[UR] [V2] Add wait before enqueue in command buffer #17709
Conversation
| if (phEvent == nullptr) { | ||
| phEvent = &internalEvent; | ||
| } | ||
| UR_CALL(hCommandBuffer->awaitExecution(commandListLocked)); |
There was a problem hiding this comment.
This will block on host. The currentExecution event, if not null, should be simply added to the wait listwhen enqueuing the command list.
unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
Show resolved
Hide resolved
unified-runtime/source/adapters/level_zero/v2/command_buffer.cpp
Outdated
Show resolved
Hide resolved
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
Show resolved
Hide resolved
unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
Outdated
Show resolved
Hide resolved
unified-runtime/source/adapters/level_zero/v2/command_buffer.cpp
Outdated
Show resolved
Hide resolved
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
|
The failure on CI is not related to code changes and has been reported in #17066. @intel/llvm-gatekeepers please merge. |
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
Taken from intel#17709 Co-authored-by: Mikołaj Komar <[email protected]>
|
@intel/llvm-gatekeepers bump |
There is two CI failures on this PR. Please review them, fix them if they are related to the changes in this PR or add an explanation as to why the CI failure is unrelated before we can merge this PR. |
|
@sommerlukas Changes in this PR only affect unified-runtime L0 v2 adapter and are completely unrelated to the failures on CI (Jenkins failed due to internal error, and e2e failed on some unstable test, previously it was the one described in #17066, now it unexpectedly passes some different test) |
|
Moreover it passed CI on commit before, and the changes in the last commit are functionally cosmetical, without any influence on program logic |
|
Merged this now. @Xewar313 Can you please file a Github issue for the unexpectedly passing test, so we track this? |
|
I have filed the issue #17842 |
This PR disables flaky fill test for command buffer that was enabled in #17709. The issue is connected to the bug in the driver that is patched in new version, but the CI machines still have the old one, which causes it to sometimes fail (for example https://github.com/intel/llvm/actions/runs/14250564960/job/39942652796?pr=17836)
According to @MichalMrozek, zeCommandListImmediateAppendCommandListsExp has the same requirements as zeCommandQueueExecuteCommandLists, because of this, the command list must not be referenced by device when it is enqueued. This PR fixes this issue by adding event to synchronize append and execution