[UR][CUDA] Fix urUSMContextMemcpyExp synchronization issue#21602
Draft
[UR][CUDA] Fix urUSMContextMemcpyExp synchronization issue#21602
Conversation
cuMemcpy is synchronous with respect to the host, but it does not synchronize with device operations in other streams. This can lead to race conditions where urUSMContextMemcpyExp reads stale data if there are pending operations on the source or destination buffers. The issue manifests as sporadic test failures in CI where host_mem reads as 0 instead of the expected value (42), indicating the copy happened before the fill operations completed. Fix: Add cuCtxSynchronize() before cuMemcpy to ensure all pending device operations in the context have completed. This guarantees data consistency at the cost of a device-wide synchronization. Since urUSMContextMemcpyExp is not performance-critical and should provide strong consistency guarantees, this trade-off is acceptable. Fixes #19688 Test: exp_usm_context_memcpy/urUSMContextMemcpyExpTestDevice.Success now passes consistently on CUDA.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
cuMemcpy is synchronous with respect to the host, but it does not synchronize with device operations in other streams. This can lead to race conditions where urUSMContextMemcpyExp reads stale data if there are pending operations on the source or destination buffers.
The issue manifests as sporadic test failures in CI where host_mem reads as 0 instead of the expected value (42), indicating the copy happened before the fill operations completed.
Fix: Add cuCtxSynchronize() before cuMemcpy to ensure all pending device operations in the context have completed. This guarantees data consistency at the cost of a device-wide synchronization.
Since urUSMContextMemcpyExp is not performance-critical and should provide strong consistency guarantees, this trade-off is acceptable.
Fixes #19688
Test: exp_usm_context_memcpy/urUSMContextMemcpyExpTestDevice.Success now passes consistently on CUDA.