Skip to content

Fix-11260-launchhost#1127

Merged
pvelesko merged 2 commits intomainfrom
fix-11260-launchhost
Feb 16, 2026
Merged

Fix-11260-launchhost#1127
pvelesko merged 2 commits intomainfrom
fix-11260-launchhost

Conversation

@pvelesko
Copy link
Collaborator

@pvelesko pvelesko commented Dec 20, 2025

@pvelesko pvelesko force-pushed the fix-11260-launchhost branch 2 times, most recently from 3098147 to b7c539b Compare January 13, 2026 17:16
@pvelesko pvelesko force-pushed the fix-11260-launchhost branch 4 times, most recently from 3bb2d3b to 0fe4c20 Compare January 29, 2026 02:05
- Adds explicit host synchronization on queue sync events before adding barriers that wait on them.

- When addDependenciesQueueSync appends a signal event to another queue's
command list, the GPU may not process it before we call
zeEventHostSynchronize. Add a zero-timeout flush to force the driver
to submit the signal command immediately.

- Modified the synchronization mechanism for queue events to use a short timeout instead of blocking indefinitely. This change addresses potential hangs by ensuring that the GPU-side barrier maintains properxq ordering without waiting indefinitely for events to be ready.
@pvelesko pvelesko force-pushed the fix-11260-launchhost branch from 0fe4c20 to 769ddaf Compare February 12, 2026 11:07
addDependenciesQueueSync and addCallback previously flushed only the
active queue (get()). When a stream switches to profiling mode after
hipEventRecord, ClProfilingQueue_ depends on ClRegularQueue_ via
switchModeTo; both must be flushed for cross-queue deps to work on
Mali. Fixes TestHipLaunchHostFuncMultiStream timeout on Mali-G52.
@pvelesko pvelesko merged commit fa2d862 into main Feb 16, 2026
19 checks passed
@pvelesko pvelesko deleted the fix-11260-launchhost branch February 16, 2026 10:16
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.

hipLaunchHostFunc Hangs Intermittently on Aurora with Level Zero Backend

1 participant

Comments