Global Memory Shared between Multiprocessors

[ksampath]

Suppose I allocate a certain amount of global memory which contains data that is intended to be shared across all multiprocessors, is it guaranteed that whenever the data is updated in global memory the data will be ready and available immediately across all multiprocessors. If this is not the case what is the best way to share data between multiprocessors via global memory?

[AastaLLL]

Hi,

What kind of memory do you want to share between processors?
CPU buffer or GPU buffer?

Thanks.

[ksampath]

Hi AastaLLL,

A GPU buffer.

[AastaLLL]

Hi,

The IPC library will handle the synchronization for you.
So you can assume the data will reflash immediately.

There are some possible IPC mechanisms on the Orin:

• NvBufSurface: Jetson Linux API Reference: Buffer Management API module | NVIDIA Docs

• NvSci: https://developer.nvidia.com/embedded/downloads#?search=nvsci

• cuMemExportToShareableHandle/cuMemImportFromShareableHandle:
  CUDA Driver API :: CUDA Toolkit Documentation

Thanks.

[ksampath]

Thank you for this information.