Skip to content

btl/self is not accelerator aware #12230

Open
@edgargabriel

Description

@edgargabriel

While debugging another issue on our systems, I realized that the btl/self component is not accelerator aware. As of right now, a send-to-self operation from gpu memory to gpu memory, data is copied first to host memory and then back to gpu memory. We should probably fix that and perform a direct accelerator.memcpy() operation, at least for contiguous buffers

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions