Merge branch 'master' into bob/async

doc/sphinx/conf.py

-
-
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 #
@@ -73,13 +71,11 @@ author = 'Intel Corporation'
 # built documents.
 #
 # The short X.Y version.
-version = '0.22'
+version = '0.23'
 
 # The Documentation full version, including alpha/beta/rc tags. Some features
-# available in the latest code will not necessarily be documented first.
-# rc syntax may be tagged; this documentation supports various rc-naming conventions 
-
-release = '0.22.0'
+# available in the latest code will not necessarily be documented first
+release = '0.23.0'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.
@@ -107,38 +103,18 @@ todo_include_todos = True
 
 # -- Options for HTML output ----------------------------------------------
 
-html_title = 'nGraph Compiler stack Documentation'
+html_title = "Documentation for the nGraph Library and Compiler stack"
 
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
 html_theme = 'ngraph_theme'
 
+# Theme options are theme-specific and customize the look and feel of a theme
+# further.  For a list of options available for each theme, see the
+# documentation.
 # html_theme_options = {}
-html_theme_path = ["../"]
-
-if tags.has('release'):
-    is_release = True
-    docs_title = 'Docs / %s' %(version) %(release)
-else:
-    is_release = False
-    docs_title = 'Docs / Latest'
-
-    # borrow this from the zephyr docs theme
-    html_context = {
-        # 'show_license': html_show_license, we have custom footers to attribute 
-        # RTD, WTD, and Sphinx contributors; so we do not enable this
-        'docs_title': docs_title,
-        'is_release': is_release,
-        'theme_logo_only': False,
-        'current_version': version,
-        'versions': ( ("latest", "../"),
-                     ("0.20.0", "/0.20.0/"), #not yet sure how we'll do this
-                     ("0.19.0", "/0.19.0/"), 
-                     ("0.18.0", "/0.18.0/"),
-                     ("0.17.0", "/0.17.0/"),
-                     ("0.16.0", "/0.16.0/"),
-                    )
-    }
-
-html_logo = '../ngraph_theme/static/favicon.ico'
+html_logo = '../ngraph_theme/static/logo.png'
 
 # The name of an image file (within the static path) to use as favicon of the
 # docs.  This file should be a Windows icon file (.ico) being 16x16 or 32x32
@@ -151,6 +127,7 @@ html_favicon = '../ngraph_theme/static/favicon.ico'
 html_static_path = ['../ngraph_theme/static']
 
 # Add any paths that contain custom themes here, relative to this directory.
+html_theme_path = ["../"]
 
 # Custom sidebar templates, must be a dictionary that maps document names
 # to template names.
@@ -165,26 +142,6 @@ html_sidebars = {
 }
 
 
-# Custom added feature to allow redirecting old URLs
-#
-# list of tuples (old_url, new_url) for pages to redirect
-# (URLs should be relative to document root, only)
-html_redirect_pages = [
-        ('backend-support', 'backends/index'),
-        ('core/core', 'core/overview.rst'),
-        ('core/fusion', 'core/fusion/index'),
-        ('frameworks/mxnet', 'frameworks/mxnet_intg.rst'),
-        ('frameworks/onnx', 'frameworks/onnx_intg.rst'),
-        ('frameworks/tensorflow', 'frameworks/tensorflow_connect.rst'),
-        ('frameworks/paddle', 'frameworks/paddle_integ.rst'),
-        ('inspection/inspection', 'inspection/index'),
-        ('releases/release-notes', 'releases/index'),
-        # ('getting_started/getting_starting', 'getting_started/index'),
-        #  mv to framework-specific helper directory
-        ('project/project', 'project/index'),
-        ('python_api/', 'python_api/index'),
-]
-
 # -- Options for HTMLHelp output ------------------------------------------
 
 # Output file base name for HTML help builder.
@@ -195,11 +152,11 @@ htmlhelp_basename = 'IntelnGraphlibrarydoc'
 latex_elements = {
     # The paper size ('letterpaper' or 'a4paper').
     #
-    # 'papersize': 'letterpaper',
+    'papersize': 'letterpaper',
 
     # The font size ('10pt', '11pt' or '12pt').
     #
-    # 'pointsize': '10pt',
+    'pointsize': '10pt',
 
     # Additional stuff for the LaTeX preamble.
     #
@@ -214,11 +171,10 @@ latex_elements = {
 # (source start file, target name, title,
 #  author, documentclass [howto, manual, or own class]).
 latex_documents = [
-    (master_doc, 'nGraphCompilerStack.tex', 'nGraph Compiler Stack Documentation',
-     'Intel Corporation', 'manual'),
+    (master_doc, 'nGraphCompilerStack.tex', u'nGraph Compiler Stack Documentation',
+     u'Intel Corporation', 'manual'),
 ]
 
-
 # -- Options for manual page output ---------------------------------------
 
 # One entry per manual page. List of tuples
@@ -244,22 +200,7 @@ breathe_projects = {
 }
 
 rst_epilog = u"""
-.. |codename| replace:: Intel nGraph
-.. |project| replace:: Intel nGraph Library
-.. |InG| replace:: Intel® nGraph
-.. |copy|   unicode:: U+000A9 .. COPYRIGHT SIGN
-   :ltrim:
-.. |deg|    unicode:: U+000B0 .. DEGREE SIGN
-   :ltrim:
-.. |plusminus|  unicode:: U+000B1 .. PLUS-MINUS SIGN
-   :rtrim:
-.. |micro|  unicode:: U+000B5 .. MICRO SIGN
-   :rtrim:
-.. |trade|  unicode:: U+02122 .. TRADEMARK SIGN
-   :ltrim:
-.. |reg|    unicode:: U+000AE .. REGISTERED TRADEMARK SIGN
-   :ltrim:   
-
+.. include:: /replacements.txt
 """
 
 # -- autodoc Extension configuration --------------------------------------

doc/sphinx/ngraph_theme/ngversions.html

@@ -6,16 +6,14 @@
     </span>
     <div class="rst-other-versions">
       <dl>
-        <dt>{{ _('Previous Versions') }}</dt>
+        <dt>{{ _('Recent Versions') }}</dt>
         <dd><!-- Until our https://docs.ngraph.ai/ publishing is set up, we link to GitHub -->  
           <ul>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.20.0">0.20.0-rc.0</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.19.0-rc.2">0.19.0-rc.2</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.18.1">0.18.1</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.17.0-rc.1">0.17.0-rc.1</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.3">0.16.0-rc.3</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.2">0.16.0-rc.2</a></li>
-           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.1">0.16.0-rc.1</a></li>
+           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.23.0">0.23.0</a></li>
+           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.22.0">0.22.0</a></li>
+           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.21.0">0.21.0</a></li>
+           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.20.0">0.20.0</a></li>
+           <li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.19.0">0.19.0</a></li>
          </ul></dd>
       </dl>
       <dl>
@@ -28,4 +26,4 @@
           </dd>
       </dl>
     </div>
-</div>
+</div>
\ No newline at end of file

doc/sphinx/source/backends/backend-api/index.rst

@@ -7,4 +7,3 @@ Backend
 .. doxygenclass:: ngraph::runtime::Backend
    :project: ngraph
    :members:
-

doc/sphinx/source/backends/cpp-api.rst

@@ -7,9 +7,10 @@ Backend APIs
    :maxdepth: 1
    
    backend-api/index
-   executable-api/index
-   hosttensor-api/index
+   dynamicbackend-api/index
    plaidml-ng-api/index
+   executable-api/index
+   
 
 As of version ``0.15``, there is a new backend API to work with functions that 
 can be compiled as a runtime ``Executable``. Where previously ``Backend`` used a 

doc/sphinx/source/backends/hosttensor-api/index.rst → doc/sphinx/source/backends/dynamicbackend-api/index.rst

 .. backends/hosttensor-api/index.rst:
 
 
-HostTensor
-==========
+DynamicBackend
+==============
 
-.. doxygenclass:: ngraph::runtime::HostTensor
+.. doxygenclass:: ngraph::runtime::dynamic::DynamicBackend
    :project: ngraph
    :members:

doc/sphinx/source/backends/executable-api/index.rst

@@ -7,7 +7,6 @@ Executable
 The ``compile`` function on an ``Executable`` has more direct methods to 
 actions such as ``validate``, ``call``, ``get_performance_data``, and so on. 
 
-
 .. doxygenclass:: ngraph::runtime::Executable
    :project: ngraph
    :members: 

doc/sphinx/source/backends/index.rst

@@ -6,6 +6,7 @@ Developer Resources for Backends
 
 * :ref:`what_is_backend`
 * :ref:`how_to_use`
+* :ref:`miscellaneous_resources`
 
 
 .. _what_is_backend:
@@ -60,3 +61,57 @@ interface; each backend implements the following five functions:
   for later execution.
 * And, finally, the ``call()`` method is used to invoke an nGraph function 
   against a particular set of tensors.
+
+
+.. _miscellaneous_resources: 
+
+Miscellaneous resources
+=======================
+
+Additional resources for device or framework-specific configurations:
+
+OpenCL
+------
+
+OpenCL is needed for the :doc:`plaidml-ng-api/index`; this is not needed if 
+you have only a CPU backend.  
+
+#. Install the latest Linux driver for your system. You can find a list 
+   of drivers at https://software.intel.com/en-us/articles/opencl-drivers;
+   You may need to install `OpenCL SDK`_ in case of an ``libOpenCL.so`` absence.
+
+#. Any user added to "video" group: 
+
+   .. code-block:: console 
+
+      sudo usermod –a –G video <user_id>
+
+   may, for example, be able to find details at the ``/sys/module/[system]/parameters/`` location. 
+
+
+nGraph Bridge from TensorFlow\*
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+When specified as the generic backend -- either manually or automatically 
+from a framework --  ``NGRAPH`` defaults to CPU, and it also allows for 
+additional device configuration or selection. 
+
+Because nGraph can select backends, specifying the ``INTELGPU`` 
+backend as a runtime environment variable also works if one is 
+present in your system: 
+
+:envvar:`NGRAPH_TF_BACKEND="INTELGPU"`
+
+An `axpy.py example`_ is optionally available to test; outputs will vary 
+depending on the parameters specified. 
+
+.. code-block:: console
+
+   NGRAPH_TF_BACKEND="INTELGPU" python3 axpy.py
+
+* ``NGRAPH_INTELGPU_DUMP_FUNCTION`` -- dumps nGraph’s functions 
+  in dot format.
+
+
+.. _axpy.py example: https://github.com/tensorflow/ngraph-bridge/blob/master/examples/axpy.py
+.. _OpenCL SDK: https://software.intel.com/en-us/opencl-sdk

doc/sphinx/source/conf.py

@@ -73,11 +73,11 @@ author = 'Intel Corporation'
 # built documents.
 #
 # The short X.Y version.
-version = '0.21'
+version = '0.22'
 
 # The Documentation full version, including alpha/beta/rc tags. Some features
 # available in the latest code will not necessarily be documented first
-release = '0.21.0'
+release = '0.22.0'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.

doc/sphinx/source/frameworks/generic-configs.rst

@@ -5,8 +5,8 @@ Integrating new frameworks
 
 This section details some of the *configuration options* and some of the 
 *environment variables* that can be used to tune for optimal performance when 
-your system already has a version of nGraph installed with one of our supported
-backends. 
+your system already has a version of nGraph installed with one or more of our 
+supported :doc:`../backends/index`.
 
 Regardless of the framework, after the :doc:`../buildlb` step, a good place 
 to start usually involves making the libraries available to the framework. On 
@@ -19,15 +19,59 @@ something like:
    export LD_LIBRARY_PATH=path/to/ngraph_dist/lib/
 
 
-Find or display nGraph Version
--------------------------------
-
+Find or display version
+=======================
 
+If you're working with the :doc:`../python_api/index`, the following command 
+may be useful:
 
 .. code-block:: console
 
    python3 -c "import ngraph as ng; print('nGraph version: ',ng.__version__)";
 
+To manually build a newer version than is available from the latest `PyPI`_
+(:abbr:`Python Package Index (PyPI)`), see our nGraph Python API `BUILDING.md`_ 
+documentation.
+
+
+Activate logtrace-related environment variables
+===============================================
+
+Another configuration option is to activate ``NGRAPH_CPU_DEBUG_TRACER``,
+a runtime environment variable that supports extra logging and debug detail. 
+
+This is a useful tool for data scientists interested in outputs from logtrace 
+files that can, for example, help in tracking down model convergences. It can 
+also help engineers who might want to add their new ``Backend`` to an existing 
+framework to compare intermediate tensors/values to references from a CPU 
+backend.
+
+To activate this tool, set the ``env`` var ``NGRAPH_CPU_DEBUG_TRACER=1``.
+It will dump ``trace_meta.log`` and ``trace_bin_data.log``. The names of the 
+logfiles can be customized.
+
+To specify the names of logs with those flags:
+
+:: 
+
+  NGRAPH_TRACER_LOG = "meta.log"
+  NGRAPH_BIN_TRACER_LOG = "bin.log"
+
+The meta_log contains::
+ 
+  kernel_name, serial_number_of_op, tensor_id, symbol_of_in_out, num_elements, shape, binary_data_offset, mean_of_tensor, variance_of_tensor
+
+A line example from a unit-test might look like::
+
+  K=Add S=0 TID=0_0 >> size=4 Shape{2, 2} bin_data_offset=8 mean=1.5 var=1.25
+
+The binary_log line contains::
+
+  tensor_id, binary data (tensor data)
+
+A reference for the implementation of parsing these logfiles can also be found 
+in the unit test for this feature.
+
 
 FMV
 ---
@@ -37,7 +81,7 @@ number of generic ways to patch or bring architecture-based optimizations to
 the :abbr:`Operating System (OS)` that is handling your ML environment. See 
 the `GCC wiki for details`_.
 
-If your nGraph build is a Neural Network configured on Clear Linux* OS 
+If your nGraph build is a Neural Network configured on Clear Linux\* OS 
 for Intel® Architecture, and it includes at least one older CPU, the 
 `following article may be helpful`_.
 
@@ -63,11 +107,14 @@ For CPU (and most cuDNN) backends, the preferred layout is currently ``NCHW``.
 Intel® Math Kernel Library for Deep Neural Networks 
 ---------------------------------------------------
 
--The following `KMP options`_ were originally optimized for models using the 
+.. important:: Intel® MKL-DNN is automatically enabled as part of an
+   nGraph default :doc:`build <../buildlb>`; you do *not* need to add it 
+   separately or as an additional component to be able to use these 
+   configuration settings.
+
+The following `KMP`_ options were originally optimized for models using the 
 Intel® `MKL-DNN`_ to train models with the ``NCHW`` data layout; however, other 
-configurations can be explored. MKL-DNN is automatically enabled as part of an 
-nGraph compilation; you do *not* need to add MKL-DNN separately or as an 
-additional component to be able to use these configuration settings.   
+configurations can be explored.    
 
 * ``KMP_BLOCKTIME`` Sets the time, in milliseconds, that a thread should wait 
   after completing the execution of a parallel region, before sleeping.
@@ -128,10 +175,11 @@ Convolution shapes
 ``OMP_NUM_THREADS``
 ^^^^^^^^^^^^^^^^^^^
 
-The best resource for this configuration option is the `gnu.org site`_ 
-``OMP_NUM_THREADS`` defaults to the number of logical cores. To check the 
-number of cores on your system, you can run the following on the command-line to 
-see the details of your CPU: 
+The best resource for this configuration option is the Intel® OpenMP\* docs 
+at the following link: `Intel OpenMP documentation`_. ``OMP_NUM_THREADS`` 
+defaults to the number of logical cores. To check the number of cores on your 
+system, you can run the following on the command-line to see the details 
+of your CPU:
 
 .. code-block:: console
 
@@ -171,11 +219,12 @@ achieve the best performance for NN workloads on CPU platforms. The nGraph
 Compiler stack runs on transformers handled by Intel® Architecture (IA), and 
 thus can make more efficient use of the underlying hardware.
 
-
-.. _KMP options: https://software.intel.com/en-us/cpp-compiler-developer-guide-and-reference-controlling-thread-allocation
-.. KMP options: https://software.intel.com/en-us/node/522691
+.. _PyPI: https://pypi.org/project/ngraph-core
+.. _KMP: https://software.intel.com/en-us/node/522691
 .. _MKL-DNN: https://github.com/intel/mkl-dnn
-.. _gnu.org site: https://gcc.gnu.org/onlinedocs/libgomp/Environment-Variables.html
+.. _Intel OpenMP documentation: https://www.openmprtl.org/documentation
 .. _Movidius: https://www.movidius.com/
+.. _BUILDING.md: https://github.com/NervanaSystems/ngraph/blob/master/python/BUILDING.md
 .. _GCC wiki for details: https://gcc.gnu.org/wiki/FunctionMultiVersioning
 .. _following article may be helpful: https://clearlinux.org/documentation/clear-linux/tutorials/fmv
+

doc/sphinx/source/frameworks/mxnet_integ.rst

 .. frameworks/mxnet_integ.rst:
 
 MXNet\* bridge
-===============
+==============
+
+.. deprecated:: 0.21.0
+
 
 * See the nGraph-MXNet `Integration Guide`_ on the nGraph-MXNet repo.
 

doc/sphinx/source/frameworks/onnx_integ.rst

@@ -10,7 +10,7 @@ nGraph's internal representation and converted to ``Function`` objects, which
 can be compiled and executed on one of nGraph's backends.
 
 You can use nGraph's Python API to run an ONNX model and nGraph can be used 
-as an ONNX backend using the add-on package `nGraph-ONNX <ngraph_onnx>`_.
+as an ONNX backend using the add-on package `nGraph ONNX`_.
 
 
 .. note:: In order to support ONNX, nGraph must be built with the 
@@ -33,8 +33,7 @@ for nGraph, ONNX and NumPy:
 Importing an ONNX model
 -----------------------
 
-You can download models from the `ONNX Model Zoo <onnx_model_zoo_>`_.
-For example ResNet-50:
+You can download models from the `ONNX Model Zoo`_. For example, ResNet-50:
 
 ::
 
@@ -92,9 +91,9 @@ data:
 
 
 Find more information about nGraph and ONNX in the 
-`nGraph-ONNX <ngraph_onnx>`_ GitHub repository.
+`nGraph ONNX`_ GitHub repository.
 
 
-.. _ngraph_onnx: https://github.com/NervanaSystems/ngraph-onnx/
-.. _ngraph_onnx_building: https://github.com/NervanaSystems/ngraph-onnx/blob/master/BUILDING.md
-.. _onnx_model_zoo: https://github.com/onnx/models
+.. _ngraph ONNX: https://github.com/NervanaSystems/ngraph-onnx
+.. _ngraph ONNX building: https://github.com/NervanaSystems/ngraph-onnx/blob/master/BUILDING.md
+.. _ONNX model zoo: https://github.com/onnx/models

doc/sphinx/source/index.rst

-.. ---------------------------------------------------------------------------
 .. Copyright 2018-2019 Intel Corporation
 .. Licensed under the Apache License, Version 2.0 (the "License");
 .. you may not use this file except in compliance with the License.

doc/sphinx/source/inspection/index.rst

@@ -6,11 +6,11 @@ Visualization Tools
 nGraph provides serialization and deserialization facilities, along with the 
 ability to create image formats or a PDF. 
 
-When visualization is enabled, a ``dot`` file gets generated, along with a 
-``png``. The default can be adjusted by setting the 
-``NGRAPH_VISUALIZE_TREE_OUTPUT_FORMAT`` flag to another format, like PDF. 
+When visualization is enabled, ``svg`` files for your graph get generated. The 
+default can be adjusted by setting the ``NGRAPH_VISUALIZE_TRACING_FORMAT`` 
+flag to another format, like PNG or PDF. 
 
-.. note:: Large graphs are usually not legible with formats like PDF.  
+.. note:: Large graphs are usually not legible with formats like PDF.
 
 Large graphs may require additional work to get into a human-readable format. 
 On the back end, very long edges will need to be cut to make (for example) a 

doc/sphinx/source/project/doc-contributor-README.rst

@@ -18,9 +18,36 @@
 Contributing to documentation
 =============================
 
-.. important:: Read this for changes affecting **anything** in ``ngraph/doc``
+.. note:: Tips for contributors who are new to the highly-dynamic 
+   environment of documentation in AI software:
 
-For updates to the nGraph Library ``/doc`` repo, please submit a PR with 
+   * A good place to start is "document something you figured out how to 
+     get working". Content changes and additions should be targeted at 
+     something more specific than "developers". If you don't understand 
+     how varied and wide the audience is, you'll inadvertently break or block 
+     things.
+
+   * There are experts who work on all parts of the stack; try asking how 
+     documentation changes ought to be made in their respective sections.
+
+   * Start with something small. It is okay to add a "patch" to fix a typo 
+     or suggest a word change; larger changes to files or structure require 
+     research and testing first, as well as some logic for why you think 
+     something needs changed.
+
+   * Most documentation should wrap at about ``80``. We do our best to help 
+     authors source-link and maintain their own code and contributions; 
+     overwriting something already documented doesn't always improve it.
+
+   * Be careful editing files with links already present in them; deleting 
+     links to papers, citations, or sources is discouraged.
+
+   * Please do not submit Jupyter* notebook code to the nGraph Library
+     or core repos; best practice is to maintain any project-specific 
+     examples, tests, or walk-throughs in a separate repository and to link 
+     back to the stable ``op`` or Ops that you use in your project.
+
+For updates within the nGraph Library ``/doc`` repo, please submit a PR with 
 any changes or ideas you'd like integrated. This helps us maintain trackability 
 with respect to changes made, additions, deletions, and feature requests.
 
@@ -36,7 +63,6 @@ it for a specific use case. Add a note on our wiki to show us what you
 did; new and novel applications may have their projects highlighted on an 
 upcoming `ngraph.ai`_ release.   
 
-
 .. note:: Please do not submit Jupyter* notebook code to the nGraph Library 
    or core repos; best practice is to maintain any project-specific examples, 
    tests, or walk-throughs in a separate repository.
@@ -45,9 +71,11 @@ upcoming `ngraph.ai`_ release.
 Documenting source code examples 
 --------------------------------
 
-When **verbosely** documenting functionality of specific sections of code -- whether 
-they are entire code blocks within a file, or code strings that are **outside** 
-the nGraph Library's `documentation repo`_, here is an example of best practice: 
+When **verbosely** documenting functionality of specific sections of code -- 
+whether they are entire code blocks within a file, or code strings that are 
+**outside** the nGraph Library's `documentation repo`_, here is an example 
+of best practice: 
+
 
 Say a file has some interesting functionality that could benefit from more 
 explanation about one or more of the pieces in context. To keep the "in context" 
@@ -73,8 +101,7 @@ the folder where the ``Makefile`` is that generates the documentation you're
 writing. 
 
 See the **note** at the bottom of this page for more detail about how 
-this works in the current |version| version of Intel nGraph library 
-documentation. 
+this works in the current |version| version of nGraph Library documentation.
 
 
 Adding captions to code blocks 

doc/sphinx/source/project/release-notes.rst

@@ -6,34 +6,50 @@ Release Notes
 nGraph is provided as source code, APIs, build scripts, and some binary formats 
 for various Compiler stack configurations and use cases. 
 
+For downloads formatted as ``.zip`` and ``tar.gz``, see 
+https://github.com/NervanaSystems/ngraph/releases.
+
 This page includes additional documentation updates.
 
 We are pleased to announce the release of version |version|-doc.
 
 
-0.21-doc
---------
+Core updates for |version|
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-Documentation updates
-~~~~~~~~~~~~~~~~~~~~~
++ PlaidML support
++ More ONNX ops
++ Elementwise divide defaults to Python semantics
++ GenerateMask seed optional
 
-Summary of documentation-related changes:
 
-+ Update :doc:`doc-contributor-README` for new community-based contributions. 
-+ Added instructions on how to test or display the installed nGraph version.
-+ Added instructions on building nGraph bridge (ngraph-bridge).
-+ Updated Backend Developer Guides and ToC structure.
-+ Tested documentation build on Clear Linux OS; it works.
-+ Fixed a few links and redirs affected by filename changes.
-+ Some coding adjustments for options to render math symbols, so they can be 
-  documented more clearly and without excessive JS (see replacements.txt).
-+ Consistent filenaming on all BE indexes.
-+ Remove deprecated TensorAPI.
-+ 
+Latest doc updates
+~~~~~~~~~~~~~~~~~~
 
++ Document new debug tool
++ Note deprecation of MXNet's ``ngraph-mxnet`` PyPI
++ Note default change to `svg` files for graphs and visualization
++ Add more prominent tips for contributors who find the doc-contributor-README
+
+
+.. important:: Pre-releases (``-rc-0.*``) have newer features, and are less stable.  
+
+
+Changelog on Previous Releases
+==============================
+
+0.22
+----
+
++ More ONNX ops
++ Optimizations
++ Don't reseed RNG on each use
++ Initial doc and API for IntelGPU backend 
++ DynamicBackend API
 
-Core updates for |version|
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+0.21
+----
 
 + The offset argument in tensor reads and writes has been removed
 + Save/load API
@@ -43,17 +59,19 @@ Core updates for |version|
 + Provenance improvements
 + offset arg for tensor creation is deprecated
 + static linking support
++ Initial test of 0.21-doc
++ Updated :doc:`doc-contributor-README` for new community-based contributions. 
++ Added instructions on how to test or display the installed nGraph version.
++ Added instructions on building nGraph bridge (ngraph-bridge).
++ Updated Backend Developer Guides and ToC structure.
++ Tested documentation build on Clear Linux OS; it works.
++ Fixed a few links and redirs affected by filename changes.
++ Some coding adjustments for options to render math symbols, so they can be 
+  documented more clearly and without excessive JS (see replacements.txt).
++ Consistent filenaming on all BE indexes.
++ Removed deprecated TensorAPI.
 
 
-.. important:: Pre-releases (``-rc-0.*``) have newer features, and are less stable.  
-
-
-Changelog on Previous Releases
-==============================
-
-For downloads formatted as ``.zip`` and ``tar.gz``, see 
-https://github.com/NervanaSystems/ngraph/releases.
-
 0.20
 ----
 
@@ -67,8 +85,6 @@ https://github.com/NervanaSystems/ngraph/releases.
 0.19
 ----
 
-**Download** `0.19.0-rc.2`_
-
 + More dynamic shape preparation
 + Distributed interface factored out
 + fp16 and bfloat16 types
@@ -88,9 +104,6 @@ https://github.com/NervanaSystems/ngraph/releases.
 0.18
 ----
 
-**Download** `0.18.1`_
-
-
 + Python formatting issue
 + mkl-dnn work-around
 + Event tracing improvements
@@ -103,8 +116,6 @@ https://github.com/NervanaSystems/ngraph/releases.
 0.17
 ----
 
-**Download** `0.17.0-rc.1`_
-
 + Allow negative padding in more places
 + Add code generation for some quantized ops
 + Preliminary dynamic shape support
@@ -116,11 +127,6 @@ https://github.com/NervanaSystems/ngraph/releases.
 0.16
 ----
 
-* **Download**: `0.16.0-rc.3`_
-* **Download** `0.16.0-rc.2`_
-* **Download** `0.16.0-rc.1`_
-
-
 + NodeInput and NodeOutput classes prepare for simplifications of Node
 + Test improvements
 + Additional quantization ops
@@ -128,11 +134,3 @@ https://github.com/NervanaSystems/ngraph/releases.
 + Fix memory leak
 + Concat optimization
 + Doc updates
-
-.. _0.20.0-rc.0: https://github.com/NervanaSystems/ngraph/releases/tag/v0.20.0-rc.0_
-.. _0.19.0-rc.2: https://github.com/NervanaSystems/ngraph/releases/tag/v0.19.0-rc.2_
-.. _0.18.1: https://github.com/NervanaSystems/ngraph/releases/tag/v0.18.1_
-.. _0.17.0-rc.1: `https://github.com/NervanaSystems/ngraph/releases/tag/v0.17.0-rc.1
-.. _0.16.0-rc.3: https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.3
-.. _0.16.0-rc.2: https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.2
-.. _0.16.0-rc.1: https://github.com/NervanaSystems/ngraph/releases/tag/v0.16.0-rc.1

python/BUILDING.md

@@ -2,10 +2,10 @@
 
 ## Building nGraph Python Wheels
 
-If you want to try a newer version of nGraph's Python API than is available from
-PyPI, you can build your own latest version from the source code. This 
-process is very similar to what is outlined in our [ngraph_build] instructions 
-with two important differences:
+If you want to try a newer version of nGraph's Python API than is available 
+from PyPI, you can build the latest version from source code. This process is 
+very similar to what is outlined in our [ngraph_build] instructions with two 
+important differences:
 
 1. You must specify: `-DNGRAPH_PYTHON_BUILD_ENABLE=ON` and `-DNGRAPH_ONNX_IMPORT_ENABLE=ON` 
    when running `cmake`.
@@ -18,7 +18,7 @@ with two important differences:
 
 After this procedure completes, the `ngraph/build/python/dist` directory should 
 contain the Python packages of the version you cloned. For example, if you 
-checked out and built `0.21`, you may see something like: 
+checked out and built `0.21` for Python 3.7, you might see something like: 
 
     $ ls python/dist/
     ngraph-core-0.21.0rc0.tar.gz  

python/README.md

@@ -18,8 +18,8 @@ boost compared to native implementations.
 
 nGraph can be used directly with the [Python API][api_python] described here, or 
 with the [C++ API][api_cpp] described in the [core documentation]. Alternatively, 
-its performance benefits can be realized through a frontend such as 
-[TensorFlow][frontend_tf], [MXNet][frontend_mxnet], and [ONNX][frontend_onnx].
+its performance benefits can be realized through frontends such as 
+[TensorFlow][frontend_tf], [PaddlePaddle][paddle_paddle] and [ONNX][frontend_onnx].
 You can also create your own custom framework to integrate directly with the 
 [nGraph Ops] for highly-targeted graph execution.
 
@@ -77,7 +77,7 @@ print('Result = ', result)
 
 [up to 45X]: https://ai.intel.com/ngraph-compiler-stack-beta-release/
 [frontend_onnx]: https://pypi.org/project/ngraph-onnx/
-[frontend_mxnet]: https://pypi.org/project/ngraph-mxnet/ 
+[paddle_paddle]: https://ngraph.nervanasys.com/docs/latest/frameworks/paddle_integ.html 
 [frontend_tf]: https://pypi.org/project/ngraph-tensorflow-bridge/
 [ngraph_github]: https://github.com/NervanaSystems/ngraph "nGraph on GitHub"
 [ngraph_building]: https://github.com/NervanaSystems/ngraph/blob/master/python/BUILDING.md "Building nGraph"

src/ngraph/CMakeLists.txt

@@ -370,7 +370,6 @@ set (SRC
     op/util/index_reduction.hpp
     op/util/logical_reduction.cpp
     op/util/logical_reduction.hpp
-    op/util/reshape.hpp
     op/util/rnn_cell_base.cpp
     op/util/rnn_cell_base.hpp
     op/util/unary_elementwise_arithmetic.cpp
@@ -458,6 +457,8 @@ set (SRC
     rank.hpp
     runtime/aligned_buffer.cpp
     runtime/aligned_buffer.hpp
+    runtime/allocator.cpp
+    runtime/allocator.hpp
     runtime/backend.cpp
     runtime/backend.hpp
     runtime/backend_manager.cpp

src/ngraph/distributed.cpp

@@ -46,7 +46,7 @@ static std::unique_ptr<DistributedInterface> s_distributed_interface;
 
 void ngraph::set_distributed_interface(std::unique_ptr<DistributedInterface> distributed_interface)
 {
-    NGRAPH_DEBUG << "Setting distributed interfsce to: " << distributed_interface->get_name();
+    NGRAPH_DEBUG << "Setting distributed interface to: " << distributed_interface->get_name();
     s_distributed_interface = std::move(distributed_interface);
 }
 

src/ngraph/op/fused/group_conv_transpose.cpp

@@ -203,7 +203,8 @@ shared_ptr<Node> op::GroupConvolutionTranspose::copy_with_new_args(const NodeVec
                                                       get_padding_end(),
                                                       get_output_padding(),
                                                       get_groups(),
-                                                      get_pad_type());
+                                                      get_pad_type(),
+                                                      get_output_shape());
 }
 
 Shape op::GroupConvolutionTranspose::get_data_batch_shape() const

src/ngraph/op/util/reshape.hpp

-//*****************************************************************************
-// Copyright 2017-2019 Intel Corporation
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//*****************************************************************************
-
-#pragma once
-
-#include <cstddef>
-#include <memory>
-#include <vector>
-
-#include "ngraph/builder/reshape.hpp"
-#include "ngraph/node.hpp"
-#include "ngraph/shape.hpp"
-
-namespace ngraph
-{
-    namespace op
-    {
-        namespace util
-        {
-            /// \brief      Change shape of input tensor.
-            ///
-            /// \param[in]  node   The node producing the tensor to be reshaped.
-            /// \param[in]  shape  The new shape for input tensor.
-            ///
-            /// \return     The node representing a Reshape operation.
-            ///
-            std::shared_ptr<ngraph::Node> reshape(const std::shared_ptr<ngraph::Node>& node,
-                                                  const Shape& shape)
-            {
-                return builder::reshape(node, shape);
-            }
-
-            /// \brief Permute axes according to specified axes_order parameter.
-            ///
-            /// \param node The node which axes we want to permute.
-            /// \param axes_order The permutation of node tensor axes.
-            ///
-            /// \return: New node with permuted axes.
-            std::shared_ptr<ngraph::Node> reorder_axes(const std::shared_ptr<ngraph::Node>& node,
-                                                       std::vector<std::size_t> axes_order)
-            {
-                return builder::reorder_axes(node, axes_order);
-            }
-
-            /// \brief Return transposed tensor (with axes in reversed order).
-            ///
-            /// \param node Input tensor we want to transpose
-            ///
-            /// \return: New node with reversed dimensions.
-            std::shared_ptr<ngraph::Node> transpose(const std::shared_ptr<ngraph::Node>& node)
-            {
-                return builder::transpose(node);
-            }
-
-            /// \brief Flatten the input tensor into a 2D matrix.
-            ///
-            /// \param node The tensor to be flattened.
-            /// \param axis The axis dividing shape.
-            ///
-            /// \return The new node will be a 2D matrix representing the flattened input node.
-            std::shared_ptr<ngraph::Node> flatten(const std::shared_ptr<ngraph::Node>& node,
-                                                  int axis)
-            {
-                return builder::flatten(node, axis);
-            }
-        } // namespace util
-    }     // namespace  op
-} // namespace  ngraph

src/ngraph/runtime/aligned_buffer.cpp

@@ -18,22 +18,34 @@
 #include <memory>
 
 #include "ngraph/runtime/aligned_buffer.hpp"
+#include "ngraph/runtime/allocator.hpp"
 #include "ngraph/util.hpp"
 
 using namespace ngraph;
+using namespace std;
 
 runtime::AlignedBuffer::AlignedBuffer()
-    : m_allocated_buffer(nullptr)
+    : m_allocator(nullptr)
+    , m_allocated_buffer(nullptr)
     , m_aligned_buffer(nullptr)
     , m_byte_size(0)
 {
 }
 
-runtime::AlignedBuffer::AlignedBuffer(size_t byte_size, size_t alignment)
+runtime::AlignedBuffer::AlignedBuffer(size_t byte_size, size_t alignment, Allocator* allocator)
+    : m_allocator(allocator)
+    , m_byte_size(byte_size)
 {
     m_byte_size = std::max<size_t>(1, byte_size);
     size_t allocation_size = m_byte_size + alignment;
-    m_allocated_buffer = static_cast<char*>(ngraph_malloc(allocation_size));
+    if (allocator)
+    {
+        m_allocated_buffer = static_cast<char*>(m_allocator->malloc(allocation_size, alignment));
+    }
+    else
+    {
+        m_allocated_buffer = static_cast<char*>(malloc(allocation_size));
+    }
     m_aligned_buffer = m_allocated_buffer;
     size_t mod = size_t(m_aligned_buffer) % alignment;
 
@@ -44,10 +56,12 @@ runtime::AlignedBuffer::AlignedBuffer(size_t byte_size, size_t alignment)
 }
 
 runtime::AlignedBuffer::AlignedBuffer(AlignedBuffer&& other)
-    : m_allocated_buffer(other.m_allocated_buffer)
+    : m_allocator(other.m_allocator)
+    , m_allocated_buffer(other.m_allocated_buffer)
     , m_aligned_buffer(other.m_aligned_buffer)
     , m_byte_size(other.m_byte_size)
 {
+    other.m_allocator = nullptr;
     other.m_allocated_buffer = nullptr;
     other.m_aligned_buffer = nullptr;
     other.m_byte_size = 0;
@@ -57,7 +71,14 @@ runtime::AlignedBuffer::~AlignedBuffer()
 {
     if (m_allocated_buffer != nullptr)
     {
-        ngraph_free(m_allocated_buffer);
+        if (m_allocator)
+        {
+            m_allocator->free(m_allocated_buffer);
+        }
+        else
+        {
+            free(m_allocated_buffer);
+        }
     }
 }
 
@@ -65,9 +86,11 @@ runtime::AlignedBuffer& runtime::AlignedBuffer::operator=(AlignedBuffer&& other)
 {
     if (this != &other)
     {
+        m_allocator = other.m_allocator;
         m_allocated_buffer = other.m_allocated_buffer;
         m_aligned_buffer = other.m_aligned_buffer;
         m_byte_size = other.m_byte_size;
+        other.m_allocator = nullptr;
         other.m_allocated_buffer = nullptr;
         other.m_aligned_buffer = nullptr;
         other.m_byte_size = 0;

src/ngraph/runtime/aligned_buffer.hpp

@@ -18,6 +18,8 @@
 
 #include <cstddef>
 
+#include "ngraph/runtime/allocator.hpp"
+
 namespace ngraph
 {
     namespace runtime
@@ -32,7 +34,11 @@ namespace ngraph
 class ngraph::runtime::AlignedBuffer
 {
 public:
-    AlignedBuffer(size_t byte_size, size_t alignment);
+    // Allocator objects and the allocation interfaces are owned by the
+    // creators of AlignedBuffers. They need to ensure that the lifetime of
+    // allocator exceeds the lifetime of this AlignedBuffer.
+    AlignedBuffer(size_t byte_size, size_t alignment, Allocator* allocator = nullptr);
+
     AlignedBuffer();
     ~AlignedBuffer();
 
@@ -46,6 +52,7 @@ private:
     AlignedBuffer(const AlignedBuffer&) = delete;
     AlignedBuffer& operator=(const AlignedBuffer&) = delete;
 
+    Allocator* m_allocator;
     char* m_allocated_buffer;
     char* m_aligned_buffer;
     size_t m_byte_size;

src/ngraph/runtime/allocator.cpp

+//*****************************************************************************
+// Copyright 2017-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#include "ngraph/runtime/allocator.hpp"
+
+ngraph::runtime::Allocator::~Allocator()
+{
+}
+
+class ngraph::runtime::DefaultAllocator : public ngraph::runtime::Allocator
+{
+public:
+    void* malloc(size_t size, size_t alignment)
+    {
+        // If allocation succeeds, returns a pointer to the lowest (first) byte in the
+        // allocated memory block that is suitably aligned for any scalar type.
+        // TODO(pruthvi): replace std::malloc with custom aligned_alloc implementation
+        // which is portable and work on all alignment requirement.
+        void* ptr = std::malloc(size);
+
+        // check for exception
+        if (!ptr)
+        {
+            throw ngraph::ngraph_error("malloc failed to allocate memory of size " +
+                                       std::to_string(size));
+        }
+        return ptr;
+    }
+
+    void free(void* ptr)
+    {
+        if (ptr)
+        {
+            std::free(ptr);
+        }
+    }
+};
+
+std::unique_ptr<ngraph::runtime::Allocator> ngraph::runtime::create_default_allocator()
+{
+    return std::unique_ptr<DefaultAllocator>(new DefaultAllocator());
+}

src/ngraph/runtime/allocator.hpp

+//*****************************************************************************
+// Copyright 2017-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include "ngraph/except.hpp"
+#include "ngraph/util.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        class Allocator;
+        class DefaultAllocator;
+        /// \brief Create a default allocator that calls into system
+        ///        allocation libraries
+        std::unique_ptr<Allocator> create_default_allocator();
+    }
+}
+
+/// \brief Abstract class for the allocator
+class ngraph::runtime::Allocator
+{
+public:
+    virtual ~Allocator() = 0;
+    /// \brief allocates memory with the given size and alignment requirement
+    /// \param size exact size of bytes to allocate
+    /// \param alignment specifies the alignment. Must be a valid alignment supported by the implementation.
+    virtual void* malloc(size_t size, size_t alignment) = 0;
+
+    /// \brief deallocates the memory pointed by ptr
+    /// \param ptr pointer to the aligned memory to be released
+    virtual void free(void* ptr) = 0;
+};

src/ngraph/runtime/backend.cpp

@@ -93,6 +93,13 @@ void runtime::Backend::remove_compiled_function(std::shared_ptr<Executable> exec
 {
 }
 
+bool runtime::Backend::is_device_memory(void* ptr)
+{
+    // override this method for each supported backend to determine if the passed pointer is in
+    // device pinned memory or not
+    return false;
+}
+
 std::shared_ptr<runtime::Executable> runtime::Backend::load(istream& input_stream)
 {
     throw runtime_error("load opertion unimplemented.");

src/ngraph/runtime/backend.hpp

@@ -21,10 +21,12 @@
 
 #include "ngraph/function.hpp"
 #include "ngraph/pass/pass_config.hpp"
+#include "ngraph/runtime/allocator.hpp"
 #include "ngraph/runtime/executable.hpp"
 #include "ngraph/runtime/performance_counter.hpp"
 #include "ngraph/shape.hpp"
 #include "ngraph/type/element_type.hpp"
+#include "ngraph/util.hpp"
 
 namespace ngraph
 {
@@ -142,6 +144,23 @@ public:
     /// \returns a shared pointer to the op if found, else nullptr
     virtual std::shared_ptr<ngraph::Node> get_backend_op(const std::string& op_name, ...);
 
+    /// \brief Returns memory allocator used by backend for host allocations
+    virtual Allocator* get_host_memory_allocator() { return nullptr; }
+    /// \brief Set the host memory allocator to be used by the backend
+    /// \param allocator is pointer to host memory allocator object
+    virtual void set_host_memory_allocator(std::unique_ptr<Allocator> allocator) {}
+    /// \brief Returns memory allocator used by backend for device allocations
+    virtual Allocator* get_device_memory_allocator()
+    {
+        // override this method from each supported backend to return
+        // its own device memory allocator
+        return nullptr;
+    }
+
+    /// \brief method for each supported backend to determine if the passed pointer is in device pinned memory or not
+    /// \param ptr pointer to the memory to determine if its in device memory or not
+    virtual bool is_device_memory(void* ptr);
+
     /// \brief Allows sending backend specific configuration. The map contains key, value pairs
     ///     specific to a particluar backend. The definition of these key, value pairs is
     ///     defined by each backend.

src/ngraph/runtime/cpu/cpu_backend.cpp

@@ -63,11 +63,17 @@ namespace
     } s_cpu_static_init;
 }
 
+runtime::cpu::CPU_Backend::~CPU_Backend()
+{
+    m_exec_map.clear();
+}
+
 shared_ptr<runtime::cpu::CPU_CallFrame> runtime::cpu::CPU_Backend::make_call_frame(
     const shared_ptr<runtime::cpu::CPU_ExternalFunction>& external_function,
-    ngraph::pass::PassConfig& pass_config)
+    ngraph::pass::PassConfig& pass_config,
+    Allocator* allocator)
 {
-    return external_function->make_call_frame(pass_config);
+    return external_function->make_call_frame(pass_config, allocator);
 }
 
 shared_ptr<runtime::Tensor>
@@ -114,7 +120,8 @@ shared_ptr<runtime::Executable>
             return rc;
         }
     }
-    rc = make_shared<CPU_Executable>(func, pass_config, performance_counters_enabled);
+    rc = make_shared<CPU_Executable>(
+        func, pass_config, get_host_memory_allocator(), performance_counters_enabled);
     {
         std::lock_guard<std::mutex> guard(m_exec_map_mutex);
         m_exec_map.insert({func, rc});
@@ -124,6 +131,7 @@ shared_ptr<runtime::Executable>
 
 runtime::cpu::CPU_Executable::CPU_Executable(shared_ptr<Function> func,
                                              ngraph::pass::PassConfig& pass_config,
+                                             Allocator* allocator,
                                              bool performance_counters_enabled)
 {
     FunctionInstance& instance = m_function_instance;
@@ -131,7 +139,7 @@ runtime::cpu::CPU_Executable::CPU_Executable(shared_ptr<Function> func,
     {
         instance.m_external_function = make_shared<CPU_ExternalFunction>(func);
         instance.m_external_function->m_emit_timing = performance_counters_enabled;
-        auto cf = instance.m_external_function->make_call_frame(pass_config);
+        auto cf = instance.m_external_function->make_call_frame(pass_config, allocator);
         instance.m_call_frame = dynamic_pointer_cast<CPU_CallFrame>(cf);
     }
     set_parameters_and_results(*func);
@@ -173,6 +181,28 @@ void runtime::cpu::CPU_Backend::remove_compiled_function(shared_ptr<Executable>
     }
 }
 
+runtime::Allocator* runtime::cpu::CPU_Backend::get_host_memory_allocator()
+{
+    if (!m_allocator)
+    {
+        m_allocator = create_default_allocator();
+    }
+    return m_allocator.get();
+}
+
+void runtime::cpu::CPU_Backend::set_host_memory_allocator(
+    std::unique_ptr<runtime::Allocator> allocator)
+{
+    if (m_allocator)
+    {
+        // Resources allocated with the existing allocator might still be around and expect it
+        // to be available for freeing. We cannot switch to the new allocator
+        throw ngraph_error(
+            "Allocator already exists. Changing allocators mid-execution is not permitted.");
+    }
+    m_allocator = std::move(allocator);
+}
+
 vector<runtime::PerformanceCounter> runtime::cpu::CPU_Executable::get_performance_data() const
 {
     vector<runtime::PerformanceCounter> rc;
@@ -190,6 +220,7 @@ bool runtime::cpu::CPU_Backend::is_supported(const Node& op) const
 {
     return true;
 }
+
 bool runtime::cpu::CPU_Backend::is_supported_property(const Property prop) const
 {
     if (prop == Property::memory_attach)

src/ngraph/runtime/cpu/cpu_backend.hpp

@@ -22,6 +22,7 @@
 
 #include "cpu_backend_visibility.h"
 #include "ngraph/pass/pass_config.hpp"
+#include "ngraph/runtime/allocator.hpp"
 #include "ngraph/runtime/backend.hpp"
 
 namespace ngraph
@@ -36,9 +37,12 @@ namespace ngraph
             class CPU_BACKEND_API CPU_Backend : public runtime::Backend
             {
             public:
+                ~CPU_Backend() override;
+
                 std::shared_ptr<CPU_CallFrame>
                     make_call_frame(const std::shared_ptr<CPU_ExternalFunction>& external_function,
-                                    ngraph::pass::PassConfig& pass_config);
+                                    ngraph::pass::PassConfig& pass_config,
+                                    Allocator* allocator);
 
                 std::shared_ptr<ngraph::runtime::Tensor>
                     create_tensor(const ngraph::element::Type& element_type,
@@ -60,6 +64,9 @@ namespace ngraph
 
                 void remove_compiled_function(std::shared_ptr<Executable> exec) override;
 
+                Allocator* get_host_memory_allocator() override;
+                void set_host_memory_allocator(std::unique_ptr<Allocator> allocator) override;
+
                 bool is_supported(const Node& node) const override;
                 bool is_supported_property(const Property prop) const override;
 
@@ -69,6 +76,7 @@ namespace ngraph
                 std::mutex m_exec_map_mutex;
                 std::unordered_map<std::shared_ptr<Function>, std::shared_ptr<Executable>>
                     m_exec_map;
+                std::unique_ptr<Allocator> m_allocator;
             };
 
             class CPU_BACKEND_API CPU_Executable : public runtime::Executable
@@ -76,6 +84,7 @@ namespace ngraph
             public:
                 CPU_Executable(std::shared_ptr<Function> func,
                                ngraph::pass::PassConfig& pass_config,
+                               Allocator* allocator,
                                bool performance_counters_enabled);
                 bool call(const std::vector<std::shared_ptr<runtime::Tensor>>& outputs,
                           const std::vector<std::shared_ptr<runtime::Tensor>>& inputs) override;

src/ngraph/runtime/cpu/cpu_call_frame.cpp

@@ -30,7 +30,8 @@ using namespace ngraph;
 runtime::cpu::CPU_CallFrame::CPU_CallFrame(std::shared_ptr<CPU_ExternalFunction> external_function,
                                            InitContextFuncCG compiled_init_ctx_func,
                                            DestroyContextFuncCG compiled_destroy_ctx_func,
-                                           EntryPoint compiled_function)
+                                           EntryPoint compiled_function,
+                                           runtime::Allocator* allocator)
     : m_external_function(external_function)
     , m_compiled_init_ctx_func(compiled_init_ctx_func)
     , m_compiled_destroy_ctx_func(compiled_destroy_ctx_func)
@@ -47,7 +48,7 @@ runtime::cpu::CPU_CallFrame::CPU_CallFrame(std::shared_ptr<CPU_ExternalFunction>
             std::to_string(std::thread::hardware_concurrency()) + "]");
     }
 
-    setup_runtime_context();
+    setup_runtime_context(allocator);
     if (!m_external_function->is_direct_execution())
     {
         // Invoke codegen runtime context initialization function.
@@ -178,7 +179,7 @@ void runtime::cpu::CPU_CallFrame::propagate_layouts(
     }
 }
 
-void runtime::cpu::CPU_CallFrame::setup_runtime_context()
+void runtime::cpu::CPU_CallFrame::setup_runtime_context(Allocator* allocator)
 {
     for (auto i = 0; i < m_num_ctx; i++)
     {
@@ -202,7 +203,7 @@ void runtime::cpu::CPU_CallFrame::setup_runtime_context()
         size_t alignment = runtime::cpu::CPU_ExternalFunction::s_memory_pool_alignment;
         for (auto buffer_size : m_external_function->get_memory_buffer_sizes())
         {
-            auto buffer = new AlignedBuffer(buffer_size, alignment);
+            auto buffer = new AlignedBuffer(buffer_size, alignment, allocator);
             ctx->memory_buffers.push_back(buffer);
         }
         const auto& mkldnn_emitter = m_external_function->get_mkldnn_emitter();

src/ngraph/runtime/cpu/cpu_call_frame.hpp

@@ -24,6 +24,7 @@
 #include <vector>
 
 #include "ngraph/function.hpp"
+#include "ngraph/runtime/allocator.hpp"
 #include "ngraph/runtime/cpu/cpu_layout_descriptor.hpp"
 #include "ngraph/runtime/cpu/cpu_runtime_context.hpp"
 #include "ngraph/runtime/tensor.hpp"
@@ -59,7 +60,8 @@ namespace ngraph
                 CPU_CallFrame(std::shared_ptr<CPU_ExternalFunction> external_function,
                               InitContextFuncCG compiled_init_ctx_func,
                               DestroyContextFuncCG compiled_destroy_ctx_func,
-                              EntryPoint compiled_function);
+                              EntryPoint compiled_function,
+                              runtime::Allocator* allocator);
                 ~CPU_CallFrame();
 
                 /// \brief Invoke the function with values matching the signature of the function.
@@ -71,7 +73,7 @@ namespace ngraph
                 void propagate_layouts(const std::vector<std::shared_ptr<runtime::Tensor>>& tvs,
                                        const LayoutDescriptorPtrs& layouts) const;
 
-                void setup_runtime_context();
+                void setup_runtime_context(runtime::Allocator* allocator);
                 void setup_cg_runtime_context();
                 void cleanup_runtime_context();
 

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1825,7 +1825,8 @@ bool runtime::cpu::CPU_ExternalFunction::is_codegen(const ngraph::pass::PassConf
 }
 
 shared_ptr<ngraph::runtime::cpu::CPU_CallFrame>
-    runtime::cpu::CPU_ExternalFunction::make_call_frame(ngraph::pass::PassConfig& pass_config)
+    runtime::cpu::CPU_ExternalFunction::make_call_frame(ngraph::pass::PassConfig& pass_config,
+                                                        Allocator* allocator)
 {
 #if defined(NGRAPH_DEX_ONLY)
     if (is_codegen(pass_config))
@@ -1853,7 +1854,8 @@ shared_ptr<ngraph::runtime::cpu::CPU_CallFrame>
     return make_shared<ngraph::runtime::cpu::CPU_CallFrame>(shared_from_this(),
                                                             m_compiled_init_ctx_func,
                                                             m_compiled_destroy_ctx_func,
-                                                            m_compiled_function);
+                                                            m_compiled_function,
+                                                            allocator);
 }
 
 const runtime::cpu::LayoutDescriptorPtrs&

src/ngraph/runtime/cpu/cpu_external_function.hpp

@@ -100,8 +100,7 @@ namespace ngraph
                                      bool release_function = true);
                 ~CPU_ExternalFunction();
                 std::shared_ptr<ngraph::runtime::cpu::CPU_CallFrame>
-                    make_call_frame(ngraph::pass::PassConfig& pass_config);
-
+                    make_call_frame(ngraph::pass::PassConfig& pass_config, Allocator* allocator);
                 const LayoutDescriptorPtrs& get_parameter_layout_descriptors();
                 const LayoutDescriptorPtrs& get_result_layout_descriptors();
                 const std::vector<size_t>& get_memory_buffer_sizes() const

src/ngraph/runtime/cpu/cpu_runtime_context.hpp

@@ -37,7 +37,6 @@ namespace ngraph
     {
         class AlignedBuffer;
     }
-
     class State;
 }
 

test/models/onnx/convtranspose_output_shape.prototxt

+ir_version: 3
+producer_name: "backend-test"
+graph {
+  node {
+    input: "X"
+    input: "W"
+    output: "Y"
+    op_type: "ConvTranspose"
+    attribute {
+      name: "output_shape"
+      ints: 10
+      ints: 8
+      type: INTS
+    }
+    attribute {
+      name: "strides"
+      ints: 3
+      ints: 2
+      type: INTS
+    }
+  }
+  name: "test_convtranspose_output_shape"
+  input {
+    name: "X"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 3
+          }
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+  }
+  input {
+    name: "W"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 2
+          }
+          dim {
+            dim_value: 3
+          }
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+  }
+  output {
+    name: "Y"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 2
+          }
+          dim {
+            dim_value: 10
+          }
+          dim {
+            dim_value: 8
+          }
+        }
+      }
+    }
+  }
+}
+opset_import {
+  version: 9
+}

test/onnx/onnx_import_convpool.in.cpp

@@ -30,6 +30,7 @@
 #include "util/all_close.hpp"
 #include "util/all_close_f.hpp"
 #include "util/ndarray.hpp"
+#include "util/test_case.hpp"
 #include "util/test_control.hpp"
 #include "util/test_tools.hpp"
 
@@ -409,3 +410,19 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, model_global_lp_pool_p3)
 
     EXPECT_TRUE(test::all_close_f(expected_outputs.front(), outputs.front()));
 }
+
+NGRAPH_TEST(onnx_${BACKEND_NAME}, model_convtranspose_output_shape)
+{
+    auto conv_transpose_fn = onnx_import::import_onnx_model(
+        file_util::path_join(SERIALIZED_ZOO, "onnx/convtranspose_output_shape.prototxt"));
+
+    auto test_case = ngraph::test::NgraphTestCase(conv_transpose_fn, "${BACKEND_NAME}");
+
+    test_case.add_input_from_file<float>(TEST_FILES, "onnx/convtranspose_output_shape/x.bin");
+    test_case.add_input_from_file<float>(TEST_FILES, "onnx/convtranspose_output_shape/w.bin");
+    test_case.add_expected_output_from_file<float>(
+        {1, 2, 10, 8}, TEST_FILES, "onnx/convtranspose_output_shape/y.bin");
+
+    test_case.dump_results();
+    test_case.run();
+}