Expand description
Inter-Process Communication for CUDA events and allocations.
Two producers / two consumers of cross-process state:
- Device memory:
mem_get_handle→ transmit bytes to another process →mem_open_handlegives you a device pointer aliasing the same physical memory. Close withmem_close_handle. - Events:
event_get_handle/event_open_handlefor cross-process synchronization.
Availability: Linux only in practice. Windows drivers return
CUDA_ERROR_NOT_SUPPORTED. Use external-memory / semaphore interop
for Windows IPC.
Functions§
- event_
get_ handle - Export a CUDA event for sharing with another process.
- event_
open_ handle - Open a peer-exported event handle. Returns a raw
CUevent; wrap it with [Event::from_raw] if needed. - mem_
close_ handle - Release an imported device pointer previously opened via
mem_open_handle. - mem_
get_ handle - Export a device allocation for sharing with another process.
- mem_
open_ handle - Open a peer-exported device-memory handle. Returns a device pointer valid in the current context.