[SYCL] cmd mem leak fix #17125
[SYCL] cmd mem leak fix #17125
Conversation
…e thrown. In that case, the command will have a failed EnqueueStatus. During the clean up, we don't want to reenqueue it if we know it has failed before
cperkinsintel
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
cperkinsintel
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
cperkinsintel
requested review from
againull,
KseniyaTikhomirova and
sergey-semenov
and removed request for
againull
|
ESIMD/matrix_transpose_glb.cpp is failing on other PR ( https://github.com/intel/llvm/actions/runs/13465061599/job/37630646656?pr=17126 ) I haven't been able to reproduce it locally yet, maybe Arc only? |
|
HostInteropTask/interop-task-cuda-buffer-migrate.cpp is failing in other unrelated PR also on CUDA ( https://github.com/intel/llvm/actions/runs/13510530573/job/37751750445?pr=17152 and https://github.com/intel/llvm/actions/runs/13509301482/job/37749427764?pr=17149 ) Issue is here: #17026 |
There was a problem hiding this comment.
LGTM. Please adjust the PR title ("only" wouldn't make any sense in the commit history).
I would also like @KseniyaTikhomirova to have another look too.
|
|
cperkinsintel
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
cperkinsintel
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
When enqueueing a command and its dependencies, an exception might be thrown. In that case, the command will have a failed EnqueueStatus and not stored in graph_builder DAG. But we need to make sure that dependencies also check so they are not stored there. Otherwise there will be leaks. During the clean up, we don't want to reenqueue it if we know it has failed before. This is broken out from intel#16618
When enqueueing a command and its dependencies, an exception might be thrown. In that case, the command will have a failed EnqueueStatus and not stored in graph_builder DAG. But we need to make sure that dependencies also check so they are not stored there. Otherwise there will be leaks. During the clean up, we don't want to reenqueue it if we know it has failed before.
This is broken out from #16618