[onert] Simplify NPU execution with blocking runNPU_model

[batcheu]

Replace multi-step NPU request workflow with single blocking API call.
This change removes the complex threading mechanism used to handle submitNPU_request timeout and thread-safety issues.

ONE-DCO-1.0-Signed-off-by: Jonghwa Lee jonghwa3.lee@samsung.com

[batcheu]

The workaround is no longer required for DevContext because a patch that solves the timeout issue is applied to the trix-engine.

• https://review.tizen.org/gerrit/gitweb?p=platform/adaptation/npu/trix-engine.git;a=commit;h=28a631023327c7f727f853c9e0070273a5976488

[batcheu]

It can optimize model inference time by removing this workaround (~200 us on Cortex-A72 @1.8GHz)

onert_run execution time comparision

• Single Batch

  • Before : 2.569 ms

    raw data

    $ onert_run -r 10 model.circle
    ===================================
    MODEL_LOAD   takes 2.990 ms
    PREPARE      takes 16.813 ms
    EXECUTE      takes 2.569 ms
    - MEAN     :  2.569 ms
    - MAX      :  3.555 ms
    - MIN      :  2.395 ms
    - GEOMEAN  :  2.551 ms
    ===================================
    

  • After : 2.365 ms

    raw data

    $ onert_run -r 10 model.circle
    ===================================
    MODEL_LOAD   takes 2.982 ms
    PREPARE      takes 16.882 ms
    EXECUTE      takes 2.365 ms
    - MEAN     :  2.365 ms
    - MAX      :  3.302 ms
    - MIN      :  2.179 ms
    - GEOMEAN  :  2.347 ms
    ===================================
    

• N-Batch (N : 4)

  • Before : 5.762 ms

    raw data

    $ onert_run -r 10 model.batch_4.circle
    ===================================
    MODEL_LOAD   takes 3.012 ms
    PREPARE      takes 17.089 ms
    EXECUTE      takes 5.762 ms
    - MEAN     :  5.762 ms
    - MAX      :  7.107 ms
    - MIN      :  5.536 ms
    - GEOMEAN  :  5.746 ms
    ===================================
    

  • After : 5.432 ms

    raw data

    $ onert_run -r 10 model.batch_4.circle
    ===================================
    MODEL_LOAD   takes 2.997 ms
    PREPARE      takes 16.888 ms
    EXECUTE      takes 5.432 ms
    - MEAN     :  5.432 ms
    - MAX      :  6.688 ms
    - MIN      :  5.209 ms
    - GEOMEAN  :  5.418 ms
    ===================================
    

[ragmani]

Sorry for the late reply.
I'd like to confirm whether the hanging issue has been resolved. Do these changes ensure that threads no longer become unresponsive? In other words, is there no longer any need for the previous workaround of allowing hanging threads?

[batcheu]

Sorry for the late reply. I'd like to confirm whether the hanging issue has been resolved. Do these changes ensure that threads no longer become unresponsive? In other words, is there no longer any need for the previous workaround of allowing hanging threads?

Yes, there's no hanging issue with the latest version of trix-engine.

[ragmani]

LGTM

[hseok-oh]

LGTM