This is super interesting and potentially extremely useful as well. What is the performance improvements against doing it via API remoting or translating?
It is fast because the cuda workload isn't in the hot path. It only forwards iocrl control to setup channels, launching cuda kernels isn't in that path. the guest maps the real device memory, BAR writes from userspace go straight to hardware, and launching a kernel is a write to memory the guest already has in its VA space.
This is super interesting and potentially extremely useful as well. What is the performance improvements against doing it via API remoting or translating?
It is fast because the cuda workload isn't in the hot path. It only forwards iocrl control to setup channels, launching cuda kernels isn't in that path. the guest maps the real device memory, BAR writes from userspace go straight to hardware, and launching a kernel is a write to memory the guest already has in its VA space.