cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

captian-n
Adept II

12.3 RC 8.95 - BSOD with multiple device - multiple context - multiple threads

Hi all,

I develop at the moment with new HD 7950 and I try the new 12.2 pre-certified driver now 12.3. My application is multiple context with multiple threads style and until now it worked all the time without problems. With the new driver it chrashed with a blue screen if I release the complete stuff with a call like that:

parallel_for(0,static_cast<int>(pOpenClDevices.size()),[&](int l)

{

    clReleaseKernel(pOpenClDevices.OpenCLBackPro);

    clReleaseKernel(pOpenClDevices.OpenCLBackPro2);

    clReleaseKernel(pOpenClDevices.OpenCLResetImage);

    clReleaseKernel(pOpenClDevices.OpenCLReset);

    clReleaseKernel(pOpenClDevices.OpenCLScaling);

    clReleaseContext(pOpenClDevices.Context); 

});

I use lambdas in combination with the PPL from Microsoft VS 2010. This just an example it sometimes also crashes during a parallel read back. Have any else that multi threading problem is there a workaround. Single threading is no solution it works but for HPC issues with up to 8 GPUs per machine the increasing of calculation time is to big. It runs also with one device without errors or with an older driver.

Thanks in advance

0 Likes
1 Solution
captian-n
Adept II

With the 12.4 Beta from 15. March (8.96-120315) the BSOD issue seems to be solved. I tested it now for several cases and it works fine. But now the core clock and memory clock rate of the Disabled adapters (that one without monitors) stays at idle speed with an OpenCL load. Only the adapter with monitor increase the the clock rates  Knows someone a solution how can I increase it for OpenCL, without plugin a monitor or monitor simulator? May be with OpenCL commands?

View solution in original post

0 Likes
11 Replies
captian-n
Adept II

Hi I made small demonstrator for that issue the source code is attached to this post. I hope some of you can reproduce this bug without to be angry about me, because the result is a blue-screen. You need 2 GPUs to get a blue-screen, because in my opinion it is multithreading bug in the driver.

My test equipment: HP Z800 64GB, 2x XEON X5660, 2x AMD Radeon HD 7950, WIN 7 X64 Business, VS 2010. The application should be started without development environment, because it seems to be a synchronization problem it not happens on the same point in code. The application is build as 64bit with VS2010. VS2010 is also the minimum requirement, because for multithreading PPL is used. If there are some question left please ask me. Thanks

0 Likes

Today I find another application have the same problem like me. I try to run SiSoftSandra (san2012sp2-1830.exe) benchmark, with same result. Easy to reproduce. Start SiSoftSandra go to Benchmark->GP (GPU/CPU/APU). (Type: OpenCL GP Processor) (Graphic Processor: 2x AMD Radeon HD 7900 (3584SP 56C 800MHz, 6GB)) and than press Refresh button or F5. After some calculation time same result like my own software. I hope some of the official AMD moderators would say something to that case, after a well known software has the same problem. For me it is important, we are in front of commercial software launch and I hope the problem could be solved in the next two month.

0 Likes

Thanks for reporting this, we have reproduced it internally.

Thanks for information.

0 Likes

Hi,  do you remember the BSOD number? and what is the configruation about the 2x 7950? crossfire mode or mulit-adapter mode ?

0 Likes

Sorry for the late reply. It is a non crossfire configuration (multi-adapter).  I have reproduced a BSOD for you, but the number change all the time. I hope it helps.

BSOD.jpg

0 Likes
captian-n
Adept II

With the 12.4 Beta from 15. March (8.96-120315) the BSOD issue seems to be solved. I tested it now for several cases and it works fine. But now the core clock and memory clock rate of the Disabled adapters (that one without monitors) stays at idle speed with an OpenCL load. Only the adapter with monitor increase the the clock rates  Knows someone a solution how can I increase it for OpenCL, without plugin a monitor or monitor simulator? May be with OpenCL commands?

0 Likes

0 Likes

Thanks for that hint. That is what I meant with monitor simulator. I use it for testing it works good, but that is nothing what I would recommend our customers.

0 Likes

Just a me-too comment.

I have the same crashing problem with the same drivers in a multiple 7970 system. The March 15 drivers also fix the crashing issue but show the same clock problem, core clock is locked at 500MHz on headless gpus. 500MHz is not idle speed and keeps the gpus a bit warm when not in use.

Unfortunately the slow clock on one gpu slows down everything when synchronizing multiple devices.

0 Likes

Your are right the idle core clock is about 300 MHz. And in my case the slow clocked GPUs also slows down the complete process because of synchronizing.

0 Likes