[GPU] Add event completion for optimized out instance

[wilson-seok]

Description of the issue(symptom, root-cause, how it was resolved)

• NMS cpu_impl has wrong output tensor and it causes accuracy drop in validation(MSCOCO mAP 39.3% -> 30.4%) about BMG.
• The issue is gone when get_stream().finish() is called before NMS cpu_impl execution.
• From memory dump of NMS inputs, I found that input0(boxes) is concat which is optimized out from prepare buffer fusing pass.
• Current execute_impl() of ocl_impl checks can_be_optimized() first and return aggregate_events(events). But aggregate_events() doesn't wait for dependent events when QueueTypes=in_order and SyncMethods=none from dGPU use case.
• As the concat user is cpu_impl(nms) and it has need_completion_events=true. So all events should be completed in concat. I added this consideration with wait_for_events().

The code and line that caused this issue (if it is not changed directly)

• openvino/src/plugins/intel_gpu/src/graph/impls/ocl/primitive_base.hpp

  Line 247 in fb89d8b

  return stream.aggregate_events(events, false, instance.is_output());

Problematic graph

• Execution graph about NMS, Input0 is optimized out concat.

Checklist

• Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• 176971

[Copilot]

Pull request overview

This PR fixes a critical synchronization issue in the GPU plugin where optimized-out operations (like concat) were not properly completing their dependent events before CPU implementations (like NMS) accessed their outputs, causing accuracy degradation in MSCOCO validation (39.3% → 30.4% mAP).

Key Changes:

• Added event completion check for optimized-out instances that have CPU consumers requiring synchronization
• Ensures wait_for_events() is called when an optimized instance needs completion events before returning aggregated events

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[hyunback]

LGTM

[e-ddykim]

I think that we can return here instead of calling the next line aggregate_events().

[wilson-seok]

@e-ddykim I added return here. Thanks!

[isanghao]

Did you review existing places to handle needs_completion_event? If I remember correctly, we already handled such case and it is surprising that it is not working as expected..
I found this comment in primitive_inst.cpp and this seems to be relevant.

        // Prepare dependencies events in case of OOO queue, CPU implementation,
        // or optimized_out impl which has CPU users (needs_completion_event() && !is_output() condition)

[wilson-seok]

@isanghao Yes, the dep_events is created from prepare_primitive() as concat is optimized out/needs_completion_event()/!is_output(). And this dep_events is not waited in execute() because BMG has SyncMethods=none. That's why this accuracy happens.