Since upgrading to the CUDA 4.0 release candidate I have noticed errors resulting from runtime checks to cudaGetLastError() which were not present in CUDA 3.2. I believe I have traced the problem down to an ‘invalid device pointer’ error code returned by cudaGetLastError() when used after a prior call to cufftDestroy(). It appears to only happen if the plan was created via cufftPlan2d() rather than a 1d fft. The return value from cufftDestroy() still indicates success.
The attached program should demonstrate the error when run under CUDA 4.0 RC (I haven’t tested against CUDA 4.0 RC2).
I am currently using a Tesla C1060 running Ubuntu 10.04
- Justin Voo
test.cpp (705 Bytes)