I have seen that there is a hard runner of 5 to 15 seconds before the CUDA applications fail I realize that this is not ideal for running the CUDA application, but assuming that it is the right choice for using the CUDA and due to the quantity of thread per sequential work it should be run at that time, to expand this time Is there a way to get around it?
I am not a CUDA expert, --- I am developing AMD stream with SDK, Which AFAIK is roughly comparable
You can disable the Windows Watchdog timer, but for these reasons is not highly recommended it should be clear. To disable it, you need to regedit to HKEY_LOCAL_MACHINE \ SYSTEM \ CurrentControlSet \ Control \ Watchdog \ Display \ DisableBugCheck , create a REG_DWORD and set it to 1. You may need to do anything in the NVidia control panel. See some references to "VPU Recovery" in the CUDA docs.
Ideally, you may be able to break your Kernel Operations in many ways on your data so that the operation limit on time runs.
Alternatively, you can split the problem domain so that it can calculate the output pixel command per less. Ie, 1,000,000 output pixels in one fell swoop instead of computing, 10 commands in the GPU continue to calculate 100,000 each.
The basic unit that fits within a piece of time is not your entire application, but the execution of a single order buffer. In the AMD stream SDK, a long sequence of operations can be broken into slices several times by clearly flipping the command line with CTX (flush) calls. Maybe the CUDA has something similar?
You should not to read all your data on the piece of your PC at all times; You can leave your texture in the GPU local memory etc; You occasionally complete some command buffers, to prove to the OS that you are not trapped in the infinite loop.
Finally, GPUs are fast , so if your application is not able to do useful work in that 5 or 10 seconds, I take it as a sign that something That's wrong.
[Update on March 2010 to update: ] The key to the registry top is old I think it was the key to Windows XP 64-bit Vista and Windows 7 There are new registry keys for. You can find it here: or here:
[Update on April 2015 to update: ] This is actually outdated TAD for disabled programming The easiest way to do this is to assume that you have an NVIDIA analytics tool installed, open the NSEIT monitor, click "Ensite Monitor Options" and open "WDDM TDR enabled" false under the "normal" set. This will change the registry setting for you, and no reboots in the reboot TDR registry setting will not take effect until you reboot.
Comments
Post a Comment