There is hope to expose a feature in CUDA of shared contexts such that two (or more) contexts can share the same address space and device pointers returned by one context can be used on the other?
If this contexts are running on the same device that could be done efficiently if the pointers map to device memory…
The idea is similar to OpenGL shared contexts that can share efficiently resources… Exposing it also would allow to transparently use fast (DMA?) transfers between devices without going to host once hardware/CUDA stack support it…
also this seems now that pinned memory has been made portable a generalitzation of it in device memory could be called “portable device memory”…
If this is much work at least could be useful some way to retrieve a real adress of device memory or a pointer that’s not contex specific but device specific and memory copy between this type of pointers…
Altough I don’t know if it can be broadly useful or if it makes sense at all. This can be good examples:
- As said a transparent way to take advatange of future(?) imporovements in device to device transfers (hardware accelerated and without goint to memory…)
- A library using CUDA creates its own context for avoiding the program using it of sniffing his device memory space but then wants to expose
the “results” as a pointer on device memory (for avoiding transfering back to host) that can be used by the other CUDA contexts.
Can please some at Nvidia say if it makes sense at all this proposal…