Use device managers in allocators instead of setting device directly #796
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
It seems the CUDA and HIP allocators have invoked
cudaSetDevice()
andhipSetDevice()
directly, which they should never do because it can invalidate the state of thehip_device_manager
andcuda_device_manager
, which are used to manage the active device by the rest of the runtime.This is a terrible oversight. I don't know why these function calls are used here instead of the device managers; the files even include them but then don't use them. This seems to be a very old bug that may have been as old as when we introduced the current runtime.
There might be a case for invoking both the
device_manager::activate_device()
as well ascuda/hipSetDevice()
directly, as this would make USM memory management functions more robust in backend interop scenarios where the user also sets device explicitly.I believe this fixes the issues described in #770
CC @al42and