Cuda 3.1: must cudaStreamCreate/cudaStreamDestroy be called in same (Windows) thread?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi there,

I’m currently changing from CUDA 2.3 to 3.1 and get access violations when calling [font=“Courier New”]cudaStreamDestroy()[/font]. I assume that this is because [font=“Courier New”]cudaStreamCreate()[/font] and [font=“Courier New”]cudaStreamDestroy()[/font] are called from different Windows threads.

Did the CUDA runtime change and now requires the same thread for [font=“Courier New”]cudaStream[/font] (and [font=“Courier New”]cudaEvent[/font] as well, btw) creation and destruction?! Where would that be documented? Or isn’t it the different threads that cause my crashes??

2

That’s always been true. Contexts are bound per thread unless you explicitly move them, so calling cudaStreamDestroy from a separate thread never did what you thought it did (it would create a context on device 0, try to destroy a stream, realize that the stream doesn’t exist, and fail to do so).

3

Well, ok. But in 2.3 nothing happend (neither a crash nor an error was returned)…

4

Different question: How costly is the stream creation?! Would it be feasible to create/destroy the stream every frame? Or would that drop performance noticeably?