CUDA Runtime 4 Runtime Error - After Upgrading from CUDA 3.2 to CUDA4

Question

I recently tried updating my system from CUDA 3.2 to CUDA 4.0

This does not seem to be a very smooth update. First of all, many SDK examples fail. matrixMul and FFT, and it all comes back saying that the errors were too big, so they crash.

The code I wrote also has some problems. I had to track some errors caused by NaN that I never had, and now I get crashes in the kernel launch. Usually he says

cudaSafeCall() Runtime API error 4: unspecified launch failure. 

or

 cutilCheckMsg() CUTIL CUDA error : fftshift() execution failed : (4) unspecified launch failure. 

(fftshift is my function that has never been interrupted before, in 3.2)

Is there something wrong with my configuration?

I tried to use the driver 270, which was released from 4.0 for development, after being downgraded from the last driver, which, it seems to me, can cause problems.

Thanks for any advice.

Answer 1

I'm going to assume (since you did not specify) that you are using Windows (7), and are trying to use 3.2 SDK selections using the nvcc 3.2 compiler, but with drivers that support 4.0. If so, you should be aware that version 4.0 has a significant revision regarding memory addressing (namely, unified memory addressing on 64-bit systems) in addition to a whole group of other functions.

You want to remove the CUDA 3.2 driver, SDK and Toolkit driver (clear them forever). Then download and install all versions 4.0 above. New SDK samples should work if everything is installed correctly.