New AstroPulse for GPU ( ATi & NV) released (r1316)


log in

Advanced search

Message boards : Number crunching : New AstroPulse for GPU ( ATi & NV) released (r1316)

1 · 2 · 3 · 4 . . . 11 · Next
Author Message
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256374 - Posted: 6 Jul 2012, 12:30:46 UTC

New AstroPulse binaries are available here: http://lunatics.kwsn.net/12-gpu-crunching/ap6-for-nv-ati-gpus-r1316-released.msg49053.html#msg49053

Please, report noticed issues in this thread.

Profile Fred E.
Volunteer tester
Send message
Joined: 22 Jul 99
Posts: 768
Credit: 24,136,341
RAC: 2,611
United States
Message 1256423 - Posted: 6 Jul 2012, 14:18:03 UTC

New AstroPulse binaries are available here: http://lunatics.kwsn.net/12-gpu-crunching/ap6-for-nv-ati-gpus-r1316-released.msg49053.html#msg49053

Please, report noticed issues in this thread.


Got it installed, have some questions:

1) On the Lunatics site, it says:

On hosts with BOINC supporting OpenCL app will use device supplied by BOINC. With older BOINC versions own enumeration ability will be used.

Not sure what this means - what do I need to do for 6.12.34?

2. Modified the suggested app_info.xml from ATI to NV as follows:

<app>
<name>astropulse_v6</name>
</app>
<file_info>
<name>AP6_win_x86_SSE2_OpenCL_NV_r1316.exe<app>
<executable/>
</file_info>
<app_version>
<app_name>astropulse_v6</app_name>
<version_num>604</version_num>
<avg_ncpus>0.04</avg_ncpus>
<max_ncpus>0.2</max_ncpus>
<plan_class>cuda</plan_class>
<cmdline></cmdline>
<coproc>
<type>cuda</type>
<count>.5</count>
</coproc>
<file_ref>
<file_name>AP6_win_x86_SSE2_OpenCL_NV_r1316.exe</file_name>
<main_program/>
</file_ref>
</app_version>

Will this work? I omitted the flops entry since I'm now running without them and would have to guess at a value anyway. Will add back if need arises.

Waiting for work, hope they mount some new files soon.
____________
Another Fred
Support SETI@home when you search the Web with GoodSearch or shop online with GoodShop.

rob smith
Volunteer tester
Send message
Joined: 7 Mar 03
Posts: 8140
Credit: 52,749,389
RAC: 75,174
United Kingdom
Message 1256440 - Posted: 6 Jul 2012, 15:35:32 UTC

I notice that the current offerings are all win-x86, are there plans for win-64 versions?
____________
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2570
Credit: 5,876,851
RAC: 2,712
Bulgaria
Message 1256450 - Posted: 6 Jul 2012, 15:59:56 UTC - in response to Message 1256374.

Second big change - there is ap_cmdline.txt file that can be used to add command line parameters to app.
Put params there as you would put them in corresponding tag in app_info.
App_info tag supported too so use a way that more convenient to you.

Which one will be used if they contradict?

(e.g.
* in app_info.xml
<cmdline>-unroll 3 -ffa_block 2048 -ffa_block_fetch 1024</cmdline>

* in ap_cmdline.txt
-unroll 6 -ffa_block 2048 -ffa_block_fetch 1024
)


Will this erase AP ATI 601 tasks on-board?:
<version_num>604</version_num>

I think I need second 'copy' section with:
<version_num>601</version_num>


P.S.
It will be good to supply APATI.aistub (two of them - for 32 and 64 bit) in the .7z


____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

Profile Mike
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 23381
Credit: 31,822,543
RAC: 24,117
Germany
Message 1256456 - Posted: 6 Jul 2012, 16:22:35 UTC
Last modified: 6 Jul 2012, 16:25:16 UTC

Here is an example for ATI cards.
Just use ap_cmdline.txt instead of comand line in appinfo.
So you can edit params without restarting Boinc.

Type has to be uppercase.

ATI or CUDA

<app>
<name>astropulse_v6</name>
</app>
<file_info>
<name>AP6_win_x86_SSE2_OpenCL_ATI_r1316.exe</name>
<executable/>
</file_info>
<file_info>
<name>ap_cmdline.txt</name>
</file_info>
<app_version>
<app_name>astropulse_v6</app_name>
<version_num>601</version_num>
<avg_ncpus>0.04</avg_ncpus>
<max_ncpus>0.2</max_ncpus>
<plan_class>ati13ati</plan_class>
<coproc>
<type>ATI</type>
<count>0.5</count>
</coproc>
<file_ref>
<file_name>AP6_win_x86_SSE2_OpenCL_ATI_r1316.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>ap_cmdline.txt</file_name>
</file_ref>
</app_version>

@ Fred E.

You dont need to do anything for 6.12 version.
Just check you appinfo with mine.

Mike
____________

Horacio
Send message
Joined: 14 Jan 00
Posts: 536
Credit: 69,362,203
RAC: 97,935
Argentina
Message 1256462 - Posted: 6 Jul 2012, 16:27:46 UTC

Question:
Is this new release still affected by the high cpu usage on newer drivers?

____________

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256468 - Posted: 6 Jul 2012, 16:34:56 UTC - in response to Message 1256440.

I notice that the current offerings are all win-x86, are there plans for win-64 versions?

It's GPU build. I'm not expect big improvement from x64 one. And this binary perfectly can work under x64 Windows OS

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256470 - Posted: 6 Jul 2012, 16:37:23 UTC - in response to Message 1256462.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256473 - Posted: 6 Jul 2012, 16:41:16 UTC - in response to Message 1256450.

Second big change - there is ap_cmdline.txt file that can be used to add command line parameters to app.
Put params there as you would put them in corresponding tag in app_info.
App_info tag supported too so use a way that more convenient to you.

Which one will be used if they contradict?

(e.g.
* in app_info.xml
<cmdline>-unroll 3 -ffa_block 2048 -ffa_block_fetch 1024</cmdline>

* in ap_cmdline.txt
-unroll 6 -ffa_block 2048 -ffa_block_fetch 1024
)


Will this erase AP ATI 601 tasks on-board?:
<version_num>604</version_num>

I think I need second 'copy' section with:
<version_num>601</version_num>


P.S.
It will be good to supply APATI.aistub (two of them - for 32 and 64 bit) in the .7z


1) If I remember correct text file processed after command line so params will be overwritten. It's easy to check cause app reports if default param replaced.
2) yes, better to make copy or just create this new section leaving old one untouched.
3) I think aistub will be supplied in next installer release. It's standalone app release.

Profile Mike
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 23381
Credit: 31,822,543
RAC: 24,117
Germany
Message 1256475 - Posted: 6 Jul 2012, 16:43:48 UTC

If cmdline.txt is not mentioned in appinfo it will be untouched on most hosts.

____________

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256476 - Posted: 6 Jul 2012, 16:44:52 UTC - in response to Message 1256423.

[quote]New AstroPulse binaries are available here: Not sure what this means - what do I need to do for 6.12.34?

Nothing. It was just some info about new app abilities and possible diffrence in experience under BOINC 7 and pre-7.

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256481 - Posted: 6 Jul 2012, 16:51:42 UTC
Last modified: 6 Jul 2012, 16:52:14 UTC

Regarding what to mention in app_info file: for stock users upgrading to this release it would be better to include libfftw3f-3.dll in app_info.
That's because stock version could mention it already. If you have this file in project dir then you need to mention it to save it from deletion. It's BOINC's paranoid "kill all I know" approach - quite useless cause nothing prevents to accumulate not known to BOINC trash in that folders.

In general, upgrading with big cache of tasks is not recommended cause chances to lost all accumulated work and work in progress quite big (the same "wipe all" approach in action). I would recommend to finish current GPU AP work, set to "no new task", do update then enable work fetch. In this case errors in app_info will cause minimal damage.

JohnDK
Volunteer tester
Avatar
Send message
Joined: 28 May 00
Posts: 829
Credit: 40,630,334
RAC: 69,217
Denmark
Message 1256627 - Posted: 6 Jul 2012, 20:58:51 UTC - in response to Message 1256470.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?

Profile Mike
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 23381
Credit: 31,822,543
RAC: 24,117
Germany
Message 1256637 - Posted: 6 Jul 2012, 21:36:02 UTC - in response to Message 1256627.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?


It is a bug.
It was the same with ATI drivers.

AMD fixed it.

____________

JohnDK
Volunteer tester
Avatar
Send message
Joined: 28 May 00
Posts: 829
Credit: 40,630,334
RAC: 69,217
Denmark
Message 1256642 - Posted: 6 Jul 2012, 21:50:10 UTC - in response to Message 1256637.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?


It is a bug.
It was the same with ATI drivers.

AMD fixed it.

Thanks, who will kick NV in the butt then? ;)

Profile Mike
Volunteer tester
Avatar
Send message
Joined: 17 Feb 01
Posts: 23381
Credit: 31,822,543
RAC: 24,117
Germany
Message 1256644 - Posted: 6 Jul 2012, 21:55:47 UTC - in response to Message 1256642.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?


It is a bug.
It was the same with ATI drivers.

AMD fixed it.

Thanks, who will kick NV in the butt then? ;)


I think it more depends on their strategy in particular.

____________

Profile Fred J. Verster
Volunteer tester
Avatar
Send message
Joined: 21 Apr 04
Posts: 3232
Credit: 31,585,541
RAC: 0
Netherlands
Message 1256704 - Posted: 7 Jul 2012, 0:23:31 UTC - in response to Message 1256644.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?


It is a bug.
It was the same with ATI drivers.

AMD fixed it.

Thanks, who will kick NV in the butt then? ;)


I think it more depends on their strategy in particular.



I'm gonna try this on 2 ATI 5870 GPUs, can I still use command-line parameters like unroll 15, ffa_block 10240 & fetch 5120, 1 per device)
? On AstroPulse work?
Or is it MB only? BĂȘta work?



____________


Knight Who Says Ni N!, OUT numbered.................

Profile Fred E.
Volunteer tester
Send message
Joined: 22 Jul 99
Posts: 768
Credit: 24,136,341
RAC: 2,611
United States
Message 1256916 - Posted: 7 Jul 2012, 8:31:44 UTC

Completed my first Task. It ran very fast - was on track for about an hour running with 2 MB 6.10 tasks at the same time, but exited early at the 35 minute mark with "Found 30 single pulses and 30 repeating pulses, exiting." Was only split tonight so it won't validate for a long time. CPU use was high - about 21% with driver 301.42 (I'm not currently reserving a CPU core for the GPU).

There are several warnings in the StdErr you might want to review. One seems to be about the BOINC version I'm using, and I'm not concerned after your earlier comments. The other has to do with inability to open a file:

"INFO: can't open binary kernel file: C:\BoincData/projects/setiathome.berkeley.edu\AstroPulse_Kernels_r1316.cl_GeForceGTX670.bin_V6, continue with recompile..." and another similar message labeled "Warning". Maybe because it's the first time? Or my non=standard data directory location?

Guess my app_info.xml section is okay. Anyone looking at that earlier in this thread needs to make the fix that Mike pointed out - CUDA needs to be upper case on the Type line.


____________
Another Fred
Support SETI@home when you search the Web with GoodSearch or shop online with GoodShop.

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256925 - Posted: 7 Jul 2012, 9:00:46 UTC - in response to Message 1256627.

Question:
Is this new release still affected by the high cpu usage on newer drivers?

Yes, cause nothing was done in drivers still. They still use busy-wait loop.

Since there's been many driver updates since the last working version, I've been wondering if that driver bug is a bug or just the way the drivers work now?

It can be called in different ways, but the facts are:
1) there is 2 fundamental ways for CPU to communicate with hardware: to read from port or mapped memory and to process hardware interrupt. First one result in busy wait loop.
2) Earlier drivers work on the same hardware with the same software and low CPU usage method. Current ones use lot of CPU. So, apparently, method for synching was changed.
3) Both methods are possible technically, even with new drivers, CUDA runtime demonstrate this. But NV did not expose ability to switch over driver modes into OpenCL runtime.

Call it but or call it feature - no matter.

Profile Raistmer
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 16 Jun 01
Posts: 3368
Credit: 46,103,851
RAC: 22,144
Russia
Message 1256930 - Posted: 7 Jul 2012, 9:08:10 UTC - in response to Message 1256916.


"INFO: can't open binary kernel file: C:\BoincData/projects/setiathome.berkeley.edu\AstroPulse_Kernels_r1316.cl_GeForceGTX670.bin_V6, continue with recompile..." and another similar message labeled "Warning". Maybe because it's the first time?


Yes, such output should appear at first running of new versionBut if one sees it always it would mean that kernels binary cache was not saved.
First time app compiles OpenCL kernels from CL file (and oclFFT kernels) into binary representation and saves those binary representations into project directory.
In case of NV those "binary representations" are mostly text one actually - PTX assembly saved so one who interesting can look and see what assembly instructions were generated from original OpenCL code.

1 · 2 · 3 · 4 . . . 11 · Next

Message boards : Number crunching : New AstroPulse for GPU ( ATi & NV) released (r1316)

Copyright © 2014 University of California