Questions and Answers :
GPU applications :
ATI GPU help please
Message board moderation
Author | Message |
---|---|
rwarren873 Send message Joined: 2 Sep 06 Posts: 25 Credit: 39,741,063 RAC: 87 |
I apologize for what may (most likely are) be dumb questions... I was very happy to see ATI supported by stock Boinc because the Lunatics instructions never made sense to me, but I'm having troubles. I've been lurking for years and have been actively reading (though not always understanding) the posts here and I just don't know what I can do to fix these issues. All of my machines are running Catalyst 12.6 Any assistance would be most appreciated. My computer (http://setiathome.berkeley.edu/show_host_detail.php?hostid=6585335) is happily chugging away at the new tasks. It is a 6core AMD machine with 8gig RAM and a Radeon 6700 (Juniper). My wife's computer (http://setiathome.berkeley.edu/show_host_detail.php?hostid=6733302 threw nothing but errors. http://setiathome.berkeley.edu/result.php?resultid=2562057385 and I finally just changed the "location" of that computer to a new one that does not allow the use of the GPU. Her machine is a 2core AMD machine with 4gig RAM and a Radeon 4650. Finally, my home server (very slow machine) with 1core and 3gig RAM with a Radeon HD 5450 refuses to even download the new ATI tasks. http://setiathome.berkeley.edu/show_host_detail.php?hostid=6585230 If other information is needed I can post it but I didn't want to be too complicated on first glance. |
rwarren873 Send message Joined: 2 Sep 06 Posts: 25 Credit: 39,741,063 RAC: 87 |
Disregard the last computer, my home server. Apparently there's some issue between the video card/drivers, WinServer2008 and BOINC. For some reason BOINC isn't recognizing it as an OpenCL GPU. |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
That problem will most likely be the drivers. Windows allowed drivers, even the WHQL, lack the OpenGL and OpenCL components as these are direct competitors to DirectX and Direct Compute. So you'll need drivers from AMD themselves, but here's the problem that AMD does not make GPU drivers for Windows Server versions. Not for WinServer 2008, not for R2. So then you'll need to install BOINC in Windows 7 (64bit) compatibility mode and install the drivers for Windows 7 (64bit), and then hope that the Windows architecture is about the same that the drivers will work. |
rwarren873 Send message Joined: 2 Sep 06 Posts: 25 Credit: 39,741,063 RAC: 87 |
I'm contemplating just installing Win7 on that machine just because of the driver issue with WinServer2008R2. I'm much more interested in why the 2nd machine can't seem to process a GPU unit at all. It seems it doesn't even try, a couple of seconds after starting, it errors out. The error log in my first post is typical, but I have no idea what the error means or how to fix it.
Obviously it looks like a memory problem, but the GPU has 1024MB onboard and the machine itself has 4gig RAM. I thought at first it may be a problem with the hybrid dual-GPU the machine came with that combined an on-motherboard GPU with the 4600 series card, but I disabled the on-board in bios. In the error it talks about max_kernel_wg_size, but the only other reference to wg I can find says "max WG size: 128" so I don't know if that's causing it to self-destruct and if so, how to change it. |
OzzFan Send message Joined: 9 Apr 02 Posts: 15691 Credit: 84,761,841 RAC: 28 |
The "WG" or workgroup size must be 256 or larger. If it says 128, your GPU cannot be used. At least that's the way it was when the app was being developed, so I believe this to be true unless someone else knows otherwise. |
Mystic. Send message Joined: 28 Aug 99 Posts: 2 Credit: 1,780,102 RAC: 0 |
I have the same problem with my computer: WARNING: patching required max_kernel_wg_size=64 ERROR: patching required max_kernel_wg_size=64 for binary cached kernels! Unhandled Exception Detected... - Unhandled Exception Record - Reason: Out Of Memory (C++ Exception) (0xe06d7363) at address 0x7591B9BC Only it errors out with size=64 It does state earlier that Max WG size:256 however. Where would you patch this? |
OzzFan Send message Joined: 9 Apr 02 Posts: 15691 Credit: 84,761,841 RAC: 28 |
A patch would have to be written first. I do not believe one has been written. In fact, the ATi GPU application was written by a third party and was freely given to SETI@Home to use as a "stock" application (after rigorous testing was done to ensure that the application met the standards set by the SETI@Home team). This third party person (Raistmer, I believe) would have to write new code to support low workgroup GPUs. |
BilBg Send message Joined: 27 May 07 Posts: 3720 Credit: 9,385,827 RAC: 0 |
This third party person (Raistmer, I believe) would have to write new code to support low workgroup GPUs. But his tasks show "Max work group size: 256" http://setiathome.berkeley.edu/result.php?resultid=2571256428 http://setiathome.berkeley.edu/result.php?resultid=2570688972 I don't know what "WARNING: patching required max_kernel_wg_size=64" means. I will try to PM Raistmer to direct him to this thread. The only other place I found max_kernel_wg_size (in 1 year period) is: http://setiathome.berkeley.edu/forum_thread.php?id=68675&postid=1258896#1258896 Â - ALF - "Find out what you don't do well ..... then don't do it!" :) Â |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
I have the same problem with my computer: Previous reports were from low-end GPUs with max WG size of 128 and tasks were completed OK on those hosts in spite of this precaution warning. So it was leaved as is, but apparently not in all cases this precaution warning can be safely ignored. "patching" is done inside oclFFT kernel generation routine and means that after initial kernels generation routine discovered that kernels need change to be runnable on device. Perhaps, binary cache recives wrong kernel version in this situation. I will try to do fix for this issue but this can take some time. What to try right now: 1. To stop Ati GPU processing on affected host. It's easiest but least productive. 2. Try to recive more info about issue - is it binary cache connected or not: suspend all GPU tasks, go to project directory and delete all oclFFT*.bin files, unsuspend single GPU task and let it run or error, report outcome to server, give link to result. 3. If you prefer to stay on "Stock" path try to replace AP stock executable with r455 AP gpu opt file (via renaming). That build has no binary cache as far as I can recall and can be free from this issue. But it would be easier to download and run Lunatics installer. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
The reason why patching required in this case - HD4xxx family of ATi GPUs has no real local/shared memory (they have fast LDS memory but it doesn't meet OpenCL requirements for local memory so was never exposed to OpenCL runtime. OpenCL can't use all available hardware resources for these GPUs). It emulates this kind of memory via global GPU memory. Consequence of this (besides much longer shared memory access times and performance issues in kernels that uses shared memory) - workgroup size for kernel that uses shared/local memory can not exceed 1 wavefront size, i.e., 64. oclFFT kernels use local memory so their workgroup should be 64 even if GPU itself supports max WG size of 256. |
Mystic. Send message Joined: 28 Aug 99 Posts: 2 Credit: 1,780,102 RAC: 0 |
I suspended the gpu tasks and deleted the oclFFT*.bin file and unsuspended gpu. It errored out with the same error after recompiling a new file( I think). result link is http://setiathome.berkeley.edu/result.php?resultid=2576204752 I stopped gpu tasks for now. Thanks for all the info. I'll check out the Lunatics installer. |
rwarren873 Send message Joined: 2 Sep 06 Posts: 25 Credit: 39,741,063 RAC: 87 |
I had no luck with that solution either. I'm going to swap the 5450 out of my WinServer machine that won't do GPU tasks anyway with the 4650 that's in my wife's machine this weekend. Thank you all for the help! |
OzzFan Send message Joined: 9 Apr 02 Posts: 15691 Credit: 84,761,841 RAC: 28 |
I have three HD 4650s and all of them have a workgroup size of 128 and won't crunch. I don't think you'll have any luck with that either. |
rwarren873 Send message Joined: 2 Sep 06 Posts: 25 Credit: 39,741,063 RAC: 87 |
Well, wife's machine is happy now using the stock BOINC ATI app with the HD5450 from my server. The server has the 4650 which won't do GPU, but driver issues with Windows Server 2008R2 prevented that anyway. Now if only SETI would get over what ever is preventing new GPU tasks being sent out, I'd be good to go :) Thank you all for the assistance! |
dragon1 Send message Joined: 17 Sep 05 Posts: 33 Credit: 4,438,013 RAC: 0 |
Just coming into this late and have recently installed a Radeon 4650 into my system which is a Core 2 duo on Windows XP home. Those cores are both crunching but BOINC says I have "no useable" GPU. My preferences are set to use GPU. Is there any way to get the Radeon 4650 GPU to run while running SETI@home (non BETA)? Or is it just not up to the job. The system previously ran an NVidia 8300 GPU which was not a CUDA acceptable processor apparently. |
OzzFan Send message Joined: 9 Apr 02 Posts: 15691 Credit: 84,761,841 RAC: 28 |
If BOINC isn't detecting a usable GPU, you'll need to make sure you're installing the OpenCL compatible drivers for Windows XP, of which the last version to support OpenCL for XP was Catalyst 12.1. But having three ATI Radeon HD 4650s, I can tell you that the stock app cannot use them anyway. You would need to install a third-party app that will utilize the Radeon's CAL instruction set for AstroPulse-only tasks. ..and yes, the nVidia 8300 was not CUDA compatible. Only the 8800 series and newer support nVidia's CUDA initiative. |
arkayn Send message Joined: 14 May 99 Posts: 4438 Credit: 55,006,323 RAC: 0 |
|
©2024 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.