intel
[SYCL][Graph] Leak in host task in order dependency test on Windows Gen12
Describe the bug
First identified in: https://github.com/intel/llvm/actions/runs/19515665783/job/55869141453?pr=20690.
The newly added test host_task_in_order_dependency.cpp leaks L0 objects. This issue has only shown up on gen12 in testing and the test is being marked as XFAIL.
Note that I reproduced the leak with icx 2025.2.0, so the issue predates the PR which added the test.
To reproduce
Compile the test host_task_in_order_dependency.cpp or run with llvm-lit.
Manual run on Windows Gen12:
C:\...>.\in_order_host_task_dependency.exe
Check balance of create/destroy calls
----------------------------------------------------------
zeContextCreate = 1 \---> zeContextDestroy = 0 ---> LEAK = 1
zeCommandQueueCreate = 1 \---> zeCommandQueueDestroy = 1
zeModuleCreate = 1 \---> zeModuleDestroy = 0 ---> LEAK = 1
zeKernelCreate = 1 \---> zeKernelDestroy = 0 ---> LEAK = 1
zeEventPoolCreate = 1 \---> zeEventPoolDestroy = 0 ---> LEAK = 1
zeCommandListCreateImmediate = 1 |
zeCommandListCreate = 1 \---> zeCommandListDestroy = 0 ---> LEAK = 2
zeEventCreate = 1 \---> zeEventDestroy = 1
zeFenceCreate = 1 \---> zeFenceDestroy = 1
zeImageCreate = 0 |
zeImageViewCreateExt = 0 \---> zeImageDestroy = 0
zeSamplerCreate = 0 \---> zeSamplerDestroy = 0
zeMemAllocDevice = 2 |
zeMemAllocHost = 3 |
zeMemAllocShared = 4 \---> zeMemFree = 8 ---> LEAK = 1
Passing run linux Gen12:
$ UR_L0_LEAKS_DEBUG=1 ./in_order_host_dependency
Check balance of create/destroy calls
----------------------------------------------------------
zeContextCreate = 1 \---> zeContextDestroy = 1
zeCommandQueueCreate = 1 \---> zeCommandQueueDestroy = 1
zeModuleCreate = 1 \---> zeModuleDestroy = 1
zeKernelCreate = 1 \---> zeKernelDestroy = 1
zeEventPoolCreate = 1 \---> zeEventPoolDestroy = 1
zeCommandListCreateImmediate = 1 |
zeCommandListCreate = 1 \---> zeCommandListDestroy = 2
zeEventCreate = 1 \---> zeEventDestroy = 1
zeFenceCreate = 1 \---> zeFenceDestroy = 1
zeImageCreate = 0 |
zeImageViewCreateExt = 0 \---> zeImageDestroy = 0
zeSamplerCreate = 0 \---> zeSamplerDestroy = 0
zeMemAllocDevice = 2 |
zeMemAllocHost = 3 |
zeMemAllocShared = 4 \---> zeMemFree = 9
Environment
- Windows
- Intel GPU Gen 12
- Produced the issue with as early as icpx 2025.2
- Driver version 1.6.35096. Level-zero backend should be used to check for leaks with UR_L0_LEAKS_DEBUG=1
Additional context
No response
Leak was fixed after sync with main sycl branch and xfail was removed. 2025.2 failure is also expected behavior without setting SYCL_ENABLE_DEFAULT_CONTEXTS=0. Once enabled, this test passes.
Edit: see below.
After passing in several CI runs, newer run shows the failure reappearing suggesting the test is flaky:
# | 1: Check balance of create/destroy calls
# | 2: ----------------------------------------------------------
# | 3: zeContextCreate = 1 \---> zeContextDestroy = 0 ---> LEAK = 1
# | not:imp1 !~~~ error: no match expected
# | 4: zeCommandQueueCreate = 1 \---> zeCommandQueueDestroy = 0 ---> LEAK = 1
# | 5: zeModuleCreate = 1 \---> zeModuleDestroy = 1
# | 6: zeKernelCreate = 1 \---> zeKernelDestroy = 1
# | 7: zeEventPoolCreate = 2 \---> zeEventPoolDestroy = 0 ---> LEAK = 2
# | 8: zeCommandListCreateImmediate = 1 |
The reported leaks also seem to align with https://github.com/intel/llvm/issues/14473. Propose moving to unsupported for now: https://github.com/intel/llvm/pull/20792