nppiFilterSobelHoriz_8u16s_C1R crashes after cudaDeviceReset

Our process runs fine without any errors using nppiFilterSobelHoriz_8u16s_C1R and nppiFilterSobelVert_8u16s_C1R.
However if we call cudaDeviceReset to change the specific GPU or turn GPU processing on or off, when we reinitialze the GPU with cudaSetDevice it fails with NPP_CUDA_KERNEL_EXECUTION_ERROR (-1000) even though other kernels that we wrote by hand or other NPP functions will still launch.

Does NPP keep some type of state that needs to be reset?


There was a problem with NPP and cudaDeviceReset that was fixed in the latest CUDA 10.1 release.

If you’re not using the latest CUDA version (10.1.168), please retest with that.

If you still witness the problem with the latest CUDA version, please provide a short, complete code that demonstrates it. Alternatively you may wish to file a bug, the instructions are linked at the top of the CUDA programming forum.

I have upgraded to the latest Cuda Toolkit 10.1 hoping that this would fix the issue, however, it didn’t.
I will file a bug report.