V7.1.38 cause UNKNOWN_ENUM (-10 = 0xfffffff6) on ATI

[MtM]

Sorry 11.7 = SDK 2.3, 11.9 was sdk 2.4

http://forums.amd.com/forum/messageview ... erthread=y seems SETI is also broken with 10.10 and upwards.

Great that it was reported by Raistmer. He's the developer of Lunatics optimized SETI applications for GPU (Astropulse and Multibeam which uses OpenCL on AMD GPUs).

I sent an email to Mike Houston to point him to our issues ...

Good, I would btw agree with a remark in that thread that a version which produced incorrect results ( which is even way worse then one which is bugged in that it fails to run something ) should be either removed as download or be accompanied with some warning about the broken compatibilities. Hope it gets fixed, people who use core_16 on amd might want to play the latest games as well, so for some downgrading or not upgrading isn't really a good option. Edit: removed concern about data validity concerning f@h.

I assume I can confirm it's the driver - the random blocks-iteration #38 failed with 995 (!) errors.

Should I downgrade or does FAH contact AMD to resolve the issue?
If that means I have to stay on "normal" GPU-WUs for a couple of weeks then that's fine with me..

I would downgrade simply because it getting fixed ( hopefully ) doesn't help you now, while those older drivers don't seem to have the issues reported. If you want to keep 11.11 due to some other reason, keep running core_11 work units.

[Eagle]

Thanks a lot, toTOW and MtM!

[toTOW]

I think the best solution for those who needs the latest drivers for games would be to go to individual downloads page (and to previous drivers page) and get the following :

1) Display driver 11.11
2) Catalyst Control Center 11.11
3) OpenCL Driver 11.9

And for people who doesn't game a lot, they should get the 11.9 full package ...

I'm waiting for an official word from Mike or Pande Group, but if it is confirmed, I'll make a more official post about the suggested strategy.

[Eagle]

Sounds good and, more importantly, useable to me!

[bruce]

Apparently the folks still running V6 have confirmed that it has nothing to do with V7. They've got another topic on the same Subject: early unit ends all the way since update to 11.11. The only real change is how the error is detected and what message is issued.

They pointed a link to this topic, but I suppose you folks can read the V6 discussion, too.

[toTOW]

Interesting : MemtestCL is working fine with 11.11 on my 6950 (but it can't allocate more than 768 MB of VRAM although I have 2 GB ...)

[Eagle]

ToTOW, it worked on my end, too when I ran a second 50-iterations-run a couple of hours ago.
Maybe it's not a persistent bug since it happened within the random-block-iteration?

[toTOW]

Well, the fact that the program doesn't reset my drivers is already a good point ...

I guess the bug is located in some OpenCL functions that MemtestCL doesn't uses, but which any scientific application does.

[MtM]

There are two issues here not one.

1. The issue of inaccurate results for SETI.
2. The fahcore_16 not starting.

2 fahcore_16 seems to fail to initialize the opencl driver, for Totow causing a hang/bsod for eagle causing just wu failures but no hang/bsod's ( correct me if wrong ).

After some googling I thought I found something suspicious, change in the way applications should initialize opencl functions, I even pm'd a link to MHouston asking if this was relevant, but after reading it again it was about a much older version change so...

Edit: so my comment about being worried about accurate results for f@h are misplaced, my apologies. Folding@Home doesn't use the openCL sdk in anything other then core_16 for Ati, the openMM cores for Nvidia use CUDA ( afaik.. pretty sure though ). And if the core doesn't work at all with the new drivers, I don't think data validity is a concern at this time.

[toTOW]

I've just tested the SETI application, and it is not crashing my drivers ...

So there is definitely two distinct issues here ...

[Eagle]

I switched to AMD SDK from Catalyst 11.9 (only OpenCL-package), but still:

Code: Select all

23:54:21:Unit 01:Called DecompressByteArray: compressed_data_size=44751 data_size=169787, decompressed_data_size=169787 diff=0
23:54:21:Unit 01:- Digital signature verified
23:54:21:Unit 01:
23:54:21:Unit 01:Project: 11294 (Run 7, Clone 244, Gen 121)
23:54:21:Unit 01:
23:54:21:Unit 01:Entering M.D.
23:54:23:Unit 01:Tpr hash 01/wudata_01.tpr:  1148122233 1600427989 4286651727 3553862433 1464123223
23:54:23:Unit 01:Working on PEPTIDE (1-42)
23:54:23:Unit 01:Client config unavailable.
23:54:23:Unit 01:Starting GUI Server
23:54:39:FahCore, running Unit 01, returned: UNKNOWN_ENUM (-1073740777 = 0xc0000417)

Additionally, the driver-package doesnt' show up within "Add Remove Programs & Drivers" anymore..

[toTOW]

Damned, that would mean that the issue is in the Display Driver and the OpenCL one ?

[Eagle]

[MtM]

Try and remove all ATI applications and drivers in this way -> link

Then install a clean 11.9. Run my diagnostics as it will report the driver versions so we can be sure everything went as well. Try to fold.

If it folds, try installing only the driver package from the 11.11 driver. Run the diagnostics again, and try to fold. I am not sure that installing only the driver package will leave the compute runtimes unchanged, in fact I'm sure CAL is included in the driver so maybe some of the openCL components or dependencies will be changed as well? Not that diagnostics will show that, only reported version numbers. If the changes are not shown, checking each dll version from one system state to another.. ouch.

[mhouston]

OpenCL runs uses CAL components and OpenCL is integrated into the downloadable catalyst drivers.