[UR][L0 v2] Port USM alloc to adapter v2 #18179
Conversation
unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
Outdated
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Outdated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/queue_immediate_in_order.cpp
Outdated
Show resolved
Hide resolved
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
2 times, most recently
from
8fc9615 to
4ac080d
Compare
unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp
Outdated
Show resolved
Hide resolved
staniewzki
changed the title
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
2 times, most recently
from
a25c7e7 to
803bf54
Compare
|
@staniewzki please also enable llvm tests for v2: https://github.com/intel/llvm/blob/sycl/sycl/test-e2e/AsyncAlloc/lit.local.cfg |
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
from
803bf54 to
ed39eb4
Compare
| @@ -150,6 +150,8 @@ ur_result_t ur_queue_immediate_in_order_t::queueGetNativeHandle( | |||
| ur_result_t ur_queue_immediate_in_order_t::queueFinish() { | |||
| TRACK_SCOPE_LATENCY("ur_queue_immediate_in_order_t::queueFinish"); | |||
|
|
|||
| hContext->getAsyncPool()->cleanupPoolsForQueue(this); | |||
There was a problem hiding this comment.
I'm considering the following scenario: enqueueAlloc->enqueue(Kernel/USMFill)->enqueueFree->queueFinish.
In this case, enqueueFree immediately places the allocation into the freelist along with the associated free event handle. When enqueueAlloc retrieves the allocation from the freelist, it correctly appends a wait on the related free event, so that part works as expected.
But, during queueFinish, the freelist cleanup starts before any synchronization occurs, which may free allocations that are still in use. I think we haven’t seen test failures because the allocations are returned to the UMF pool, which doesn’t release the memory immediately.
Maybe we should move the cleanup after ZE2UR_CALL(zeCommandListHostSynchronize, (commandListLocked->getZeCommandList(), UINT64_MAX));.
There was a problem hiding this comment.
Good catch! :D We need to fix this in L0v1 as well (and cherry-pick this to xmain-rel).
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
from
ed39eb4 to
bcde052
Compare
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
from
bcde052 to
80ab937
Compare
staniewzki
force-pushed
the
async-l0-adapter-v2
branch
from
80ab937 to
a4e8e50
Compare
This PR ports USM alloc enqueue API introduced to L0 adapter in #17112 to L0 adapter v2.