New runtime issues on cuda 3.2?

Hi all,

I have some code that I have been using for a while under cuda 3.1 and previous versions. After upgrading to 3.2, I can still compile, but now get an “unspecified launch failure” from my kernels.

I naturally went through the documentations and took a look at my kernels, but I can’t see anything incorrect (not surprising since the kernels have been in use for a while already). I am not using the Driver API.

Does anyone have any suggestions for what to look out for here? What changed in 3.2 that might be breaking my code?

fyi: I’m using a multi-Tesla c2050 system with ubuntu 10.04. I have the same problems on a different system with a GTX285.

Try running cuda-memcheck and see what it does. Normally those sort of unspecified launch failures come from out of bounds memory access. Compiler or runtime library differences might be enough to expose something that wasn’t noticeable with the older toolkit.

I had a similar problem too, but related to the same device. After a cuda memory reset I had no more launch failures and the correct results out of my code.

Yeah, you nailed it. Thanks!