding thread on Lunatics where Urs precisely described current state of affairs. If I've found the right thread in the hidden lair, the 'current' state of play is dated 28 April 2015, and says "Looks like Beignet 1.01 was still too buggy and 1.02 does still not deliver, but clinfo now works flawless in standalone : ..." Claggy has been doing the messaging, but hasn't yet passed on Juha's report of success on the Haswell. The only later reference is to Juda's post here when (as I interpret it) he got the runtime to work, but - for want of a native Linux app to test - used the Windows app under Wine as a 'proof of concept'. So, Juha has the runtime, and Urs has the app, but they haven't shared in either direction yet. Maybe this thread would be a place to start? |
Juha Send message Joined: 7 Mar 04 Posts: 388 Credit: 1,857,738 RAC: 0 |
I might (or might not) give Haswell another try this weekend with newer kernel and Beignet than previously. ...or the next weekend. Ok, I ran the test on an i7-4790, running kernel 4.1.0 with i915.enable_ppgtt=2 on the kernel command line, and Beignet 1.1.0 that was coincidentally released last Friday. This time I used a native version of MB7 app built from revision 2987 using configure line from sah_v7_opt/AKv8/ConfigureLine_AKv8c_OPENCL_SSE2_MBv7.txt. (Apart from the usual adjustment of paths I removed the LDSTATIC part as the app doesn't seem to be using libssl and libcrypto for anything and I compiled the app against headers and libOpenCL.so from opencl-headers and ocl-icd-opencl-dev packages.) I couldn't do any longer test runs so had to settle to running the benchmark workunits. And as I expected this combination of kernel and Beignet produced correct results. I was also going to re-test the Windows r2489 version so that I could compare the times against an earlier test run I did on Ivy Bridge CPU. Unfortunately I messed up setting up the 32-bit Beignet. A Beignet install includes both header files and a precompiled headers file and the timestamps of the original header files are included in this pch-file. Previously I moved the install around in an archive file but this time I just copied it. And unless you take special precautions you lose original timestamps when copying. Beignet doesn't like timestamps changing and simply refuses to compile anything even though the original headers are available. So that's something to watch out for. Finally, Beignet needs recent enough libdrm-intel for "userptr support", "EU total query support" and "subslice total query support". Ubuntu 15.04 has one, 14.04.0 doesn't. I used 15.04. I have no idea what those features are or does Multibeam app need them when run on Haswell GPU, but, again, something to watch out for. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.