[CUDA] Fallback Event impl when there is no hardware cpu/gpu coherency #3070
+124
−57
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
zcbenz
force-pushed
the
fallback-event
branch
from
f713650 to
57f4623
Compare
zcbenz
force-pushed
the
fallback-event
branch
from
57f4623 to
38ed731
Compare
awni
reviewed
Jan 27, 2026
Comment on lines
+222
to
+225
| // 2. hostNativeAtomicSupported == true | ||
| // => use cuda::atom_ref on pinned host memory | ||
| // 2. no hardware cpu/gpu coherency | ||
| // => use cuda::atom_ref on device memory |
Member
There was a problem hiding this comment.
Which category (2 or 3) do typical consumer gpus fall into (4090, 5090 etc)?
Collaborator
Author
There was a problem hiding this comment.
4090 is cat 3, I don't know about 5090. Cat 2 is probably only theoretically possible, but it is very simple to support.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
For 4090 there is no hardware cpu/gpu coherency, so we can not synchronize between CPU and GPU by sharing atomics. This PR provides a fallback implementation of
AtomicEventfor this case: we store the atomic in device-only memory, and do CPU waiting with a busy-wait loop that reads the value withcudaMemcpy.This is a very inefficient fallback but it mostly serves to make tests pass,
AtomicEventhas always been a slow fallback fromCudaEventfor infrequent use cases.Also includes some small changes:
atomic_refinstead ofatomicto make initialization easier for device-only memory.