SETI@home v8 beta to begin on Tuesday

Message boards : News : SETI@home v8 beta to begin on Tuesday
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 75 · 76 · 77 · 78 · 79 · 80 · 81 . . . 99 · Next

AuthorMessage
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59730 - Posted: 28 Sep 2016, 19:11:45 UTC - in response to Message 59722.  

That's why additional checks to conform with original logic required.
And any changes in that logic require discussion with Eric.
You have our mails - please send your proposals there.
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59730 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59731 - Posted: 28 Sep 2016, 19:19:28 UTC
Last modified: 28 Sep 2016, 19:20:23 UTC

I would like to report inconsistency in original pulse find code that could lead to additional rounding noise between builds:

In snr and threshold calculations (that define signal's score later) used sqrtf() in some places and sqrt((float)) in others.
In my understansing that leads to different rounding of square root results. So can influence borderline cases. Better to chose one function.

For MSVC sqrtf() defined as:

inline float sqrtf(_In_ float _X)
        {return ((float)sqrt((double)_X)); }

So it's essentially double precision one (instead of GPU case where sqrtf takes float and returns float).
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59731 · Report as offensive
petri33
Volunteer tester

Send message
Joined: 25 Feb 12
Posts: 8
Credit: 1,623,142
RAC: 0
Finland
Message 59736 - Posted: 29 Sep 2016, 21:12:25 UTC - in response to Message 59731.  

I would like to report inconsistency in original pulse find code that could lead to additional rounding noise between builds:

In snr and threshold calculations (that define signal's score later) used sqrtf() in some places and sqrt((float)) in others.
In my understansing that leads to different rounding of square root results. So can influence borderline cases. Better to chose one function.

For MSVC sqrtf() defined as:

inline float sqrtf(_In_ float _X)
        {return ((float)sqrt((double)_X)); }

So it's essentially double precision one (instead of GPU case where sqrtf takes float and returns float).


... and in some places of the code the sqrt/sqrtf can be omitted because the next value is "sqrt(previous*2.0f)" and that can be replaced by a multiplication by a constant sqrt(2) that translates to "previous*1.4142...(float or double)". And yes, I know, that imposes another kind of rounding error but it is significantly faster if it is in an often used code path.

But thanks, I'll look at the code for sqrt/sqrtf.
ID: 59736 · Report as offensive
petri33
Volunteer tester

Send message
Joined: 25 Feb 12
Posts: 8
Credit: 1,623,142
RAC: 0
Finland
Message 59737 - Posted: 29 Sep 2016, 21:17:10 UTC - in response to Message 59730.  

That's why additional checks to conform with original logic required.
And any changes in that logic require discussion with Eric.
You have our mails - please send your proposals there.


Finding a pulse is a rare event. I can use atomic_max() to GPU main memory from different threads to ensure that a real maximum is found. Another option is to gather all found pulses to a buffer and run a second kernel or a set of last threads to select the 'reportable' and 'best' pulses.
ID: 59737 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59742 - Posted: 30 Sep 2016, 16:40:59 UTC - in response to Message 59737.  
Last modified: 30 Sep 2016, 16:42:54 UTC

That's why additional checks to conform with original logic required.
And any changes in that logic require discussion with Eric.
You have our mails - please send your proposals there.


Finding a pulse is a rare event. I can use atomic_max() to GPU main memory from different threads to ensure that a real maximum is found. Another option is to gather all found pulses to a buffer and run a second kernel or a set of last threads to select the 'reportable' and 'best' pulses.

yes.
In this particular case atomic reduce would be better option IMHO cause there is no reasonable upper limit for number of signals for reduce (instead of reportable ones where we have hardwired upper limit of 30 pulses for now).

EDIT but check CUDA specification carefully if atomic global memory operations still atomic between different blocks. If I recall correctly it's not the case for OpenCL workgroups (that is, different workgroup execution order unrestricted )
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59742 · Report as offensive
boboviz
Volunteer tester

Send message
Joined: 13 Dec 14
Posts: 14
Credit: 155,885
RAC: 0
Italy
Message 59746 - Posted: 1 Oct 2016, 8:56:14 UTC - in response to Message 59709.  

I think we finished with massive testing of 8.19.
Hope Eric will have some time for release after returning from conference.
Next version not ready for beta for now, I'll make announcement when it will be ready.


Great news.
8.19 will go on Seti?
ID: 59746 · Report as offensive
Richard Haselgrove
Volunteer tester

Send message
Joined: 3 Jan 07
Posts: 1451
Credit: 3,272,268
RAC: 0
United Kingdom
Message 59747 - Posted: 1 Oct 2016, 11:29:50 UTC - in response to Message 59746.  

I think we finished with massive testing of 8.19.
Hope Eric will have some time for release after returning from conference.
Next version not ready for beta for now, I'll make announcement when it will be ready.

Great news.
8.19 will go on Seti?

It's only news once Eric okays it ;-)

Last I heard he was out of state (giving a talk at a scientific conference in Mexico)
ID: 59747 · Report as offensive
Zalster
Volunteer tester

Send message
Joined: 30 Dec 13
Posts: 258
Credit: 12,340,341
RAC: 0
United States
Message 59749 - Posted: 1 Oct 2016, 18:20:42 UTC - in response to Message 59747.  

I think we finished with massive testing of 8.19.
Hope Eric will have some time for release after returning from conference.
Next version not ready for beta for now, I'll make announcement when it will be ready.

Great news.
8.19 will go on Seti?

It's only news once Eric okays it ;-)

Last I heard he was out of state (giving a talk at a scientific conference in Mexico)


Hopefully near a beach with lots of frozen drinks ;)
ID: 59749 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 2 Jul 13
Posts: 505
Credit: 5,019,318
RAC: 0
United States
Message 59757 - Posted: 5 Oct 2016, 1:35:20 UTC

For any Mac users testing the Mac nVidia Apps, nVidia released a new CUDA Driver yesterday. Everyone with a Fermi and above GPU running Darwin 15.x and above, Especially those running Darwin 16.x, should update to the New Driver,
http://www.nvidia.com/object/macosx-cuda-8.0.46-driver.html
New Release 8.0.46
CUDA driver update to support macOS 10.12 and NVIDIA display driver 367.15.10.05f01
Recommended CUDA version(s): CUDA 8.0
Supported MAC OS X:
10.12.x
10.11.x

Apparently everyone with Darwin 14 and below are excluded from this update. For those with Darwin 13 & 14 running Fermi and above, you should be using this driver, http://www.nvidia.com/object/macosx-cuda-7.5.30-driver.html
ID: 59757 · Report as offensive
JLDun
Volunteer tester
Avatar

Send message
Joined: 23 May 07
Posts: 108
Credit: 127,898
RAC: 0
United States
Message 59758 - Posted: 5 Oct 2016, 3:59:06 UTC
Last modified: 5 Oct 2016, 4:01:04 UTC

I'd like to point out a WU on Main, just in case of something 'interesting':
2283484899
Currently 'Inconclusive' after being sent to a x86-android-linux-gnu (Android/Intel Atom x64), opencl_intel_gpu_sah (Mac x64)and opencl_nvidia_SoG (Windows).
ID: 59758 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59759 - Posted: 5 Oct 2016, 5:57:42 UTC - in response to Message 59758.  

I'd like to point out a WU on Main, just in case of something 'interesting':
2283484899
Currently 'Inconclusive' after being sent to a x86-android-linux-gnu (Android/Intel Atom x64), opencl_intel_gpu_sah (Mac x64)and opencl_nvidia_SoG (Windows).

overflow
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59759 · Report as offensive
Zalster
Volunteer tester

Send message
Joined: 30 Dec 13
Posts: 258
Credit: 12,340,341
RAC: 0
United States
Message 59773 - Posted: 7 Oct 2016, 2:11:09 UTC

I'm guessing there is a problem with the server showing recent work units?

The last one show is from yesterday evening. Only ask as I'm trying to see how the new OS, OpenCl and Cuda are reacting to the work units.

I'll check in the morning and hopefully it will update to a more recent report so I can tell if they are validating or not.
ID: 59773 · Report as offensive
Profile Jimbocous
Volunteer tester
Avatar

Send message
Joined: 9 Jan 16
Posts: 51
Credit: 1,038,205
RAC: 0
United States
Message 59774 - Posted: 7 Oct 2016, 3:47:51 UTC - in response to Message 59773.  

I'm guessing there is a problem with the server showing recent work units?

The last one show is from yesterday evening. Only ask as I'm trying to see how the new OS, OpenCl and Cuda are reacting to the work units.

I'll check in the morning and hopefully it will update to a more recent report so I can tell if they are validating or not.

Replica DB server has been offline most of the day, according to the SSP. Never really recovered from the maintenance window Tuesday.
If I can help out by testing something, please let me know.
Available hardware and software is listed in my profile here.
ID: 59774 · Report as offensive
Zalster
Volunteer tester

Send message
Joined: 30 Dec 13
Posts: 258
Credit: 12,340,341
RAC: 0
United States
Message 59775 - Posted: 7 Oct 2016, 4:21:00 UTC - in response to Message 59774.  

Yeah I was wondering if it was something like that..Well guess nothing to do but keep crunching until they are able to repair it.

Thanks

Zalster
ID: 59775 · Report as offensive
Profile Jimbocous
Volunteer tester
Avatar

Send message
Joined: 9 Jan 16
Posts: 51
Credit: 1,038,205
RAC: 0
United States
Message 59776 - Posted: 7 Oct 2016, 4:36:25 UTC - in response to Message 59775.  

Yeah I was wondering if it was something like that..Well guess nothing to do but keep crunching until they are able to repair it.

Yeah, as long as we can u/l and d/l work, it's all good :)
If I can help out by testing something, please let me know.
Available hardware and software is listed in my profile here.
ID: 59776 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59777 - Posted: 7 Oct 2016, 14:17:42 UTC

Here https://cloud.mail.ru/public/2aUP/dborYAw9G is the iGPU build with maximal possible precision options for kernel code.
Please try if it will improve iGPU precision.
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59777 · Report as offensive
SusieQ
Volunteer tester

Send message
Joined: 12 Nov 10
Posts: 1149
Credit: 32,460,657
RAC: 1
United Kingdom
Message 59778 - Posted: 7 Oct 2016, 19:14:53 UTC - in response to Message 59777.  
Last modified: 7 Oct 2016, 19:59:18 UTC

Hi Raistmer

I have just set this host going using your new iGPU app http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=78787

I'll leave it running over the weekend - hopefully that should give you some idea as to whether your changes have worked or not. Edit: Well it might if they manage to fix the results pages :)

Fingers crossed :-)
ID: 59778 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59781 - Posted: 8 Oct 2016, 7:28:33 UTC - in response to Message 59778.  

Hi Raistmer

I have just set this host going using your new iGPU app http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=78787

I'll leave it running over the weekend - hopefully that should give you some idea as to whether your changes have worked or not. Edit: Well it might if they manage to fix the results pages :)

Fingers crossed :-)

Thanks, I'll look for your host results.
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59781 · Report as offensive
SusieQ
Volunteer tester

Send message
Joined: 12 Nov 10
Posts: 1149
Credit: 32,460,657
RAC: 1
United Kingdom
Message 59783 - Posted: 10 Oct 2016, 12:57:35 UTC - in response to Message 59781.  

Hi Raistmer

I have just set this host going using your new iGPU app http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=78787

I'll leave it running over the weekend - hopefully that should give you some idea as to whether your changes have worked or not. Edit: Well it might if they manage to fix the results pages :)

Fingers crossed :-)

Thanks, I'll look for your host results.

I was going to swap this host back to stock today but, as the results pages are still not updating, I'll leave it running your test app until after this week's outage. Hopefully the results pages will be working again after that and you can confirm whether you have enough information before I swap back :-)
ID: 59783 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 59789 - Posted: 12 Oct 2016, 5:59:50 UTC - in response to Message 59783.  
Last modified: 12 Oct 2016, 6:00:41 UTC

Seems new build is too imprecise on Skylake (INTEL Intel(R) HD Graphics 530 (4859MB) OpenCL: 2.0) too.
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 59789 · Report as offensive
Previous · 1 . . . 75 · 76 · 77 · 78 · 79 · 80 · 81 . . . 99 · Next

Message boards : News : SETI@home v8 beta to begin on Tuesday


 
©2022 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.