Skip to content

Conversation

@simeonschaub
Copy link
Member

@simeonschaub
Copy link
Member Author

Looks like this is not quite enough. If we are to follow CUDA, the stream should be synchronized whenever a regular pointer is taken from unified or host memory. We'll probably have to implement something similar to https://github.com/JuliaGPU/CUDA.jl/blob/7de34f20f8842a256e439941d154f888060504f6/src/memory.jl#L494 for that to have the correct semantics.

@luraess
Copy link
Member

luraess commented Nov 11, 2025

Could be needed at this point indeed. If going this way, one should make sure to also provide an opt out mechanism from possible implicit syncing (cf JuliaGPU/CUDA.jl#2662) as this may hinder some use cases as discussed here PTsolvers/Chmy.jl#65.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

support unified memory

2 participants