.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "auto_examples_runtime/plot_parallel_for_profiling.py"
.. LINE NUMBERS ARE GIVEN BELOW.

.. only:: html

    .. note::
        :class: sphx-glr-download-link-note

        :ref:`Go to the end <sphx_glr_download_auto_examples_runtime_plot_parallel_for_profiling.py>`
        to download the full example code.

.. rst-class:: sphx-glr-example-title

.. _sphx_glr_auto_examples_runtime_plot_parallel_for_profiling.py:


.. _l-example-parallel-for-profiling:

Inspect ParallelFor profiling from Python
=========================================

Parallel-region profiling is disabled by default. This example enables it with
a bounded collector, runs an ``Abs`` model, and inspects an immutable report.
Passing ``hardware_counters=True`` explicitly requests the Linux grouped
``perf_event_open`` backend. Linux may require ``CAP_PERFMON`` or a permissive
``kernel.perf_event_paranoid`` setting. The backend rejects rather than scales
multiplexed groups, so derived metrics exist only when ``time_running`` equals
``time_enabled``. It also marks multi-thread regions unsuitable because a
thread-bound perf group cannot represent their aggregate work. Unsupported
platforms and denied counters retain portable timing; ``None`` is distinct from
a valid zero.

.. GENERATED FROM PYTHON SOURCE LINES 18-73




.. rst-class:: sphx-glr-script-out

 .. code-block:: none

    events=1, dropped=0
    /home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/kernels/kernels/math/kernel_abs.cc:108: participants requested/admitted/observed=1/1/1, wall=23263 ns, cpu_utilization=0.918, counters=permission_denied, time_running/enabled=unavailable, ipc=unavailable, llc_miss_rate=unavailable






|

.. code-block:: Python


    from __future__ import annotations

    import numpy

    from onnx_light.onnx import TensorProto
    from onnx_light.onnx_lib import parser
    from onnx_light.onnx_py import _onnxpykernels

    runtime = _onnxpykernels.runtime

    model = parser.parse_model(
        '<ir_version: 10, opset_import: ["" : 18]>\n'
        "agraph (float[N] x) => (float[N] y) {\n"
        "  y = Abs(x)\n"
        "}\n"
    )
    values = -numpy.ones(100_000, dtype=numpy.float32)
    tensor_proto = TensorProto()
    tensor_proto.name = "x"
    tensor_proto.dims.append(values.size)
    tensor_proto.data_type = int(TensorProto.FLOAT)
    tensor_proto.raw_data = values.tobytes()

    collector = runtime.ParallelRegionCollector(capacity=8, hardware_counters=True)
    options = runtime.RuntimeSessionOptions(
        parameters=runtime.RuntimeParameters(1), parallel_region_collector=collector
    )
    context = runtime.RuntimeContext(runtime.KernelContext(runtime.default_opset(18)))
    context.set("x", runtime.tensor_from_proto(tensor_proto))
    session = runtime.RuntimeSession(model, options)
    session.run(context)

    report = session.parallel_region_report()
    print(f"events={len(report.events)}, dropped={report.dropped_events}")
    for event in report.events:
        location = f"{event.file_name}:{event.line}"
        wall_time = "unavailable" if event.wall_time_ns is None else f"{event.wall_time_ns} ns"
        utilization = (
            "unavailable" if event.cpu_utilization is None else f"{event.cpu_utilization:.3f}"
        )
        ipc = "unavailable" if event.ipc is None else f"{event.ipc:.3f}"
        llc = "unavailable" if event.llc_miss_rate is None else f"{event.llc_miss_rate:.3%}"
        counter_time = (
            "unavailable"
            if event.counter_time_enabled is None
            else f"{event.counter_time_running}/{event.counter_time_enabled}"
        )
        print(
            f"{location}: participants requested/admitted/observed="
            f"{event.requested_threads}/{event.admitted_threads}/{event.observed_threads}, "
            f"wall={wall_time}, cpu_utilization={utilization}, "
            f"counters={event.counter_status}, time_running/enabled={counter_time}, "
            f"ipc={ipc}, llc_miss_rate={llc}"
        )


.. rst-class:: sphx-glr-timing

   **Total running time of the script:** (0 minutes 0.002 seconds)


.. _sphx_glr_download_auto_examples_runtime_plot_parallel_for_profiling.py:

.. only:: html

  .. container:: sphx-glr-footer sphx-glr-footer-example

    .. container:: sphx-glr-download sphx-glr-download-jupyter

      :download:`Download Jupyter notebook: plot_parallel_for_profiling.ipynb <plot_parallel_for_profiling.ipynb>`

    .. container:: sphx-glr-download sphx-glr-download-python

      :download:`Download Python source code: plot_parallel_for_profiling.py <plot_parallel_for_profiling.py>`

    .. container:: sphx-glr-download sphx-glr-download-zip

      :download:`Download zipped: plot_parallel_for_profiling.zip <plot_parallel_for_profiling.zip>`


.. include:: plot_parallel_for_profiling.recommendations


.. only:: html

 .. rst-class:: sphx-glr-signature

    `Gallery generated by Sphinx-Gallery <https://sphinx-gallery.github.io>`_