Astropulse 6.02 Windows for OpenCL on ATI released.

Message boards : News : Astropulse 6.02 Windows for OpenCL on ATI released.
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 . . . 45 · Next

AuthorMessage
arkayn
Volunteer tester
Avatar

Send message
Joined: 16 Jan 07
Posts: 155
Credit: 194,400
RAC: 0
United States
Message 42264 - Posted: 18 Apr 2012, 21:36:42 UTC - in response to Message 42263.  

I thought that OpenCL support was older than that... Or are other projects shipping OpenCL apps disguised as normal ATI or CUDA. Is there a driver version for each for which we could assume OpenCL would function?


They are being disguised as ATI or CUDA, just look at Milkyway for proof.

BOINC 7.x is the first client to actually send OpenCL info to the server.
ID: 42264 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42265 - Posted: 18 Apr 2012, 21:41:30 UTC - in response to Message 42264.  

I thought that OpenCL support was older than that... Or are other projects shipping OpenCL apps disguised as normal ATI or CUDA. Is there a driver version for each for which we could assume OpenCL would function?


They are being disguised as ATI or CUDA, just look at Milkyway for proof.

BOINC 7.x is the first client to actually send OpenCL info to the server.


And for anonymous platform we using that trick long ago already.

Unfortunately, though it will be OK for NV, there could be problems with ATi where not all GPU OpenCl compatible. Some are Brook+ - only....
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42265 · Report as offensive
zombie67 [MM]
Volunteer tester
Avatar

Send message
Joined: 18 May 06
Posts: 280
Credit: 26,339,733
RAC: 917
United States
Message 42266 - Posted: 18 Apr 2012, 21:43:58 UTC - in response to Message 42233.  

For the folks that can't get work, did you try detaching/re-attaching? That worked for me. I also quit BOINC in between, just to make sure the project info was complete removed before re-attaching.

Did you do any analysis, say of your work fetch debug logs before/after, to indicate why this course of action should make a difference?


It wasn't me that came up with the idea. It was suggested earlier in this thread. And it worked for me. That is all I know.
Dublin, California
Team: SETI.USA

ID: 42266 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42267 - Posted: 18 Apr 2012, 21:44:20 UTC
Last modified: 18 Apr 2012, 21:44:32 UTC

Estimates are too far from reality (as usual for AP :( )
My host completes task in ~hour while estimated time for fresh download is 103h.
BTW, first task had estimate ~90h.
Ok, they big, but why the hell they got even bigger ????
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42267 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42268 - Posted: 18 Apr 2012, 21:47:40 UTC

And another observation: while it runs (0.978 CPUs + 1 ATI GPU), host running 4 MB tasks. SO, all cores still busy, no reserved core on my host. And it's good.
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42268 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42269 - Posted: 18 Apr 2012, 21:47:52 UTC - in response to Message 42264.  

Looks like they're using ati14 plan classes which seem like they're API version 1.4.0 or higher. Does OpenCL always come with that API?
ID: 42269 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42270 - Posted: 18 Apr 2012, 21:49:26 UTC - in response to Message 42267.  

It was probably my changes to the scheduler that messed up the time computation. If it gets too far out of whack, I'll zero them out.


ID: 42270 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42271 - Posted: 18 Apr 2012, 21:51:01 UTC - in response to Message 42265.  

We could probably put a Brook+ build online if it becomes a problem. It's not like you didn't put a lot of work into that code, too.
ID: 42271 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42272 - Posted: 18 Apr 2012, 21:51:36 UTC - in response to Message 42270.  

It was probably my changes to the scheduler that messed up the time computation. If it gets too far out of whack, I'll zero them out.



2 orders of magnitude surely less than distance to habitated star, but... ;)
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42272 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42273 - Posted: 18 Apr 2012, 21:53:02 UTC - in response to Message 42271.  
Last modified: 18 Apr 2012, 21:55:55 UTC

We could probably put a Brook+ build online if it becomes a problem. It's not like you didn't put a lot of work into that code, too.


The problem - how we will distinguish between hosts that Brook+ only and OpenCL w/o help of BOINC ?
For anonymous platform there is no problem cause operator behind keyboard usually sensible enough for that....
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42273 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42274 - Posted: 18 Apr 2012, 21:54:22 UTC - in response to Message 42269.  

Looks like they're using ati14 plan classes which seem like they're API version 1.4.0 or higher. Does OpenCL always come with that API?


Early ATi drivers needed SDK to be installed to enable OpenCL support...
Probably we have to limit driver version to be sure if OpenCL supported.
Then we need way to pick capable GPU .
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42274 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42275 - Posted: 18 Apr 2012, 21:58:38 UTC - in response to Message 42273.  
Last modified: 18 Apr 2012, 22:01:37 UTC

The problem - how we will distinguish between hosts that Brook+ only and OpenCL w/o help of BOINC ?


We would put the Brook+ version on an ATI plan class with an API version of whatever the minimum is... (1.2? 1.3?), so anyone with a compatible ATI board will get it, down to whatever BOINC version supported ATI.

The OpenCL plan class will go only to BOINC client 7.0.25 or higher. Of course, at first those machines will also get the Brook+ version until the scheduler decides which is best for their machine. That should be the OpenCL version. (But may not be on some boards.)
ID: 42275 · Report as offensive
arkayn
Volunteer tester
Avatar

Send message
Joined: 16 Jan 07
Posts: 155
Credit: 194,400
RAC: 0
United States
Message 42276 - Posted: 18 Apr 2012, 22:02:38 UTC
Last modified: 18 Apr 2012, 22:04:07 UTC

I have finished 2 ATI OpenCL WU, 1 was a too much blanking and the other took 6400 seconds on my low end HD7750.

The Nvidia WU is at 41% in 34 minutes.
ID: 42276 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42277 - Posted: 18 Apr 2012, 22:03:37 UTC - in response to Message 42275.  

The problem - how we will distinguish between hosts that Brook+ only and OpenCL w/o help of BOINC ?


We would put the Brook+ version on an ATI plan class with an API version of whatever the minimum is... (1.2? 1.3?), so anyone with a compatible ATI board will get it, down to whatever BOINC version supported ATI.

The OpenCL plan class will go only to BOINC client 7.0.25 or higher. Of course, at first those machines will also get the Brook+ version until the scheduler decides which is best for their machine. That should be the OpenCL version. (But may not be on some boards.)


This surely will improve overall performance... but it will not solve initial problem: OpenCL support only for BOINC 7+. Many peoples refuse to update BOINC.
I stayed with modded 6.10.58 as long as was possible to...

News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42277 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42278 - Posted: 18 Apr 2012, 22:05:20 UTC - in response to Message 42275.  

(But may not be on some boards.)


I'm quite sure that OpenCL will be faster everywhere where it can run.
Brook+ lack of FFT so only FFA moved to GPU. AP's main loop completely on CPU there...
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42278 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42279 - Posted: 18 Apr 2012, 22:15:29 UTC
Last modified: 18 Apr 2012, 22:18:51 UTC

Here's what is currently getting checked. It looks like we are checking a driver version. And a "version_num". I don't know if that's an API version number or a hardware version number. I'd probably have to go into the core client code to find out.

This is also where the crappy FLOPS estimates come from.

Preexisting plan classes come for version numbers 1.0, 1.3.0, 1.3.186, and 1.4.0. The 1.3.186 appears to be when some DLL names changed from ATI to AMD or vice versa.

static bool ati_check(COPROC_ATI& c, HOST_USAGE& hu,
    int min_driver_version,
    bool need_amd_libs,
    double min_ram,
    double ndevs,       // # of GPUs used; can be fractional
    double cpu_frac,    // fraction of FLOPS performed by CPU
    double flops_scale
) {
    if (c.version_num) {
        ati_requirements.update(min_driver_version, min_ram);
    }

    if (need_amd_libs) {
        if (!c.amdrt_detected) {
            return false;
        }
    } else {
        if (!c.atirt_detected) {
            return false;
        }
    }
    if (c.version_num < min_driver_version) {
        return false;
    }
    if (c.available_ram < min_ram) {
        return false;
    }

    hu.gpu_ram = min_ram;
    hu.natis = ndevs;

    coproc_perf(
        capped_host_fpops(),
        flops_scale * hu.natis*c.peak_flops,
        cpu_frac,
        hu.projected_flops,
        hu.avg_ncpus
    );
    hu.peak_flops = hu.natis*c.peak_flops + hu.avg_ncpus*capped_host_fpops();
    hu.max_ncpus = hu.avg_ncpus;
    return true;
}

ID: 42279 · Report as offensive
Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 15 Mar 05
Posts: 1547
Credit: 26,967,897
RAC: 1,380
United States
Message 42280 - Posted: 18 Apr 2012, 22:26:02 UTC - in response to Message 42279.  

David's favorite solutions for problems like this is to let it fail. I don't like that solution for obvious reasons one of which is that I get a lot of email. Where an app version fails repeatedly on a host, in theory the scheduler decides to stop sending that app version. So we could do the Brook version as plan class 'ati13ati' and the OpenCL version as 'ati14'. Cards that fail on OpenCL would stop getting the ati14 version.

Just throwing it out as an option.
ID: 42280 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42281 - Posted: 18 Apr 2012, 22:29:49 UTC - in response to Message 42280.  

David's favorite solutions for problems like this is to let it fail. I don't like that solution for obvious reasons one of which is that I get a lot of email. Where an app version fails repeatedly on a host, in theory the scheduler decides to stop sending that app version. So we could do the Brook version as plan class 'ati13ati' and the OpenCL version as 'ati14'. Cards that fail on OpenCL would stop getting the ati14 version.

Just throwing it out as an option.


Ok, but how long they will fail? AP tasks quite big - it could create big load on already very limited bandwidth...
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42281 · Report as offensive
Profile Raistmer
Volunteer tester
Avatar

Send message
Joined: 18 Aug 05
Posts: 2423
Credit: 15,878,738
RAC: 0
Russia
Message 42282 - Posted: 18 Apr 2012, 22:42:23 UTC
Last modified: 18 Apr 2012, 22:42:55 UTC

Looks like 11.1 - the first Catalyst with integrated OpenCL support (thanks goes to arkayn for discovering this).

Catalyst Software Suite 91.6 MB 11.1 1/26/2011 Download
Description:
Package Includes:
Display Driver
OpenCL Driver
ATI Integrated Driver
Catalyst Control Center
(English Language Only)
For the 64 bit version of Windows Vista and Windows 7

http://support.amd.com/us/gpudownload/windows/previous/11/Pages/radeon.aspx?os=Windows%20Vista%20-%2064-Bit%20Edition&rev=11.1
News about SETI opt app releases: https://twitter.com/Raistmer
ID: 42282 · Report as offensive
Richard Haselgrove
Volunteer tester

Send message
Joined: 3 Jan 07
Posts: 1451
Credit: 3,266,428
RAC: 0
United Kingdom
Message 42283 - Posted: 18 Apr 2012, 22:48:24 UTC - in response to Message 42282.  
Last modified: 18 Apr 2012, 22:54:58 UTC

Looks like 11.1 - the first Catalyst with integrated OpenCL support (thanks goes to arkayn for discovering this).

That's been known for a long time - it's even in the Lunatics installer ReadMe.

Catalyst 10.10 Accelerated Parallel Processing (APP) Technology Edition

11.1 was the first driver where the APP version was promoted to default download.

Edit - SETI@home user HAL9000 maintains a useful cheat sheet with ATI driver version equivalences: http://www.hal6000.com/seti/boinc_ati_gpu_cheat_sheet.htm
ID: 42283 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 . . . 45 · Next

Message boards : News : Astropulse 6.02 Windows for OpenCL on ATI released.


 
©2019 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.