Strangley all the evidence I’m amassing from several attempted implementations seems to suggest no! If not why not? It seems especially strange especially when considering that in order to use the async API you have to do host memory allocs using CUDA and even these cannot be done on another thread.
I hope I’m not missing something in the picture completely but all the evidence I’ve seen seems to point to CUDA not recognising handles, memory, etc. that are fine when used in the original thread that created them.
Any help is much appreciated, thanks,