SETI@home v7 6.98 for ATI OpenCL released.

log in

Advanced search

Message boards : News : SETI@home v7 6.98 for ATI OpenCL released.

Previous · 1 · 2 · 3 · 4 · 5 . . . 7 · Next
Author Message
Profile Mike
Volunteer tester
Avatar
Send message
Joined: 16 Jun 05
Posts: 2096
Credit: 889,977
RAC: 3,745
Message 44096 - Posted: 17 Oct 2012, 15:59:16 UTC - in response to Message 44092.
Last modified: 17 Oct 2012, 16:01:37 UTC

We're just guessing at the requirements for this version, so point me at any hosts that are repeatedly failing with this app version.


The app is fully OpenCL 1.0 compatible.

I tested on HD 4770 (RV 770) HD 4850 with driver 12.1.

Work group size 256 required.

I suggest driver version 12.1. and newer.

Running Windows XP driver 12.1 is the last driver with OpenCL support.

Mike


And I'm running my HD4850 on Vista with driver version 12.4, which also is an OpenCL 1.0 driver on the 4850. Apparently HD4850 doesn't support other than OpenCL 1.0, so no matter which newer driver you install, you will not get anything more than OpenCL 1.0 on that card.

But you knew that already Mike, I'm just a bit chatty :-)


HD 4xxx only has OpenCL beta Level support and now is officially depreciated by AMD.
http://developer.amd.com/TOOLS/HC/AMDAPPSDK/pages/DriverCompatibility.aspx

Which means it might work but must not.
I found out RV 770 works O.K.
____________

Profile cAnDYmanS@H-Beta
Volunteer tester
Avatar
Send message
Joined: 24 May 12
Posts: 38
Credit: 436,379
RAC: 0
Message 44097 - Posted: 17 Oct 2012, 15:59:53 UTC
Last modified: 17 Oct 2012, 16:01:11 UTC

Can't believe it's finally here! I almost gave up hope on running S@H on my ATI.
Already completed successfully 4 units or so. Took about 20 minutes per task. Crunching on...but only have 2 tasks on my machine at a time. Speaking of which: is there some sort of limitation or have I been knocked down to unreliable?

Win 7 Ultimate x64
Asus ATI 6870 1GB stock frequencies
Boinc 7.0.28 (x64)

Cheers!
____________

Per aspera, ad astra!

Profile Mike
Volunteer tester
Avatar
Send message
Joined: 16 Jun 05
Posts: 2096
Credit: 889,977
RAC: 3,745
Message 44098 - Posted: 17 Oct 2012, 16:07:38 UTC - in response to Message 44097.

Can't believe it's finally here! I almost gave up hope on running S@H on my ATI.
Already completed successfully 4 units or so. Took about 20 minutes per task. Crunching on...but only have 2 tasks on my machine at a time. Speaking of which: is there some sort of limitation or have I been knocked down to unreliable?

Win 7 Ultimate x64
Asus ATI 6870 1GB stock frequencies
Boinc 7.0.28 (x64)

Cheers!


Maybe you need to increase your preferences here.
Just go to your account and change keep enough work for x days.

____________

Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar
Send message
Joined: 15 Mar 05
Posts: 1333
Credit: 15,718,393
RAC: 17,177
Message 44099 - Posted: 17 Oct 2012, 16:10:37 UTC - in response to Message 44076.


There has to be some difference though between our results.

wuid=4153634


The first listed result (from host 59866) is missing a best_spike entry in its result file. I'm not sure why that would happen.

The name of that workunit alone indicates a problem in the splitter that I will have to track down and fix.
____________

CH3-CH2-OH
Volunteer tester
Send message
Joined: 10 Mar 12
Posts: 810
Credit: 3,092,953
RAC: 786
Message 44100 - Posted: 17 Oct 2012, 16:14:01 UTC - in response to Message 44096.

We're just guessing at the requirements for this version, so point me at any hosts that are repeatedly failing with this app version.


The app is fully OpenCL 1.0 compatible.

I tested on HD 4770 (RV 770) HD 4850 with driver 12.1.

Work group size 256 required.

I suggest driver version 12.1. and newer.

Running Windows XP driver 12.1 is the last driver with OpenCL support.

Mike


And I'm running my HD4850 on Vista with driver version 12.4, which also is an OpenCL 1.0 driver on the 4850. Apparently HD4850 doesn't support other than OpenCL 1.0, so no matter which newer driver you install, you will not get anything more than OpenCL 1.0 on that card.

But you knew that already Mike, I'm just a bit chatty :-)


HD 4xxx only has OpenCL beta Level support and now is officially depreciated by AMD.
http://developer.amd.com/TOOLS/HC/AMDAPPSDK/pages/DriverCompatibility.aspx

Which means it might work but must not.
I found out RV 770 works O.K.


Well, I guess I should upgrade my HD4850 to something newer. Reluctantly though, since it served me well, since 2008. I have to take a look at what I can get without totally emptying my bank account :-)

Richard Haselgrove
Volunteer tester
Send message
Joined: 3 Jan 07
Posts: 1119
Credit: 2,898,641
RAC: 1,288
Message 44101 - Posted: 17 Oct 2012, 16:21:16 UTC - in response to Message 44099.

...
The name of that workunit alone indicates a problem in the splitter that I will have to track down and fix.

WUs with that sort of name were first noticed on the main project, shortly after maintenance two weeks ago.

new WU numbering scheme

CH3-CH2-OH
Volunteer tester
Send message
Joined: 10 Mar 12
Posts: 810
Credit: 3,092,953
RAC: 786
Message 44102 - Posted: 17 Oct 2012, 16:22:21 UTC - in response to Message 44099.


There has to be some difference though between our results.

wuid=4153634


The first listed result (from host 59866) is missing a best_spike entry in its result file. I'm not sure why that would happen.

The name of that workunit alone indicates a problem in the splitter that I will have to track down and fix.


The name yes, I've been wondering about the new naming of WU's both here and on main. There seems to have been a change of how WU's are named. I have tons and tons of WU's with different naming types from what we've been used to, both here and on main. That's been going on for a couple of weeks now, and the new type of naming is now the norm rather than the exception

Example old type could be named:

05ap10al.31522.890.6.14.185_1

New naming example:

05ap10al.31522.1299.140733193388038.14.77_0

Profile cAnDYmanS@H-Beta
Volunteer tester
Avatar
Send message
Joined: 24 May 12
Posts: 38
Credit: 436,379
RAC: 0
Message 44103 - Posted: 17 Oct 2012, 16:29:03 UTC - in response to Message 44098.

Hi Mike,

Just as I came back because I had seen the "error of my ways", you had already replied. Indeed, being used to always keeping a cache of at least a day of work, I forgot that I recently changed my preferences to 0, because of the recent WCG HCC GPU app release.

Thanks anyway!
____________

Per aspera, ad astra!

Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar
Send message
Joined: 15 Mar 05
Posts: 1333
Credit: 15,718,393
RAC: 17,177
Message 44104 - Posted: 17 Oct 2012, 16:35:12 UTC - in response to Message 44102.

That huge number is supposed to be the ID of the receiver we used, which is obtained from a library routine. I'm guessing that at some point the library was compiled with a version of g++ that doesn't have the same storage format for the data type used.

Fortunately the correct values is used in the workunit header, so we don't have to do all this work over.
____________

Richard Haselgrove
Volunteer tester
Send message
Joined: 3 Jan 07
Posts: 1119
Credit: 2,898,641
RAC: 1,288
Message 44105 - Posted: 17 Oct 2012, 16:39:45 UTC - in response to Message 44104.

That huge number is supposed to be the ID of the receiver we used, which is obtained from a library routine. I'm guessing that at some point the library was compiled with a version of g++ that doesn't have the same storage format for the data type used.

Fortunately the correct values is used in the workunit header, so we don't have to do all this work over.

On the main board, Joe posted

The 140733193388045 is hexadecimal 7FFF0000000D so the excess is only in the upper 32 bits, I'm reasonably sure the proper channel indicator would be hex D or decimal 13 for that WU. For mental conversion, subtract 32 from the last two digits.

which perhaps points to a 32 bit/64 bit problem.

Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar
Send message
Joined: 15 Mar 05
Posts: 1333
Credit: 15,718,393
RAC: 17,177
Message 44106 - Posted: 17 Oct 2012, 17:29:56 UTC - in response to Message 44105.
Last modified: 17 Oct 2012, 17:31:20 UTC

Yes. So my assumption of a bad compile of the library was wrong. A fresh compile has the same problem.

It appears the g++ is taking the liberty of storing enums as 32 bit quantities, but when converted back to integer, the upper 32 bits of the integer get undefined values. Either that, or when presented with 'enum { A=0, B }' is assigning 0xffffffff00000001 to B rather than 1.

Or there could be a problem in the std::map container that is mangling the upper 32 bits of integers.
____________

Profile Raistmer
Volunteer tester
Avatar
Send message
Joined: 18 Aug 05
Posts: 1657
Credit: 10,492,706
RAC: 28,259
Message 44107 - Posted: 17 Oct 2012, 17:33:28 UTC

Back to topic:
As we can't compare results "on the fly" due to lack of info CUDA app provides (we discuss this on Lunatics intencively now) I propose to post links to suspictious workunits with ATi app involved (not results, workunits, please, to save time of reviewer) in separate thread. To organize testing more effectively.

I will start such thread soon. We need to form test case base for further investigation.

Profile Raistmer
Volunteer tester
Avatar
Send message
Joined: 18 Aug 05
Posts: 1657
Credit: 10,492,706
RAC: 28,259
Message 44109 - Posted: 17 Oct 2012, 17:39:55 UTC

Eric, ATI HD4650 should be excluded (as having workgroup size of 128 -the is known incompatibility with such cards).

TRuEQ & TuVaLu
Volunteer tester
Avatar
Send message
Joined: 28 Jan 11
Posts: 581
Credit: 1,675,967
RAC: 655
Message 44112 - Posted: 17 Oct 2012, 17:53:06 UTC - in response to Message 44100.

We're just guessing at the requirements for this version, so point me at any hosts that are repeatedly failing with this app version.


The app is fully OpenCL 1.0 compatible.

I tested on HD 4770 (RV 770) HD 4850 with driver 12.1.

Work group size 256 required.

I suggest driver version 12.1. and newer.

Running Windows XP driver 12.1 is the last driver with OpenCL support.

Mike


And I'm running my HD4850 on Vista with driver version 12.4, which also is an OpenCL 1.0 driver on the 4850. Apparently HD4850 doesn't support other than OpenCL 1.0, so no matter which newer driver you install, you will not get anything more than OpenCL 1.0 on that card.

But you knew that already Mike, I'm just a bit chatty :-)


HD 4xxx only has OpenCL beta Level support and now is officially depreciated by AMD.
http://developer.amd.com/TOOLS/HC/AMDAPPSDK/pages/DriverCompatibility.aspx

Which means it might work but must not.
I found out RV 770 works O.K.


Well, I guess I should upgrade my HD4850 to something newer. Reluctantly though, since it served me well, since 2008. I have to take a look at what I can get without totally emptying my bank account :-)


6670 with DDR5 memory if it isn't too slow.

Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar
Send message
Joined: 15 Mar 05
Posts: 1333
Credit: 15,718,393
RAC: 17,177
Message 44113 - Posted: 17 Oct 2012, 17:55:59 UTC - in response to Message 44107.

Back to topic:
As we can't compare results "on the fly" due to lack of info CUDA app provides (we discuss this on Lunatics intencively now) I propose to post links to suspictious workunits with ATi app involved (not results, workunits, please, to save time of reviewer) in separate thread. To organize testing more effectively.

I will start such thread soon. We need to form test case base for further investigation.


Good idea, thanks!
____________

Profile Eric J Korpela
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar
Send message
Joined: 15 Mar 05
Posts: 1333
Credit: 15,718,393
RAC: 17,177
Message 44114 - Posted: 17 Oct 2012, 17:58:28 UTC - in response to Message 44106.
Last modified: 17 Oct 2012, 17:58:53 UTC

Either that, or when presented with 'enum { A=0, B }' is assigning 0xffffffff00000001 to B rather than 1.


It's worse than that. When presented with 'enum { A=0, B }', g++ is assigning 0xffffffff00000000 to A and 0xffffffff00000001 to B.
____________

Profile Raistmer
Volunteer tester
Avatar
Send message
Joined: 18 Aug 05
Posts: 1657
Credit: 10,492,706
RAC: 28,259
Message 44116 - Posted: 17 Oct 2012, 18:05:06 UTC - in response to Message 44113.

Back to topic:
As we can't compare results "on the fly" due to lack of info CUDA app provides (we discuss this on Lunatics intencively now) I propose to post links to suspictious workunits with ATi app involved (not results, workunits, please, to save time of reviewer) in separate thread. To organize testing more effectively.

I will start such thread soon. We need to form test case base for further investigation.


Good idea, thanks!


Here: http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=1966

Profile cAnDYmanS@H-Beta
Volunteer tester
Avatar
Send message
Joined: 24 May 12
Posts: 38
Credit: 436,379
RAC: 0
Message 44119 - Posted: 17 Oct 2012, 18:41:41 UTC
Last modified: 17 Oct 2012, 18:43:01 UTC

Hey, hey! Got my first validated unit http://setiweb.ssl.berkeley.edu/beta/workunit.php?wuid=4157040! Everything seems to be working smooth here.

Good job!
____________

Per aspera, ad astra!

Profile Raistmer
Volunteer tester
Avatar
Send message
Joined: 18 Aug 05
Posts: 1657
Credit: 10,492,706
RAC: 28,259
Message 44120 - Posted: 17 Oct 2012, 19:00:13 UTC - in response to Message 44119.

Hey, hey! Got my first validated unit http://setiweb.ssl.berkeley.edu/beta/workunit.php?wuid=4157040! Everything seems to be working smooth here.

Good job!


LoL, one of still quite rare ATi vs ATi validations :)

CH3-CH2-OH
Volunteer tester
Send message
Joined: 10 Mar 12
Posts: 810
Credit: 3,092,953
RAC: 786
Message 44123 - Posted: 17 Oct 2012, 21:22:53 UTC

I must confess that I have been cheating...

I did one OpenCL 6.04 just now.

Sorry..

LOL

Previous · 1 · 2 · 3 · 4 · 5 . . . 7 · Next

Message boards : News : SETI@home v7 6.98 for ATI OpenCL released.


Main page · Your account · Message boards


Copyright © 2014 University of California
AstroPulse is funded in part by the NSF through grant AST-0307956