oneapi-src / oneapi-src/unified-memory-framework
Mapping to host
Nobody has claimed this yet.
- Dominant language
- C
- Stars
- 98
- Forks
- 48
- Avg merge
- 3d 20h
- Merged PRs (30d)
- 6
Description
Rationale
Today, the MPI library for Level Zero and CUDA can map the device memory to the host. It is done for performance reasons: In case of short messages mapping to host and calling host memcpy works faster than copying via L0 APIs.
Description
TBD
API Changes
Need to decided on the right API
Implementation details
In case of Level Zero the Level Zero IPC API can be used. Today MPI extracts file descriptor from the IPC handle and use mmap function to map memory to the host.
In case of CUDA the gdrcopy library is used.
Meta
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reviewing the existing MPI paths for Level Zero and CUDA device-memory transfers, then compare the Level Zero IPC plus mmap approach with the gdrcopy integration described here. The API is still undecided, so define the supported mapping interface and its scope before implementation. Done requires an agreed API and working host mapping for both backends, with performance behavior validated for short messages.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- c
- Domain
- performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100