Vista driver restart

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m running the 174_55 beta drivers for CUDA on Vista 32 bit.

I’m currently trying to port some existing algorithms to CUDA, and in the process I’m getting alot of GPU crashes.

If I was writing everything from scratch, I don’t think the crashes would be too hard to track down, but since I’m moving a larger piece of code it’s more difficult to isolate them. (Everything runs fine in the emulator, so it’s device only crashes - probably memory access related).

In any case, when I crash the card / driver, my screens flicker and Vista eventually restart the driver.
The problem is, the driver is restarted in a bad state. After a driver restart, any use of Direct3D will in turn crash the driver again (my application requires both CUDA and Direct3D, the Vista desktop also suffers occasional crashes). Running some of the CUDA standalone samples that don’t present any 3D graphics seem fine.
I assume this is probably a known issue, but I’m quite curiuos why the driver upon reload doesn’t start clean, or whether it could be an issue with Direct3D.

2

I have seen this behavior in Linux when writing past the end of an allocated array on the GPU. I attribute it to the out of bounds memory write messing with some memory on the GPU that the driver expects to be initialized.

Unfortunately, since you are on Vista, you can’t use valgrind on the emulation mode to debug these problems. Is there another memory bounds checking debugger you can use?

3

same problem here on Vista64… :(