[maleadt]

Fixes #1027.

When appendCommandLists has nothing to dispatch from an idle in-order immediate command list, and the regular command list's batch buffer is dispatched as primary, the queue passes the regular list's first command buffer to flushTaskHeapful as an exactly full task stream (firstCmdListStream), with the immediate list's own stream as optionalEpilogueCmdStream. If the CSR replaced its command buffer during an earlier flush, ucResourceRequiresTagUpdate is set, and the tag update was written into that full task stream, so LinearStream::getSpace aborts. This crashes the oneAPI.jl test suite sporadically, since it submits kernel launches in batches of regular command lists.

This writes the tag update into the epilogue stream when there is one, as the barrier with post-sync in flushTaskHeapful and the UC tag update in flushBcsTask already do. The tag update then follows the task in the same submission. The immediate stream already has room for it: appendCommandLists reserves space there for the epilogue commands, including the barrier.

The new ULT fills the task stream, sets ucResourceRequiresTagUpdate and passes an epilogue stream to flushTask; it aborts without the change and finds the tag update in the epilogue stream with it.

flushTaskHeapless also writes the tag update into the task stream unconditionally but doesn't handle an epilogue stream at all, so I left it alone; I only have heapful hardware (Tiger Lake) to test with.

Tested on 26.35.39758.10, since master doesn't build in our older-glibc environment, for Gen12LP:

• shared ULTs: 12054 passed, 844 skipped, 0 failed with the change; the new test fails without it (the same linear_stream.h:90 abort).
• L0 core ULTs: 5134 passed, 478 skipped, 0 failed with the change.
• The reproducer from [GSD-13815] Appending a regular command list to an idle immediate command list aborts in LinearStream::getSpace when a UC tag update is pending #1027 runs 20,000 iterations with the change (without, it aborts within the first thousand). The same patch, shipped in Julia's NEO_jll 26.35.39758+2, fixes the crash in oneAPI.jl.

[jwarchul]

Hi @maleadt,

Thank you for the contribution.

Merged in commit 443de44.