CUDA streams and error handling

[striker159]

cudaStreamSynchronize does not reset the error state. Use cudaGetLastError().

Kernel errors are sticky errors and cannot be recovered.

@Robert_Crovella

Sticky errors are those errors that corrupt the CUDA context. In the CUDA runtime API, a corrupted CUDA context is non-recoverable. The only method to recover from it is to allow the owning process to terminate.

Yes, all asynchronous errors are sticky. Only asynchronous errors are sticky.

Asynchronous errors are those errors that occur as a result of device code execution itself.

It is true in some sense that a call to cudaGetLastError() “resets” the error. However in practice, it is self-evident that it does not reset the sticky error in any practical way. The same error will reoccur (i.e. be reported again) if any non-trivial usage of the CUDA runtime API is attempted. For all practical purposes the context is corrupted by the sticky error and is useless.

What is the exact error message you are experiencing?