[SYCL] Remove OwnZeMemHandle from USMAllocator

[igchor]

OwnZeMemHandle is used to track whether an allocation should be freed by the runtime or not.

There is no point in passing this flag to USMAllocator and then to USMFreeHelper since the allocation cannot come from the USMAllocator (it can only be from zeMemAlloc*).

I've thought about moving the memory allocation code to UR first, but there are a lot of dependencies: we would have to move the implementation of context, device, event, and queue at least. It seemed that doing this small change first made more sense.

[igchor]

I'm not entirely sure if doing RefCount.decrementAndTest for a MemAlloc Context which does not own the memory makes sense here: https://github.com/intel/llvm/pull/7853/files#diff-15dd1eb076d2164bd9e87d9057f05f652a716498e8cdf5975e564c65309a0985R8403) or if it would be better to just skip calling USMFreeHelper at all but for now, I just left it as it is for consistency.

[igchor]

@smaslov-intel could you please take a look at this PR?

[smaslov-intel]

LGTM

[jandres742]

+1

[igchor]

@intel/llvm-gatekeepers I've just rebased the PR - I think it's ready for merge.

[igchor]

@intel/llvm-gatekeepers @againull @smaslov-intel is there something pending to merge this?

[bader]

@igchor, review from @intel/llvm-reviewers-runtime is missing.

[aelovikov-intel]

@igchor, review from https://github.com/orgs/intel/teams/llvm-reviewers-runtime is missing.

But why is it needed? Should we change codeowner for the Unified Runtime directories?

[igchor]

@igchor, review from https://github.com/orgs/intel/teams/llvm-reviewers-runtime is missing.

But why is it needed? Should we change codeowner for the Unified Runtime directories?

I think the idea was that the common part of Unified Runtime (including usm_allocator) would be eventually owned by Codeplay and the Intel UR team (people from https://github.com/orgs/oneapi-src/teams/unified-runtime-maintain)

[smaslov-intel]

@igchor, review from https://github.com/orgs/intel/teams/llvm-reviewers-runtime is missing.

But why is it needed? Should we change codeowner for the Unified Runtime directories?

I think the idea was that the common part of Unified Runtime (including usm_allocator) would be eventually owned by Codeplay and the Intel UR team (people from https://github.com/orgs/oneapi-src/teams/unified-runtime-maintain)

I suggest that we keep current ownership until https://github.com/intel/llvm/tree/sycl/sycl/plugins/unified_runtime/ur is moved into a separate repo, where the new ownership would be more natural. ETA EoQ1'23

[igchor]

@igchor, review from https://github.com/orgs/intel/teams/llvm-reviewers-runtime is missing.

But why is it needed? Should we change codeowner for the Unified Runtime directories?

I think the idea was that the common part of Unified Runtime (including usm_allocator) would be eventually owned by Codeplay and the Intel UR team (people from https://github.com/orgs/oneapi-src/teams/unified-runtime-maintain)

I suggest that we keep current ownership until https://github.com/intel/llvm/tree/sycl/sycl/plugins/unified_runtime/ur is moved into a separate repo, where the new ownership would be more natural. ETA EoQ1'23

Do you think it would make sense to move just the usm_allocator to the UR repo before EoQ1? We already moved the core UMA abstractions there. This would minimize changes required in the llvm repo - (we would only need to integrate L0 adapter with UMA)

[smaslov-intel]

@igchor, review from https://github.com/orgs/intel/teams/llvm-reviewers-runtime is missing.

But why is it needed? Should we change codeowner for the Unified Runtime directories?

I think the idea was that the common part of Unified Runtime (including usm_allocator) would be eventually owned by Codeplay and the Intel UR team (people from https://github.com/orgs/oneapi-src/teams/unified-runtime-maintain)

I suggest that we keep current ownership until https://github.com/intel/llvm/tree/sycl/sycl/plugins/unified_runtime/ur is moved into a separate repo, where the new ownership would be more natural. ETA EoQ1'23

Do you think it would make sense to move just the usm_allocator to the UR repo before EoQ1? We already moved the core UMA abstractions there. This would minimize changes required in the llvm repo - (we would only need to integrate L0 adapter with UMA)

Yeah, let's do that.

[bader]

@igchor, could you resolve merge conflicts, please?

[igchor]

@igchor, could you resolve merge conflicts, please?

Sure, done.

[bader]

Failed Tests (1):
SYCL :: XPTI/kernel/basic.cpp

This issue was caused by 18899cc, which was reverted in 01511a3.