Skip to content

[UR][L0 v2] Work around a GPU driver timestamp write on some platforms - #23186

Open
againull wants to merge 1 commit into
intel:syclfrom
againull:tag_wa
Open

againull wants to merge 1 commit into
intel:syclfrom
againull:tag_wa

Conversation

@againull

Copy link
Copy Markdown
Contributor

On some platforms the device's 8-byte write of a profiling tag's recorded timestamp also rewrites the rest of the aligned 64-byte block around it, with stale contents. It only shows up when the two happen at the same time: the device landing a timestamp write while the host writes the event state sitting next to it, which is what recycling a pooled event for a new recording does. Those host writes are then lost - including the flag marking the event as timestamp-recording, so the fresh recording reports UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE.

Give the timestamp a 64-byte block of its own to work around this. Note there is no race on the timestamp itself: recordings sharing a destination are serialized - by the queue for an in-order queue, and by the barriers for an out-of-order one.

Driver issue is tracked as GSD-13449.

Assisted-By: Claude

On some platforms the device's 8-byte write of a profiling tag's recorded
timestamp also rewrites the rest of the aligned 64-byte block around it,
with stale contents. It only shows up when the two happen at the same time:
the device landing a timestamp write while the host writes the event state
sitting next to it, which is what recycling a pooled event for a new
recording does. Those host writes are then lost - including the flag marking
the event as timestamp-recording, so the fresh recording reports
UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE.

Give the timestamp a 64-byte block of its own to work around this. Note there is no race on the
timestamp itself: recordings sharing a destination are serialized - by the
queue for an in-order queue, and by the barriers SYCL brackets each tag with
for an out-of-order one - so the last write to land is the current
recording's.

Driver issue is tracked as GSD-13449.

Assisted-By: Claude
@againull
againull requested review from a team as code owners September 15, 2026 18:28
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.

1 participant