Merge branch 'master' into master

.gitignore

@@ -115,3 +115,7 @@ python/pybind11/
 CMakeCache.txt
 CMakeFiles/
 CMakeSettings.json
+
+# don't add dot-save files!
+
+.save

README.md

@@ -96,7 +96,7 @@ to improve it:
 [framework integration guides]: http://ngraph.nervanasys.com/docs/latest/framework-integration-guides.html
 [release notes]: https://ngraph.nervanasys.com/docs/latest/project/release-notes.html
 [Github issues]: https://github.com/NervanaSystems/ngraph/issues
-[contrib guide]: https://ngraph.nervanasys.com/docs/latest/project/code-contributor-README.html
+[contrib guide]: https://ngraph.nervanasys.com/docs/latest/project/contribution-guide.html
 [pull request]: https://github.com/NervanaSystems/ngraph/pulls
 [how to import]: https://ngraph.nervanasys.com/docs/latest/howto/import.html
 [ngraph_wireframes_with_notice]: doc/sphinx/source/graphics/readme_stack.png "nGraph wireframe"

doc/sphinx/ngraph.doxyfile

@@ -15,7 +15,7 @@
 
 
 
-PROJECT_NAME     = "Intel® nGraph Library and API docs"
+PROJECT_NAME     = "Intel® nGraph Compiler stack and API docs"
 PROJECT_BRIEF    = "Code reference for the Intel® nGraph C++ Library"
 
 OUTPUT_DIRECTORY = @CMAKE_CURRENT_BINARY_DIR@
@@ -240,7 +240,7 @@ EXTENSION_MAPPING      =
 # case of backward compatibilities issues.
 # The default value is: YES.
 
-# MARKDOWN_SUPPORT       = NO
+MARKDOWN_SUPPORT       = NO
 
 # When the TOC_INCLUDE_HEADINGS tag is set to a non-zero value, all headings up
 # to that level are automatically included in the table of contents, even if
@@ -405,7 +405,7 @@ EXTRACT_LOCAL_METHODS  = NO
 # the file that contains the anonymous namespace. By default anonymous namespace
 # are hidden.
 
-EXTRACT_ANON_NSPACES   = YES
+EXTRACT_ANON_NSPACES   = NO
 
 # If the HIDE_UNDOC_MEMBERS tag is set to YES, doxygen will hide all
 # undocumented members inside documented classes or files. If set to NO these
@@ -615,7 +615,7 @@ LAYOUT_FILE            =
 # LATEX_BIB_STYLE. To use this feature you need bibtex and perl available in the
 # search path. See also \cite for info how to create references.
 
-CITE_BIB_FILES         =
+# CITE_BIB_FILES         =
 
 #---------------------------------------------------------------------------
 # Configuration options related to warning and progress messages
@@ -626,7 +626,7 @@ CITE_BIB_FILES         =
 # messages are off.
 # The default value is: NO.
 
-QUIET                  = NO
+QUIET                  = YES
 
 # The WARNINGS tag can be used to turn on/off the warning messages that are
 # generated to standard error (stderr) by doxygen. If WARNINGS is set to YES
@@ -635,14 +635,14 @@ QUIET                  = NO
 # Tip: Turn warnings on while writing the documentation.
 # The default value is: YES.
 
-WARNINGS               = YES
+WARNINGS               = NO
 
 # If the WARN_IF_UNDOCUMENTED tag is set to YES then doxygen will generate
 # warnings for undocumented members. If EXTRACT_ALL is set to YES then this flag
 # will automatically be disabled.
 # The default value is: YES.
 
-WARN_IF_UNDOCUMENTED   = YES
+WARN_IF_UNDOCUMENTED   = NO
 
 # If the WARN_IF_DOC_ERROR tag is set to YES, doxygen will generate warnings for
 # potential errors in the documentation, such as not documenting some parameters
@@ -786,7 +786,8 @@ EXCLUDE_SYMLINKS       = NO
 # Note that the wildcards are matched against the file with absolute path, so to
 # exclude all test directories for example use the pattern */test/*
 
-EXCLUDE_PATTERNS       =
+EXCLUDE_PATTERNS       = */tools/*
+                         */resource/*  
 
 # The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
 # (namespaces, classes, functions, etc.) that should be excluded from the
@@ -1176,7 +1177,7 @@ DOCSET_PUBLISHER_ID    = org.doxygen.Publisher
 # The default value is: Publisher.
 # This tag requires that the tag GENERATE_DOCSET is set to YES.
 
-DOCSET_PUBLISHER_NAME  = Publisher
+DOCSET_PUBLISHER_NAME  = AIPG Intel Corporation
 
 # If the GENERATE_HTMLHELP tag is set to YES then doxygen generates three
 # additional HTML index files: index.hhp, index.hhc, and index.hhk. The
@@ -1897,7 +1898,7 @@ HAVE_DOT               = YES
 # speed.
 # This tag requires that the tag HAVE_DOT is set to YES.
 
-DOT_NUM_THREADS        = 32
+DOT_NUM_THREADS        = 0
 
 # When you want a differently looking font in the dot files that doxygen
 # generates you can specify the font name using DOT_FONTNAME. You need to make
@@ -2144,4 +2145,4 @@ GENERATE_LEGEND        = YES
 # The default value is: YES.
 # This tag requires that the tag HAVE_DOT is set to YES.
 
-DOT_CLEANUP            = YES
\ No newline at end of file
+DOT_CLEANUP            = YES

doc/sphinx/ngraph_theme/layout.html

@@ -77,9 +77,6 @@
   {# Keep modernizr in head - http://modernizr.com/docs/#installing #}
   <script src="{{ pathto('_static/js/modernizr.min.js', 1) }}"></script>
 
-  {# swagger-ui loading for API doc display #}
-  <link rel="stylesheet" href="{{ pathto('_static/css/swagger.css', 1) }}" type="text/css" />
-
 </head>
 
 <body class="wy-body-for-nav" role="document">

doc/sphinx/ngraph_theme/static/css/theme.css

@@ -1638,9 +1638,9 @@ html {
 body {
   font-family: "RobotoSlab", Sans, sans-serif;
   font-weight: normal;
-  color: #38403f;
+  color: #3f3c3d;
   min-height: 100%;
-  overflow-x: hidden;
+  overflow-x: visible;
   background: #fcfcfc;
 }
 
@@ -1731,11 +1731,13 @@ p {
 
 h1 {
   font-size: 153%;
+  color: #585351;
 }
 
 h2, .rst-content .toctree-wrapper p.caption {
   font-size: 139%;
   font-weight: lighter;
+  color: #826c62;
 }
 
 h3 {
@@ -1854,7 +1856,7 @@ div[class^='highlight'] td.code {
 
 code, p.caption { 
   font-family: "NeoSansIntel-Light", sans, monospace;
-  color: #A79992;
+  color: #585351;
   font-size: 0.99em;
   line-height: 1.39em;
   }
@@ -1875,7 +1877,7 @@ caption-text {
     padding-top: 0.29em;
     padding-left: 0.11em;
     padding-bottom: 0.23em;
-    text-align: rig6ht; 
+    text-align: right; 
     }
 
 div[class^='highlight'] pre {
@@ -2771,7 +2773,7 @@ span[id*='MathJax-Span'] {
   padding: 0 1.618em;
   margin-bottom: 0;
   display: block;
-  font-family: NeoSansIntel, sans;
+  font-family: "NeoSansIntel", sans;
   font-weight: bold;
   text-transform: uppercase;
   font-size: 80%;
@@ -2925,11 +2927,12 @@ span[id*='MathJax-Span'] {
   padding: 0.809em;
   margin-bottom: 0.809em;
   z-index: 200;
-  background-color: #0071c5;
+  background-color: #fcfcfc;
   text-align: center;
   padding: 0.809em;
   display: block;
-  color: #fcfcfc;
+  color: #585351;
+  font-family: "NeoSansIntel-Regular", sans;
   margin-bottom: 0.809em;
 }
 .wy-side-nav-search input[type=text] {
@@ -2937,14 +2940,14 @@ span[id*='MathJax-Span'] {
     line-height: 1.3em;
   border-radius: 50px;
   padding: 6px 12px;
-  border-color: #8eb0af;
+  border-color: #aba09b;
 }
 .wy-side-nav-search img {
   display: block;
   margin: auto auto 0.809em auto;
   height: 45px;
   width: 45px;
-  background-color: #8eb0af;
+  background-color: #aba09b;
   padding: 5px;
   border-radius: 100%;
 }
@@ -2975,7 +2978,7 @@ span[id*='MathJax-Span'] {
   margin-top: -0.4045em;
   margin-bottom: 0.809em;
   font-weight: normal;
-  color: rgba(255, 255, 255, 0.3);
+  color: #3f3c3d;
 }
 
 .wy-nav .wy-menu-vertical header {
@@ -2985,7 +2988,7 @@ span[id*='MathJax-Span'] {
   color: #dadada;
 }
 .wy-nav .wy-menu-vertical a:hover {
-  background-color: #8eb0af;
+  background-color: #aba09b;
   color: #fff;
 }
 
@@ -3035,7 +3038,7 @@ span[id*='MathJax-Span'] {
   overflow-x: hidden;
   overflow-y: hidden;
   min-height: 100%;
-  background: #5f5f5f;
+  background: #585351;
   z-index: 200;
 }
 
@@ -3049,7 +3052,7 @@ span[id*='MathJax-Span'] {
 
 .wy-nav-top {
   display: none;
-  background: #0071c5;
+  background: #fcfcfc;
   color: #fff;
   padding: 0.4045em 0.809em;
   position: relative;
@@ -3066,14 +3069,14 @@ span[id*='MathJax-Span'] {
   clear: both;
 }
 .wy-nav-top a {
-  color: #fff;
-  font-weight: bold;
+  color: #585351;
+  font-family: "NeoSansIntel-Regular", sans;
 }
 .wy-nav-top img {
   margin-right: 12px;
   height: 45px;
   width: 45px;
-  background-color: #0071c5;
+  background-color: #fcfcfc;
   padding: 5px;
   border-radius: 100%;
 }
@@ -3092,15 +3095,15 @@ span[id*='MathJax-Span'] {
 .wy-nav-content {
   padding: 1.618em 3.236em;
   height: 100%;
-  max-width: 850px !important;
+  max-width: 979px !important;
   margin: auto;
 }
 
 .wy-body-mask {
   position: fixed;
-  width: 100%;
+  width: auto;
   height: 100%;
-  background: rgba(0, 0, 0, 0.2);
+  background: ;
   display: none;
   z-index: 499;
 }
@@ -3137,11 +3140,11 @@ footer span.commit code, footer span.commit .rst-content tt, .rst-content footer
 
 #search-results .search li {
   margin-bottom: 24px;
-  border-bottom: solid 1px #e1e4e5;
+  border-bottom: solid 1px #f3cdba;
   padding-bottom: 24px;
 }
 #search-results .search li:first-child {
-  border-top: solid 1px #e1e4e5;
+  border-top: solid 1px #f3cdba;
   padding-top: 24px;
 }
 #search-results .search li a {

doc/sphinx/ngraph_theme/static/theme.css

@@ -2554,9 +2554,10 @@ div[class^='highlight'] pre {
 }
 
 .function {
-    border-right: dashed 0.19em #f4f2f0;
-    border-bottom: solid 0.23em #f4f2f0;
-    border-top: solid 0.31em #77a29f;
+    border-right: dashed 0.19em #f3cdba;
+    border-bottom: solid 0.23em #826c62;
+    border-top: solid 0.31em #826c62;
+    background: #aba09b;
 }
 
 

doc/sphinx/ngraph_theme/theme.conf

@@ -9,4 +9,4 @@ sticky_navigation = True
 logo_only =
 collapse_navigation = False
 display_version = True
-use_bower = FALSE
+use_bower = False

doc/sphinx/source/backend-support/cpp-api.rst

+.. backend-support/cpp-api.rst:
+
+Runtime Backends 
+################
+
+.. figure:: ../graphics/backend-dgm.png
+   :width: 650px
+
+   Various backends are accessible via nGraph core APIs
+
+
+Backend
+=======
+
+.. doxygenclass:: ngraph::runtime::Backend
+   :project: ngraph
+   :members:
+
+
+
+Tensor
+======
+
+.. doxygenclass:: ngraph::runtime::Tensor
+   :project: ngraph
+   :members:
+
+
+
+HostTensor
+==========
+
+.. doxygenclass:: ngraph::runtime::HostTensor
+   :project: ngraph
+   :members:
+
+
+
+PlaidML
+=======
+
+.. doxygenclass:: ngraph::runtime::plaidml::PlaidML_Backend
+   :project: ngraph
+   :members:
\ No newline at end of file

doc/sphinx/source/programmable/index.rst → doc/sphinx/source/backend-support/index.rst

-.. index.rst
+.. backend-support/index.rst
 
 
-#######################
-Interact with Backends 
-#######################
+Transformer, CPU, GPU, PlaidML
+###############################
 
-Backend
-========
+
+* :ref:`hybrid_transformer`
+* :ref:`cpu_backend`
+* :ref:`plaidml_backend`
+* :ref:`gpu_backend`
+
+
+
+
+What is a backend?
+------------------
 
 Backends are responsible for function execution and value allocation. They 
 can be used to :doc:`carry out a programmed computation<../howto/execute>`
@@ -14,32 +22,38 @@ from a framework by using a CPU or GPU; or they can be used with an *Interpreter
 mode, which is primarily intended for testing, to analyze a program, or for a 
 framework developer to develop customizations. Experimental APIs to support 
 current and future nGraph Backends are also available; see, for example, the 
-section on :ref:`plaidml_`.
+section on :ref:`plaidml_backend`.
+
 
 
-.. figure:: ../graphics/backend-dgm.png
-   :width: 650px
+.. _hybrid_transformer:
 
+Hybrid Transformer
+==================
 
-.. doxygenclass:: ngraph::runtime::Backend
-   :project: ngraph
-   :members:
+Lorem ipsum
 
 
+.. _cpu_backend:
 
-Tensor
-=======
+CPU Backend
+===========
 
-.. doxygenclass:: ngraph::runtime::Tensor
-   :project: ngraph
-   :members:
+Lorem ipsum
 
 
+.. _gpu_backend:
 
-.. _plaidml_:
+GPU Backend
+===========
 
-PlaidML
-========
+Lorem ipsum
+
+
+.. _plaidml_backend:
+
+PlaidML Backend
+===============
 
 The nGraph ecosystem has recently added initial (experimental) support for `PlaidML`_,
 which is an advanced :abbr:`Machine Learning (ML)` library that can further
@@ -47,10 +61,4 @@ accelerate training models built on GPUs. When you select the ``PlaidML`` option
 as a backend, it behaves as an advanced tensor compiler that can further speed up
 training with large data sets.
 
-.. doxygenclass:: ngraph::runtime::plaidml::PlaidML_Backend
-   :project: ngraph
-   :members:
-
-
-
 .. _PlaidML: https://github.com/plaidml

doc/sphinx/source/buildlb.rst

@@ -220,12 +220,13 @@ paths for what you'll want to do next: either compile a framework to run a DL
 training model, or load an import of an "already-trained" model for inference 
 on an Intel nGraph-enabled backend.
 
-For the former case, this early |version|, :doc:`framework-integration-guides`, 
+For the former case, this early |version|, :doc:`frameworks/index`, 
 can help you get started with a training a model on a supported framework. 
 
-* :doc:`MXNet<framework-integration-guides>` framework,  
-* :doc:`TensorFlow<framework-integration-guides>` framework, and
-* :doc:`neon<framework-integration-guides>` framework,  
+* :doc:`MXNet<frameworks/tensorflow_integ>` framework,  
+* :doc:`TensorFlow<frameworks/mxnet_integ>` framework,
+* :doc:`ONNX & ONNXIFI<frameworks/onnx_integ>`, and
+* :doc:`PaddlePaddle<frameworks/paddle_integ>` framework.
 
 
 For the latter case, if you've followed a tutorial from `ONNX`_, and you have an 

doc/sphinx/source/conf.py

@@ -73,11 +73,11 @@ author = 'Intel Corporation'
 # built documents.
 #
 # The short X.Y version.
-version = '0.12'
+version = '0.13'
 
 # The Documentation full version, including alpha/beta/rc tags. Some features
 # available in the latest code will not necessarily be documented first
-release = '0.12.1'
+release = '0.13-0'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.

doc/sphinx/source/core/overview.rst

+.. core/overview.rst:
+
+
+Overview
+========
+

doc/sphinx/source/diagnostics/debug.rst

+.. debug:
+
+
+Debugging 
+#########
+
+
+Error hints
+===========
+
+
+Common mistakes
+===============
+
+
+Wiki, other help
+================
+
+
+
+

doc/sphinx/source/diagnostics/nbench.rst

+.. nbench:
+
+###################
+Working with nBench
+###################
+
+
+
+
+
+
+
+
+
+
+
+
+
+

doc/sphinx/source/diagnostics/performance-profile.rst

+.. performance-profile:
+
+Performance profiling
+#####################
+
+
+.. FW-specific validation performance comparisons via nbench
\ No newline at end of file

doc/sphinx/source/diagnostics/visualize.rst

+.. visualize: 
+
+Visualization
+#############
+

doc/sphinx/source/framework-integration-guides.rst

-.. framework-integration-guides:
-
-###############################
-Integrate Supported Frameworks
-###############################
-
-* :ref:`mxnet_intg`
-* :ref:`tensorflow_intg`
-
-A framework is "supported" when there is a framework :term:`bridge` that can be 
-cloned from one of our GitHub repos and built to connect to nGraph device backends, 
-all the while maintaining the framework's programmatic or user interface. Bridges 
-currently exist for the TensorFlow\* and MXNet\* frameworks. 
-
-.. figure:: graphics/bridge-to-graph-compiler.png
-    :width: 733px
-    :alt: JiT compiling of a computation
-
-    :abbr:`Just-in-Time (JiT)` Compiling for computation
-
-Once connected via the bridge, the framework can then run and train a deep 
-learning model with various workloads on various backends using nGraph Compiler 
-as an optimizing compiler available through the framework.  
-
-
-.. _mxnet_intg:
-
-MXNet\* bridge
-===============
-
-* See the README on `nGraph-MXNet`_ Integration.
-
-* **Testing latency for Inference**:  See the :doc:`frameworks/testing-latency` 
-  doc for a fully-documented example how to compile and test latency with an 
-  MXNet-supported model.     
-
-* **Training**: For experimental or alternative approaches to distributed training
-  methodologies, including data parallel training, see the MXNet-relevant sections
-  of the docs on :doc:`distr/index` and :doc:`How to <howto/index>` topics like
-  :doc:`howto/distribute-train`. 
-
-
-.. _tensorflow_intg:
-
-TensorFlow\* bridge
-===================
-
-See the `ngraph tensorflow bridge README`_ for how to install the `DSO`_ for the 
-nGraph-TensorFlow bridge.
-
-
-
-.. _nGraph-MXNet: https://github.com/NervanaSystems/ngraph-mxnet/blob/master/README.md
-.. _MXNet: http://mxnet.incubator.apache.org
-.. _DSO: http://csweb.cs.wfu.edu/%7Etorgerse/Kokua/More_SGI/007-2360-010/sgi_html/ch03.html
-.. _being the fastest: https://github.com/soumith/convnet-benchmarks
-.. _ngraph tensorflow bridge README: https://github.com/NervanaSystems/ngraph-tf/blob/master/README.md

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

 .. frameworks/generic-configs.rst:
 
-
-Configurations available to any framework
-#########################################
-
-
-Enabling Deep Learning paradigms  
-================================
-
-Framework architects or engineers who can't quite find what they need among 
-the existing DL tools may need to build something new off a "stock" framework, 
-or someting entirely from scratch. For this category of developer, we have 
-:doc:`documented several ways <../howto/index>` you can incorporate built-in 
-compiler support for users of your framework; this includes out-of-box support 
-for things like Intel® MKL-DNN and PlaidML when your framework supports nGraph 
-as a "backend" or engine. 
-
-   .. important:: nGraph does not provide an interface for "users" of frameworks 
-      (for example, we cannot dictate or control how Tensorflow* or MXNet* presents 
-      interfaces to users). Please keep in mind that designing and documenting 
-      the :abbr:`User Interface (UI)` of step 3 above is entirely in the realm 
-      of the framework owner or developer and beyond the scope of the nGraph 
-      Compiler stack. However, any framework can be designed to make direct use 
-      of nGraph Compiler stack-based features and then expose an accompanying UI, 
-      output message, or other detail to a user.
- 
-The nGraph :abbr:`IR Intermediate Representation` is format that can understand 
-inputs from a framework. Today, there are two primary tasks that can be accomplished 
-in the “bridge code” space of the nGraph IR: 
-
-#. Compiling a dataflow graph 
-#. Executing a pre-compiled graph. 
-
-See the :doc:`../framework-integration-guides` for how we built bridges with our 
-initially-supported frameworks. For more in-depth help in writing things like 
-graph optimizations and bridge code, we provide articles on how to 
-:doc:`../fusion/index`, and programmatically :doc:`../howto/execute` that can 
-target various compute resources using nGraph when a framework provides some 
-inputs to be computed.
-
-.. note:: Configuration options can be added manually on the command line or via 
-   scripting. Please keep in mind that fine-tuning of parameters is as much of 
-   an art as it is a science; there are virtually limitless ways to do so and 
-   our documentation provides only a sampling.  
-
-
-Integrating nGraph with new frameworks
-======================================
+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 
@@ -78,7 +33,7 @@ something like:
 
 
 FMV
-===
+---
 
 FMV stands for :abbr:`Function Multi-Versioning`, and it can also provide a 
 number of generic ways to patch or bring architecture-based optimizations to 
@@ -90,9 +45,8 @@ for Intel® Architecture, and it includes at least one older CPU, the
 `following article may be helpful`_.
 
 
-
 Training Deep Neural Networks
-==============================
+-----------------------------
 
 Before tweaking various environment variables, be aware that how the computation 
 gets executed depends upon the ordering of the data format that the model is 
@@ -108,6 +62,7 @@ For CPU (and most cuDNN) backends, the preferred layout is currently ``NCHW``.
 * **H** -- Height of the image
 * **W** -- Width of the image
 
+
 Intel® Math Kernel Library for Deep Neural Networks 
 ---------------------------------------------------
 
@@ -127,7 +82,7 @@ additional component to be able to use these configuration settings.
 
 
 nGraph-enabled Intel® Xeon® 
-============================
+---------------------------
 
 The list below includes recommendations on data layout, parameters, and 
 application configuration to achieve best performance running DNN workloads on 

doc/sphinx/source/frameworks/index.rst

 .. frameworks/index.rst: 
 
-#####################
-Connecting Frameworks   
-#####################
+
+.. TODO update CODEOWNERS for this new structure
+
+
+Current framework integrations  
+==============================
+
+.. toctree::
+   :maxdepth: 1
+
+   tensorflow_integ.rst
+   mxnet_integ.rst
+   onnx_integ.rst
+   paddle_integ.rst
+   validated/testing-latency.rst
+
+
+A framework is "supported" when there is a framework :term:`bridge` that can be 
+cloned from one of our GitHub repos and built to connect to nGraph device backends, 
+all the while maintaining the framework's programmatic or user interface. Bridges 
+currently exist for the TensorFlow\* and MXNet\* frameworks. 
+
+.. figure:: ../graphics/bridge-to-graph-compiler.png
+    :width: 733px
+    :alt: JiT compiling of a computation
+
+    :abbr:`Just-in-Time (JiT)` Compiling for computation
+
+Once connected via the bridge, the framework can then run and train a deep 
+learning model with various workloads on various backends using nGraph Compiler 
+as an optimizing compiler available through the framework.  
 
 While a :abbr:`Deep Learning (DL)` :term:`framework` is ultimately meant for 
 end use by data scientists, or for deployment in cloud container environments, 
 nGraph Core ops and the nGraph C++ Library are designed for framework builders 
 themselves. We invite anyone working on new and novel frameworks or neural 
-network designs to explore our highly-modular stack of components that can 
+network designs to explore our highly-modularized stack of components that can 
 be implemented or integrated in virtually limitless ways.
 
 Please read the articles in this section if you are considering incorporating 
@@ -17,61 +45,10 @@ design. Articles here are also useful if you are working on something
 built-from-scratch, or on an existing framework that is less widely-supported 
 than the popular frameworks like TensorFlow and PyTorch. 
 
-.. toctree::
-   :maxdepth: 1 
-
-   generic-configs.rst
-   testing-latency.rst
-   validation.rst
-
-
-Understanding users of frameworks
-=================================
-
-A data scientist or ML engineer may not initially know which framework is the 
-"best" framework to use to start working on his or her problem set. While there 
-are several to choose from, it can be daunting and time consuming to scope the 
-wide array of features and customization options offered by some of the more 
-popular frameworks:
-
-#. First **find** a tested and working DL model that does something *similar* 
-   to what the data scientist or ML engineer wants to do. To assist with this 
-   stage, we've already provided organized tables of :doc:`validation` examples.
-#. Next, **replicate** that result using well-known datasets to confirm that the 
-   model does indeed work. To assist with this stage, we've released several  
-   :doc:`pip installation options <../framework-integration-guides>` that can 
-   be used to test basic examples.
-#. Finally, **modify** some aspect: add new datasets, or adjust an algorithm's 
-   parameters to hone in on specifics that can better train, forecast, or predict 
-   scenarios modeling the real-world problem. This is also the stage where it 
-   makes sense to `tune the workload to extract best performance`_.
-
-   .. important:: nGraph does not provide an interface for "users" of frameworks 
-      (for example, we cannot dictate or control how Tensorflow* or MXNet* presents 
-      interfaces to users). Please keep in mind that designing and documenting 
-      the :abbr:`User Interface (UI)` is entirely in the realm of the framework owner 
-      or developer and beyond the scope of the nGraph Compiler stack. However, any 
-      framework can be designed to make direct use of nGraph Compiler stack-based 
-      features and then expose an accompanying UI, output message, or other detail 
-      to a user.
-
-Clearly, one challenge of the framework developer is to differentiate from 
-the pack by providing a means for the data scientist to obtain reproducible 
-results. The other challenge is to provide sufficient documentation, or to 
-provide sufficient hints for how to do any "fine-tuning" for specific use cases. 
-With the nGraph Compiler stack powering your framework, it becomes much easier 
-to help your users get reproducible results with nothing more complex than the 
-CPU that powers their operating system.   
-
-In general, the larger and more complex a framework is, the harder it becomes 
-to navigate and extract the best performance; configuration options that are 
-enabled by "default" from the framework side can sometimes slow down compilation 
-without the developer being any the wiser. Sometimes only `a few small`_ 
-adjustments can increase performance. Likewise, a minimalistic framework that 
-is designed around one specific kind of model can sometimes offer significant 
-performance-improvement opportunities by lowering overhead. 
-
-See :doc:`generic-configs` to get started.   
+.. figure:: ../graphics/translation-flow-to-ng-fofx.png
+   :width: 725px
+   :alt: Translation flow to nGraph function graph
+
 
 
 .. _tune the workload to extract best performance: https://ai.intel.com/accelerating-deep-learning-training-inference-system-level-optimizations

doc/sphinx/source/frameworks/mxnet_integ.rst

+.. mxnet_integ.rst:
+
+MXNet\* bridge
+===============
+
+* See the `README`_ on nGraph-MXNet repo.
+
+* **Testing inference latency**:  See the :doc:`validated/testing-latency` 
+  doc for a fully-documented example how to compile and test latency with an 
+  MXNet-supported model.     
+
+* **Training**: For experimental or alternative approaches to distributed 
+  training methodologies, including data parallel training, see the 
+  MXNet-relevant sections of the docs on :doc:`../distr/index` and 
+  :doc:`How to <../howto/index>` topics like :doc:`../howto/distribute-train`. 
+
+
+.. _README: https://github.com/NervanaSystems/ngraph-mxnet/blob/master/README.md
\ No newline at end of file

doc/sphinx/source/frameworks/onnx_integ.rst

+.. onnx_integ.rst:
+
+
+ONNX & ONNXIFI
+==============
\ No newline at end of file

doc/sphinx/source/frameworks/paddle_integ.rst

+.. paddle_integ.rst:
+
+PaddlePaddle
+============
+

doc/sphinx/source/frameworks/tensorflow_integ.rst

+.. tensorflow_integ.rst:
+
+TensorFlow\* bridge
+===================
+
+See the `ngraph tensorflow bridge README`_ for how to install the `DSO`_ for the 
+nGraph-TensorFlow bridge.
+
+
+.. _DSO: http://csweb.cs.wfu.edu/%7Etorgerse/Kokua/More_SGI/007-2360-010/sgi_html/ch03.html
+.. _ngraph tensorflow bridge README: https://github.com/NervanaSystems/ngraph-tf/blob/master/README.md

doc/sphinx/source/frameworks/validation.rst → doc/sphinx/source/frameworks/validated/list.rst

-.. frameworks/validation.rst: 
+.. frameworks/validated/list.rst: 
 
-##############################
-Validated Models and Workloads
-##############################
+#################################
+Validated workloads by framework
+#################################
 
-We validated performance [#f1]_ for the following TensorFlow\* and MXNet\* workloads:
+We validated performance [#f1]_ for the following TensorFlow\* and MXNet\* 
+workloads:
 
+* :ref:`tensorflow_valid`
+* :ref:`mxnet_valid`
+* :ref:`onnx_valid`
+* :doc:`testing-latency`
+
+
+
+
+.. _tensorflow_valid:
 
 TensorFlow 
 ==========
@@ -36,6 +46,9 @@ TensorFlow
    DRAW, Image generation
    A3C, Reinforcement learning
 
+
+.. _mxnet_valid:
+
 MXNet
 =====
 
@@ -63,8 +76,11 @@ MXNet
    DCGAN, Generative adversarial network
    A3C, Reinforcement learning
 
+
+.. _onnx_valid:
+
 ONNX
-=====
+====
 
 Additionally, we validated the following workloads are functional through 
 `nGraph ONNX importer`_: 

doc/sphinx/source/frameworks/testing-latency.rst → doc/sphinx/source/frameworks/validated/testing-latency.rst

-.. frameworks/testing_latency:
+.. frameworks/validated/testing-latency.rst:
 
 
 Testing latency  
@@ -23,7 +23,7 @@ reasons outlined in our `features`_ documentation.
    the experimental backend if you already use the ngraph-mxnet Github repo 
 
 
-.. figure:: ../graphics/ngraph-mxnet-models.png
+.. figure:: ../../graphics/ngraph-mxnet-models.png
     :width: 533px
     :alt: Up to 45X faster 
 
@@ -85,14 +85,14 @@ into a static graph. Also note that any model with a saved checkpoint can be
 considered a "static graph" in nGraph. For this example, we'll presume that the 
 model is pre-trained.   
 
-.. literalinclude:: ../../../examples/subgraph_snippets/mxnet-gluon-example.py
+.. literalinclude:: ../../../../examples/subgraph_snippets/mxnet-gluon-example.py
    :language: python
    :lines: 17-32
 
 
 To load the model into nGraph, we simply bind the symbol into an Executor. 
 
-.. literalinclude:: ../../../examples/subgraph_snippets/mxnet-gluon-example.py
+.. literalinclude:: ../../../../examples/subgraph_snippets/mxnet-gluon-example.py
    :language: python
    :lines: 34-35
 
@@ -101,7 +101,7 @@ the graph, and in the case of Resnet, sends the entire graph to nGraph for
 compilation. This produces a single call to an NNVM ``NGraphSubgraphOp`` embedded 
 with the compiled model. At this point, we can test the model's performance.
 
-.. literalinclude:: ../../../examples/subgraph_snippets/mxnet-gluon-example.py
+.. literalinclude:: ../../../../examples/subgraph_snippets/mxnet-gluon-example.py
    :language: python
    :lines: 40-48
 

doc/sphinx/source/fusion/index.rst

-.. fusion/index.rst: 
-
-
-Optimize Graphs 
-===============
-
-with nGraph Compiler fusions
-----------------------------
-
-The nGraph Compiler is an optimizing compiler. As such, it provides a way to 
-capture a given :term:`function graph` and perform a series of optimization 
-passes over that graph. The result is a semantically-equivalent graph that, when 
-executed using any |InG| :doc:`backend <../programmable/index>`, has optimizations 
-inherent at the hardware level: superior runtime characteristics to increase 
-training performance or reduce inference latency.   
-
-There are several ways to describe what happens when we capture and translate 
-the framework's output of ops into an nGraph graph. :term:`Fusion` is the term 
-we shall use in our documentation; the action also can be described as: 
-*combining*, *folding*, *squashing*, *collapsing*, or *merging* of graph 
-functions. 
-
-Optimization passes may include algebraic simplifications, domain-specific 
-simplifications, and fusion. Most passes share the same mode of operation (or 
-the same operational structure) and consist of various stages (each one a 
-:term:`step`) where a developer can experiment with the intercepted or dynamic 
-graph. These steps may be cycled or recycled as needed: 
-
-#. Locate a list of potentially-transformable subgraphs in the given graph.
-#. Transform the selected candidates into semantically-equivalent subgraphs 
-   that execute faster, or with less memory (or both). 
-#. Verify that the optimization pass performs correctly, with any or all expected 
-   transformations, with the ``NGRAPH_SERIALIZE_TRACING`` option, which 
-   serializes a graph in the `json` format after a pass.
-#. Measure and evaluate your performance improvements with ``NGRAPH_CPU_TRACING``, 
-   which produces timelines compatible with ``chrome://tracing``.
-
-Optimizations can be experimented upon without using any backend by registering 
-a pass with pass manager (``Manager``), calling ``run_passes`` on a function, and 
-then inspecting the transformed graph. 
-
-Optimization passes can be programmed ahead of time if you know or can predict 
-what your graph will look like when it's ready to be executed (in other words: 
-which `ops` can be automatically translated into :doc:`nGraph Core ops <../ops/index>`). 
-
-The ``Interpreter`` is simply a backend providing reference implementations of 
-ngraph ops in C++, with the focus on simplicity over performance.
- 
-
-Example
--------
+.. fusion/index.rst:
+
+Pattern matcher
+###############
+
+* :ref:`overview` 
+* :ref:`passes_list`
+* :ref:`more_detail` 
+* :ref:`passes_examples`
+* :doc:`optimize-graphs` 
+
+
+.. _overview:
+
+Generic graph optimizers: Optimization passes
+=============================================
+
+The pass manager infrastructure in nGraph makes it easy to reuse and mix the 
+generic optimization passes. It also permits you to roll your own device-specific 
+optimizations; that is, the same unified interface and APIs may be used to 
+cover both things.
+
+Invoking these passes is fairly straightforward:  
+
+#. Create a "pass manager" object. 
+#. Populate it with the desired passes. 
+#. Pass to it a pointer to your unoptimized graph, and it’ll return a pointer 
+   to an optimized graph.
+
+nGraph Core includes a large library of hardware-agnostic passes -- passes useful 
+for almost any kind of hardware backend. Some of these passes should be familiar 
+to people who are comfortable with classical compiler designs. Others, like the 
+reshape/transpose elimination and sinking passes, are quite specific to deep 
+learning.
+
+Let’s take a look at some of these passes.
+
+
+.. _passes_list:
+
+List of Passes
+==============
+
+* :ref:`algebraic_simpl`
+* :ref:`common_subex_elim`
+* :ref:`constant_fold`
+* :ref:`reshape_transpose_elim`
+* :ref:`reshape_transpose_sink`
+
+
+.. _algebraic_simpl: 
+
+Algebraic Simplification
+------------------------
+
+The **Algebraic Simplification** pass implements what amounts to a "grab bag" of 
+algebraic simplification rules. It does some basic things like rewrite "zero 
+times x" to simply "zero", or "zero plus x" to plain "x".
+
+It can also do a number of tricks more specific to deep learning. For example,
+if we discover that a tensor is being sliced up by adjacent segments, only to 
+have those slices concatenated back together again, we can skip the slicing and 
+concatting altogether. 
+
+Or, if a tensor is being padded, but the actual width of the padding is zero 
+all around, we can skip the padding step entirely.
+
+Several other transformations like this are implemented in the algebraic 
+simplification pass. And while none of these transformations might seem 
+particularly impressive on their own, when everything comes together the 
+results of this pass often yield improvement even on the initial graph straight 
+out of the bridge. This pass is also quite important as a "glue" pass that can 
+be used to clean up and/or re-simplify after other passes have done their own 
+tricks.
+
+
+.. _common_subex_elim: 
+
+Common Subexpression Elimination
+--------------------------------
+
+
+.. _constant_fold:
+
+Constant Folding
+----------------
+
+
+.. _core_fusion:
+
+Core Fusion
+-----------
+
+
+.. _reshape_transpose_elim:
+
+Reshape/Transpose Elimination
+-----------------------------
+
+The pass called **Reshape/Transpose Elimination** will find and optimize where 
+we can "push" two ``Transpose`` ops through a matrix multiplication. For example, 
+if you have two matrices (say, *foo* and *bar*), both of these matrices will be 
+transposed (to produce *foo.t* and *bar.t*, respectively), aftew which *foo.t* 
+and *bar.t* get multiplied together.
+
+Often a more efficient way to implement this is to switch the order of the 
+arguments *foo* and *bar*, multiply them together, and then transpose the output 
+of the matmul. Effectively, this cuts two `Transpose` operations down to just 
+one, where the **Reshape/Transpose** elimination will do that rewrite for you.
+
+Another common pattern can be optimized via nGraph is the case where two 
+transpositions cancel each other out. One example of this is taking the 
+"Transpose" of the transpose of a matrix, though actually a more common case is 
+when the graph is translating among different batch formats. We can often move 
+these operations around through a process called **Reshape sinking/swimming**, 
+and in cases where two transposes wind up canceling each other out, we can cut 
+them both out of the graph.
+
+
+
+.. _reshape_transpose_sink:
+
+``Reshape/Transpose Sinking``
+-----------------------------
+
+
+
+
+
+.. _elementzero_tensor_elim:
+
+``Zero-Element Tensor Elimination``
+-----------------------------------
+
+
+
+
+
+.. _more_detail:
+
+More detail
+-----------
 
 Let us first consider a simple example. A user would like to execute a graph 
 that describes the following arithmetic expression:
@@ -80,4 +171,35 @@ optimizer.
 
    graph-rewrite.rst
    passes-that-use-matcher.rst
+   optimize-graphs.rst
    
+
+.. _passes_examples:
+
+Examples of Passes
+==================
+
+The effectiveness of these passes is more striking to look at in terms of an 
+actual input graph, such as one from the framework bridge.
+
+*Figure 0* shows an excerpt from ``MobileNet v1``, a topology which makes heavy 
+use of group convolution.
+
+.. _figure-mobilenet-gc:
+
+.. figure:: ../graphics/mobilenet-group-conv.png
+   :width: 700px
+   :alt: 
+
+   Figure 0: Each of these grouped convolution complexes -- the 
+   operations within the rectangles on the left -- is very wide; each is too 
+   wide to fit legibly on the illustration.
+
+
+The group convolution fusion is able to replace each of those giant subgraphs 
+with a single CPU group convolution node. This ends up being a win in several 
+ways: 
+
+* sheer node count, 
+* mappability to MKL-DNN (which has an accelerated group convolution implementation), 
+* elimination of unnecessary temporaries, and so on.
\ No newline at end of file

doc/sphinx/source/fusion/mg/fusion_pattern.dot

-digraph G {
-    Label_0 -> Max_2
-    Constant_1 -> Max_2
-    Label_0 [shape=ellipse color=black]
-    Constant_1 [shape=ellipse color=black]
-    Max_2 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_graph1.dot

-digraph G {
-    Parameter_0 -> Negative_1 -> Negative_2;
-    Parameter_0 [shape=box color=blue]
-    Negative_1 [shape=ellipse color=black]
-    Negative_2 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_graph2.dot

-digraph G {
-    Parameter_0 -> Abs_1 -> Negative_2 -> Negative_3;
-    Parameter_0 [shape=box color=blue]
-    Abs_1 [shape=ellipse color=black]
-    Negative_2 [shape=ellipse color=black]
-    Negative_3 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_graph3.dot

-digraph G {
-    Parameter_0 -> Add_2
-    Parameter_1 -> Add_2
-    Add_2 -> Abs_3 -> Negative_4 -> Negative_5
-    Parameter_0 [shape=box color=blue]
-    Parameter_1 [shape=box color=blue]
-    Add_2 [shape=ellipse color=black]
-    Abs_3 [shape=ellipse color=black]
-    Negative_4 [shape=ellipse color=black]
-    Negative_5 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_graph4.dot

-digraph G {
-    Parameter_0 -> Add_2
-    Parameter_1 -> Add_2
-    Add_2 -> Negative_3 -> Negative_4
-    Parameter_0 [shape=box color=blue]
-    Parameter_1 [shape=box color=blue]
-    Add_2 [shape=ellipse color=black]
-    Negative_3 [shape=ellipse color=black]
-    Negative_4 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_graph5.dot

-digraph G {
-    Parameter_0 -> Sub_2
-    Parameter_1 -> Sub_2
-    Sub_2 -> Negative_3 -> Negative_4
-    Parameter_0 [shape=box color=blue]
-    Parameter_1 [shape=box color=blue]
-    Sub_2 [shape=ellipse color=black]
-    Negative_3 [shape=ellipse color=black]
-    Negative_4 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_pattern.dot

-digraph G {
-    Parameter_1 -> Negative_2 -> Negative_3;
-    Parameter_1 [shape=box color=blue]
-    Negative_2 [shape=ellipse color=black]
-    Negative_3 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr1_pattern2.dot

-digraph G {
-    Label_0 -> Negative_1 -> Negative_2;
-    Label_0 [shape=ellipse color=black]
-    Negative_1 [shape=ellipse color=black]
-    Negative_2 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr2_graph1.dot

-digraph G {
-    Parameter_0 -> Add_2
-    Constant_1 -> Add_2
-    Parameter_0 [shape=box color=blue]
-    Constant_1 [shape=ellipse color=black]
-    Add_2 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr2_graph2.dot

-digraph G {
-    Parameter_0 -> Add_3
-    Constant_1 -> Broadcast_2
-    Broadcast_2 -> Add_3
-    Parameter_0 [shape=box color=blue]
-    Constant_1 [shape=ellipse color=black]
-    Broadcast_2 [shape=ellipse color=black]
-    Add_3 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr2_graph3.dot

-digraph G {
-    Parameter_0 -> Add_2
-    Constant_1 -> Broadcast_2
-    Constant_1 -> Add_3
-    Parameter_0 [shape=box color=blue]
-    Constant_1 [shape=ellipse color=black]
-    Broadcast_2 [shape=ellipse color=black]
-    Add_3 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/pr2_pattern2.dot

-digraph G {
-    Constant_1 -> Skip_2
-    Label_3 -> Add_4
-    Skip_2 -> Add_4
-    Constant_1 [shape=ellipse color=black]
-    Skip_2 [shape=ellipse color=black]
-    Label_3 [shape=ellipse color=black]
-    Add_4 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/rp_graph1.dot

-digraph G {
-    Parameter_0 -> Add_2
-    Constant_1 -> Add_2
-    Add_2 -> Add_3
-    Constant_2 -> Add_3
-    Add_3 -> Add_4
-    Constant_3 -> Add_4
-    Parameter_0 [shape=box color=blue]
-    Constant_1 [shape=ellipse color=black]
-    Constant_2 [shape=ellipse color=black]
-    Constant_3 [shape=ellipse color=black]
-    Add_2 [shape=ellipse color=black]
-    Add_3 [shape=ellipse color=black]
-    Add_4 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/mg/rp_pattern.dot

-digraph G {
-    Label_0 -> Add_2
-    Constant_1 -> Add_2
-    Label_0 [shape=ellipse color=black]
-    Constant_1 [shape=ellipse color=black]
-    Add_2 [shape=ellipse color=black]
-}
-
-

doc/sphinx/source/fusion/optimize-graphs.rst

+.. fusion/optimize-graphs: 
+
+
+Optimize Graphs 
+===============
+
+with nGraph Compiler fusions
+----------------------------
+
+The nGraph Compiler is an optimizing compiler. As such, it provides a way to 
+capture a given :term:`function graph` and perform a series of optimization 
+passes over that graph. The result is a semantically-equivalent graph that, when 
+executed using any :doc:`backend <../backend-support/index>`, has optimizations 
+inherent at the hardware level: superior runtime characteristics to increase 
+training performance or reduce inference latency.   
+
+There are several ways to describe what happens when we capture and translate 
+the framework's output of ops into an nGraph graph. :term:`Fusion` is the term 
+we shall use in our documentation; the action also can be described as: 
+*combining*, *folding*, *squashing*, *collapsing*, or *merging* of graph 
+functions. 
+
+Optimization passes may include algebraic simplifications, domain-specific 
+simplifications, and fusion. Most passes share the same mode of operation (or 
+the same operational structure) and consist of various stages (each one a 
+:term:`step`) where a developer can experiment with the intercepted or dynamic 
+graph. These steps may be cycled or recycled as needed: 
+
+#. Locate a list of potentially-transformable subgraphs in the given graph.
+#. Transform the selected candidates into semantically-equivalent subgraphs 
+   that execute faster, or with less memory (or both). 
+#. Verify that the optimization pass performs correctly, with any or all expected 
+   transformations, with the ``NGRAPH_SERIALIZE_TRACING`` option, which 
+   serializes a graph in the `json` format after a pass.
+#. Measure and evaluate your performance improvements with ``NGRAPH_CPU_TRACING``, 
+   which produces timelines compatible with ``chrome://tracing``.
+
+Optimizations can be experimented upon without using any backend by registering 
+a pass with pass manager (``Manager``), calling ``run_passes`` on a function, and 
+then inspecting the transformed graph. 
+
+Optimization passes can be programmed ahead of time if you know or can predict 
+what your graph will look like when it's ready to be executed (in other words: 
+which `ops` can be automatically translated into :doc:`nGraph Core ops <../ops/index>`). 
+
+The ``Interpreter`` is simply a backend providing reference implementations of 
+ngraph ops in C++, with the focus on simplicity over performance.

doc/sphinx/source/glossary.rst

@@ -41,6 +41,25 @@ Glossary
       of a graph's functional operations (``ops``) into one or more of
       nGraph's core ops.   
 
+   ISA 
+
+      An acronym for "Instruction Set Architecture," an ISA is machine code that  
+      is compatible with the underlying silicon architecture. A realization of 
+      an ISA is called an *implementation*. An ISA permits multiple 
+      implementations that may vary in performance, physical size, memory use or 
+      reuse, and monetary cost among other things. An ISA defines everything a 
+      machine-language programmer needs to know in order to program a particular 
+      backend device. What an ISA defines will differ among ISAs; in general, it
+      defines things like:
+
+         - supported *data types*; 
+         - physical *states* available, such as the main memory and registers; 
+         - *semantics*, such as the memory consistency and addressing modes; 
+         - *low-level machine instructions* that comprise a machine language; 
+         - and the *input/output model*.
+
+      Be careful to not confuse ISAs with microarchitectures.    
+
    op
 
       An op represents an operation. Ops are stateless and have zero

doc/sphinx/source/howto/derive-for-training.rst

@@ -137,7 +137,7 @@ update computation for ``N`` will be given by the node
 
 The different update nodes will share intermediate computations. So to
 get the updated values for the weights as computed with the specified 
-:doc:`backend <../programmable/index>`, 
+:doc:`backend <../backend-support/index>`.
 
 .. literalinclude:: ../../../examples/mnist_mlp/mnist_mlp.cpp
    :language: cpp

doc/sphinx/source/howto/distribute-train.rst

@@ -6,11 +6,11 @@ Train using multiple nGraph CPU backends with data parallel
 
 In the :doc:`previous section <../howto/derive-for-training>`, we described the 
 steps needed to create a "trainable" nGraph model. Here we demonstrate how to 
-train a data parallel model by distributing the graph across devices.
+train a data parallel model by distributing the graph to more than one device.
 
+As of release version 0.12, the default build is with OpenMPI.  To use the 
+`Intel MLSL`_ library, set the following compilation flag at build time: 
 
-
-To use this mode of training, create an nGraph build with the cmake flag 
 ``-DNGRAPH_DISTRIBUTED_ENABLE=TRUE``.
 
 To deploy data-parallel training on backends supported by nGraph API, the 
@@ -20,14 +20,13 @@ To deploy data-parallel training on backends supported by nGraph API, the
 .. literalinclude:: ../../../examples/mnist_mlp/dist_mnist_mlp.cpp
    :language: cpp
    :lines: 180-196
-   :emphasize-lines: 9-12
+   :emphasize-lines: 8-11
 
 We need to initialize and finalize distributed training with ``Distributed`` object;
 see the `full raw code`_. 
 
-Finally, to run the training using two nGraph devices, invoke :command:`mpirun` which is a distributed with 
-`Intel MLSL`_ library.
-This will launch two nGraph CPU backends.
+Finally, to run the training using two nGraph devices, invoke :command:`mpirun` which 
+is distributed with `Intel MLSL`_ library.  This will launch two nGraph CPU backends.
 
 
 .. code-block:: console 

doc/sphinx/source/howto/execute.rst

@@ -18,9 +18,9 @@ directory. We'll be deconstructing the steps that must happen (either programmat
 or manually) in order to successfully execute a computation:
 
 * :ref:`define_cmp`
-* :ref:`specify_bkd`
+* :ref:`specify_backend`
 * :ref:`compile_cmp`
-* :ref:`allocate_bkd_storage`
+* :ref:`allocate_backend_storage`
 * :ref:`initialize_inputs`
 * :ref:`invoke_cmp`
 * :ref:`access_outputs`
@@ -108,7 +108,7 @@ function, in the order they are to be passed to the compiled function. A
    the computation of the results.
 
 
-.. _specify_bkd:
+.. _specify_backend:
 
 Specify the backend upon which to run the computation
 =====================================================
@@ -126,10 +126,10 @@ There are two backends for the CPU: the optimized ``"CPU"`` backend, which uses
 the `Intel MKL-DNN`_, and the ``"INTERPRETER"`` backend, which runs reference 
 versions of kernels that favor implementation clarity over speed. The 
 ``"INTERPRETER"`` backend can be slow, and is primarily intended for testing. 
-See the documentation on :doc:`runtime options for various backends <../programmable/index>` 
+See the documentation on :doc:`runtime options for various backends <../backend-support/index>` 
 for additional details.
 
-To continue with our original example and select the ``"CPU"`` backend: 
+To continue with our original example and select the ``"CPU_Backend"``: 
 
 .. literalinclude:: ../../../examples/abc/abc.cpp
    :language: cpp
@@ -149,7 +149,7 @@ thread needs to execute the function at the same time, create multiple
 ``CallFrame`` objects from the ``ExternalFunction``.
 
 
-.. _allocate_bkd_storage:
+.. _allocate_backend_storage:
 
 Allocate backend storage for the inputs and outputs
 ===================================================
@@ -175,7 +175,8 @@ the three parameters and the return value.
 Each tensor is a shared pointer to a :term:`Tensorview`, which is the interface 
 backends implement for tensor use. When there are no more references to the 
 tensor view, it will be freed when convenient for the backend. See the 
-:doc:`../programmable/index` documentation for details on ``Tensor``.
+:doc:`../backend-support/cpp-api` documentation for details on how to work 
+with ``Tensor``.
 
 
 .. _initialize_inputs:

doc/sphinx/source/howto/index.rst

 .. howto/index: 
 
-How to 
-======
+Constructing Graphs
+===================
 
 .. toctree::
    :maxdepth: 1

doc/sphinx/source/index.rst

@@ -17,126 +17,75 @@
 .. This documentation is available online at
 .. https://ngraph.nervanasys.com/docs/latest
 
+######################
+nGraph Compiler stack 
+######################
 
-Welcome
-=======
-
-See the latest :doc:`project/release-notes`.
-
-nGraph is an open-source C++ library, compiler stack, and runtime accelerator
-for software and neural network engineering within the :abbr:`Deep Learning (DL)`
-ecosystem. nGraph simplifies development and makes it possible to design, write,
-compile, and deploy :abbr:`Deep Neural Network (DNN)`-based solutions that can
-be adapted and deployed across many frameworks and backends. See our project
-:doc:`project/about` and `ecosystem`_ for more details.
-
-.. figure:: graphics/ngcompiler-ecosystem.png
-   :width: 650px
-   :alt: ecosystem
-
-   The Intel nGraph Compiler stack supports a broad ecosystem of frameworks and backends.
-
-
-.. _quickstart:
-
-Quick Start
-===========
-
-We have many documentation pages to help you get started.
-
-* **TensorFlow or MXNet users** can get started with
-  :doc:`framework-integration-guides`.
-
-   * `TensorFlow bridge to nGraph`_
-   * `Compiling MXNet with nGraph`_
-
-   .. note:: Note that the ``pip`` package option works only with Ubuntu 16.04
-      or greater and Intel® Xeon® CPUs. CPUs without Intel® Advanced Vector Extensions
-      512 (Intel® AVX-512) will not run these packages; the alternative is to
-      build from source. Wider support for other CPUs will be offered starting
-      in early 2019.
-
-* **Data scientists** interested in the `ONNX`_ format will find the
-  `nGraph ONNX companion tool`_ of interest.
-
-* **Framework authors and architects** will likely want to :doc:`buildlb`
-  and learn how nGraph can be used to :doc:`howto/execute`. For examples
-  of generic configurations or optimizations available when designing or
-  bridging a framework directly with nGraph, see :doc:`frameworks/index`.
-
-* To start learning about nGraph's set of **Core ops** and how they can
-  be used with Ops from other frameworks, go to :doc:`ops/index`.
-
-* **Optimization pass writers** will find :doc:`fusion/index` useful. Also
-  look for our upcoming documentation on :term:`quantization`.
-
-* For details about **PlaidML integration** and other nGraph runtime APIs,
-  see the section :doc:`programmable/index`.
-
-.. csv-table::
-   :header: "Framework", "Bridge Available?", "ONNX Support?"
-   :widths: 27, 10, 10
-
-   TensorFlow, Yes, Yes
-   MXNet, Yes, Yes
-   PaddlePaddle, Coming Soon, Yes
-   PyTorch, No, Yes
-   Other, Write your own, Custom
 
+.. toctree::
+   :maxdepth: 1 
 
-.. csv-table::
-   :header: "Backend", "Current support", "Future nGraph support"
-   :widths: 35, 10, 10
+   project/introduction.rst
 
-   Intel® Architecture Processors (CPUs), Yes, Yes
-   Intel® Nervana™ Neural Network Processor (NNPs), Yes, Yes
-   Intel® Architecture GPUs, Yes, Yes
-   AMD\* GPUs, via PlaidML, Yes
-   :abbr:`Field Programmable Gate Arrays (FPGA)` (FPGAs), Coming soon, Yes
-   NVIDIA\* GPUs, via PlaidML, Some
-   Intel Movidius™ Myriad™ 2 (VPU), Coming soon, Yes
 
+.. toctree::
+   :maxdepth: 1
+   :caption: Framework Support
+   
+   frameworks/index.rst
+   frameworks/validated/list.rst
+   frameworks/generic-configs.rst
 
-.. note:: The code in this repo is under active development as we're continually
-   adding support for more kinds of DL models and ops, compiler optimizations,
-   and backend optimizations.
 
+.. toctree::
+   :maxdepth: 1
+   :caption: nGraph Core
+
+   core/overview.rst
+   Pattern matcher <fusion/index.rst>
+   nGraph ops <ops/about.rst>
+   Graph construction <howto/index.rst>
+   Using the Python API <python_api/index.rst>
+   Compiler passes  <fusion/graph-rewrite.rst>
+   buildlb.rst
+   Using the C++ API <ops/index.rst>
+   
+.. toctree::
+   :maxdepth: 1
+   :caption: Backend support
 
-=======
+   backend-support/index.rst
+   backend-support/cpp-api.rst
 
-Contents
-========
 
 .. toctree::
    :maxdepth: 1
-   :caption: Python Ops for ONNX
+   :caption: Distributed training
+
+   distr/index.rst
 
-   python_api/index.rst
 
 .. toctree::
    :maxdepth: 1
-   :caption: Core Documentation
+   :caption: Diagnostics and visualization
+
+   diagnostics/nbench.rst
+   diagnostics/performance-profile.rst
+   diagnostics/visualize.rst
+   diagnostics/debug.rst 
 
-   buildlb.rst
-   framework-integration-guides.rst
-   frameworks/validation.rst
-   frameworks/index.rst
-   graph-basics.rst
-   howto/index.rst
-   ops/about.rst
-   ops/index.rst
-   fusion/index.rst
-   programmable/index.rst
-   distr/index.rst
 
 .. toctree::
-   :maxdepth: 2
+   :maxdepth: 1
    :caption: Project Metadata
 
-   project/index.rst
+   project/release-notes.rst
+   project/contribution-guide.rst
+   project/index.rst 
    glossary.rst
 
 
+
 Indices and tables
 ==================
 

doc/sphinx/source/project/about.rst

 .. about: 
 
-
 Architecture, Features, FAQs
 ############################
 
@@ -168,7 +167,7 @@ added with new functions that build sub-graphs from existing core ops.
 For a more detailed dive into how custom bridge code can be implemented, see our 
 documentation on how to :doc:`../howto/execute`. To learn how TensorFlow and 
 MXNet currently make use of custom bridge code, see the section on 
-:doc:`../framework-integration-guides`.
+:doc:`../frameworks/index`.
 
 .. figure:: ../graphics/bridge-to-graph-compiler.png
     :width: 733px

doc/sphinx/source/project/code-contributor-README.rst → doc/sphinx/source/project/contribution-guide.rst

-.. code-contributor-README:
+.. contribution-guide:
 
-######################
-Code Contributor Guide
-######################
+##################
+Contribution Guide
+##################
 
 License
 =======
@@ -12,6 +12,7 @@ preferably by being contributed under the Apache 2 license. Code
 contributed with another license will need the license reviewed by
 Intel before it can be accepted.
 
+
 Code formatting
 ===============
 
@@ -259,5 +260,8 @@ it is automatically enforced and reduces merge conflicts.
        int* z;
 
 
+To contribute documentation for your code, please see the :doc:`doc-contributor-README`. 
+
+
 .. _Apache 2: https://www.apache.org/licenses/LICENSE-2.0
 .. _repo wiki: https://github.com/NervanaSystems/ngraph/wiki
\ No newline at end of file

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

@@ -99,7 +99,7 @@ and the generated output will show readers of your helpful documentation
    :caption: "caption for a block of code that initializes tensors"
 
 Our documentation practices are designed around "write once, reuse" that we can 
-use to prevent code bloat.  See the :doc:`code-contributor-README` for our code 
+use to prevent code bloat.  See the :doc:`contribution-guide` for our code 
 style guide.       
 
 
@@ -127,7 +127,7 @@ To build documentation locally, run:
 
    .. code-block:: console
 
-      $ sudo apt-get install python3-sphinx
+      $ sudo apt-get install python3-sphinxcd
       $ pip3 install [-I] Sphinx==1.7.5 [--user] 
       $ pip3 install [-I] breathe numpy [--user]
       $ cd doc/sphinx/
@@ -165,3 +165,6 @@ stable reST documentation.
 .. _wiki: https://github.com/NervanaSystems/ngraph/wiki/
 .. _breathe: https://breathe.readthedocs.io/en/latest/
 .. _doxygen: http://www.doxygen.org/index.html
+
+
+.. 45555555555555555555555555555
\ No newline at end of file

doc/sphinx/source/project/governance.rst

+.. governance: 
+
+
+Governance
+##########
\ No newline at end of file

doc/sphinx/source/project/index.rst

-.. project/index.rst
+.. project/index:
 
 #################
 More about nGraph
@@ -7,9 +7,11 @@ More about nGraph
 This section contains documentation about the project and how to contribute.
 
 .. toctree::
-   :maxdepth: 1
+   :maxdepth: 2
 
-   about.rst
    release-notes.rst
-   code-contributor-README.rst
+   about.rst
+   other-efforts.rst
+   contribution-guide.rst
+   governance.rst
    doc-contributor-README.rst

doc/sphinx/source/project/introduction.rst

+.. introduction:
+
+############
+Introduction
+############
+
+The nGraph Compiler stack provides industry-standard reference and implementation 
+guidelines for working with various :abbr:`Deep Learning (DL)` (DL) models and 
+optimizing an :abbr:`Artificial Neural Network (ANN)` (often abbreviated :term:`NN`) 
+to run graph-based computations for training, inference, testing, or validation.  
+Because today's NNs make use of many custom-purpose devices (FPGAs, GPUs, CPUs, 
+and custom silicon), having such a standard simplifies what would otherwise be 
+an enormously complex and difficult-to-scale pipeline (:ref:`Figure 3 <figure-3>`) 
+from "training with your favorite framework using GPUs" (:ref:`Figure 4 <figure-4>`), 
+to deploying that (now) pre-trained model in a datacenter or production 
+environment, where infrastructure owners or software developers renting anything 
+in a datacenter ought to be mutually concerned with **efficiency per-watt**, to 
+keep costs in check.
+
+So what exactly are the motivations behind the nGraph Compiler stack? 
+
+Motivations
+===========
+
+Kernel libraries do not support graph-level optimizations
+---------------------------------------------------------
+
+A framework designed for training using GPUs requires integration with a kernel 
+library unique to that vendor's hardware. For example, after integration, a 
+kernel library can run operations that it is "familar" with optimally; however, 
+the graph itself within any larger :term:`NN` won't be optimal.
+
+.. _figure-0:
+
+.. figure:: ../graphics/framework-to-kernel-lib.png
+   :width: 555px
+   :alt: 
+
+   Figure 0: Lack of graph-level optimization makes framework-to-kernel library
+   integration enormously inefficient. The computation graph above represents 
+   the computation: "A plus B times C".
+
+
+.. _figure-1:
+
+.. figure:: ../graphics/framework-to-graph-opt.png
+   :width: 555px
+   :alt: 
+
+   Figure 1: Notice that an operation on the constant B (in this case a ``Broadcast``) 
+   can be done at compile time. This is an example of constant folding, and it 
+   is not available to a device-based kernel library.   
+
+
+.. _figure-2:
+
+.. figure:: ../graphics/ngraph-algebraic-simp.png
+   :width: 555px
+   :alt: 
+
+   Figure 2: Finally notice that the constant has value "zero" thus the add is an 
+   *identity* operation and can be eliminated. This is an example of **Algebraic 
+   simplification**, and it is not available to a device-based kernel library.
+
+
+After the two graph-level optimizations above (**Algebraic Simplification** and 
+**Constant Folding**),  we now have an optimal graph: A times C. Again, kernel 
+libraries do not support this type of optimization. Although each implementation 
+can be done individually, it will eventually yield an "exploding" number of 
+kernels the larger and more complex an :abbr:`NN (Neural Network)` becomes. For 
+some insight on why this happens, see the next section. 
+
+
+Too Many Kernels to write
+-------------------------
+
+A typical network is constructed using some kind of language-based API, which 
+translates the network or :abbr:`DL (Deep Learning)` model (statically or 
+dynamically) into serialized graphs. Those graphs can then passed through a 
+compilation process (the *Graph optimization or compilation* step in 
+*Figure 3* below), where various graph-level optimizations, like constant folding 
+or fusion can happen. These processes require unique vendor-provided libraries 
+to communicate with a driver (possibly through OpenCL\*, CUDA\*, or SYCL\*), to 
+compile and execute an implementation (kernel) for a specific 
+:abbr:`Instruction Set Architecture (ISA)`, or :term:`ISA`.
+
+Illustrated below is a simplified DL stack, showing relative complexity of 
+each component. Note that optimizing for any one on its own usually requires 
+engineering expertise that can be highly specialized to that component, and that 
+the terms have been simplified for illustrative purposes. 
+
+.. _figure-3:
+
+.. figure:: ../graphics/components-dl-stack.png
+   :width: 700px
+   :alt: A simplified DL stack
+
+   Figure 3: Components of a DL stack, simplified for illustrative purposes.
+
+There are many deep learning frameworks, each with its own strengths and user 
+bases. A setup that is common to many DL practitioners is shown in the 
+illustration below.
+
+.. _figure-4:
+
+.. figure:: ../graphics/a-common-stack.png
+   :width: 700px
+   :alt: A common implementation
+
+   Figure 4: A commonly-implemented stack uses TensorFlow\* as the frontend. 
+   The input is either optimized via Grappler, or executed directly via TensorFlow. 
+   In either case, when targeting an Nvidia\* GPU, cuDNN is called to select an 
+   optimal kernel for the operation; cuDNN then relies on CUDA\* or direct access 
+   to run code on the target; in this toy example, the target is a V100.
+
+A natural result of this approach is that the framework-level integration of 
+kernel libraries does not scale. Rather, each individual framework must be 
+manually integrated with each hardware-specific kernel library. Each integration 
+is unique to the framework and its set of deep learning operators, its view on 
+memory layout, its feature set, etc. Each of these connections, then, represents 
+significant work for what will ultimately be a brittle setup that is enormously 
+expensive to maintain.    
+
+.. _figure-5:
+
+.. figure:: ../graphics/dl-current-state.png
+   :width: 700px
+   :alt: Scalability matters
+
+   Figure 5: The number of kernels necessary to achieve optimal performance is 
+   bounded by the product of the number of chip designs one wishes to support, 
+   the number of data types supported, the number of operations, and the 
+   cardinality of each parameter for each operation.
+
+In the past, this upper bound was quite limited; however, since the industry is 
+shifting toward a more diverse future in terms of deep learning hardware, the 
+number of distinct kernels is exploding and will continue to explode.
+
+
+Get the best of both worlds
+---------------------------
+
+Integrating a framework on nGraph can be an attractive option for hardware 
+companies trying to design their own deep learning hardware or network architecture. 
+Framework integration is non-trivial amount of work, and nGraph automatically 
+does much of the heavy lifting. Furthermore, PlaidML can provide a wide range of 
+hardware coverage and optimization automatically. Any hardware that supports 
+LLVM, OpenCL, OpenGL, CUDA or Metal can be supported automatically with PlaidML 
+and nGraph.  
+
+.. _figure-6:
+
+.. figure:: ../graphics/graph-compilers-at-a-glance.png
+   :width: 700px
+   :alt: Overview of various graph and tensor compilers.
+
+   Figure 6: Overview of various graph and tensor compilers.
+
+
+.. _figure-7:
+
+.. figure:: ../graphics/tensor-compilers-at-a-glance.png
+   :width: 700px
+   :alt: A closer look at tensor compilers.
+
+   Figure 7: A closer look at tensor compilers.
+
+
+
+

doc/sphinx/source/project/other-efforts.rst

+.. _other-efforts:
+
+Other notable efforts
+=====================

src/ngraph/function.cpp

@@ -214,3 +214,13 @@ size_t Function::get_graph_size() const
     }
     return total_size;
 }
+
+size_t Function::get_placement() const
+{
+    return m_placement;
+}
+
+void Function::set_placement(size_t placement)
+{
+    m_placement = placement;
+}

src/ngraph/function.hpp

@@ -90,6 +90,9 @@ namespace ngraph
         /// graphs and should not be considered the actual memory consumption of a graph.
         size_t get_graph_size() const;
 
+        size_t get_placement() const;
+        void set_placement(size_t placement);
+
     protected:
         ResultVector m_results;
         ParameterVector m_parameters;
@@ -104,5 +107,6 @@ namespace ngraph
         size_t m_instance_id;
         std::string m_name;
         const std::string m_unique_name;
+        size_t m_placement;
     };
 }

src/ngraph/runtime/hybrid/CMakeLists.txt

@@ -17,7 +17,7 @@
 add_library(hybrid_base STATIC
     hybrid_backend.cpp
     hybrid_util.cpp
-    pass/assign_placement.cpp
+    pass/default_placement.cpp
     pass/dump.cpp
     pass/fix_get_output_element.cpp
     pass/liveness.cpp

src/ngraph/runtime/hybrid/hybrid_backend.cpp

@@ -20,7 +20,7 @@
 #include "ngraph/pass/visualize_tree.hpp"
 #include "ngraph/runtime/host_tensor.hpp"
 #include "ngraph/runtime/hybrid/hybrid_util.hpp"
-#include "ngraph/runtime/hybrid/pass/assign_placement.hpp"
+#include "ngraph/runtime/hybrid/pass/default_placement.hpp"
 #include "ngraph/runtime/hybrid/pass/dump.hpp"
 #include "ngraph/runtime/hybrid/pass/fix_get_output_element.hpp"
 #include "ngraph/runtime/hybrid/pass/liveness.hpp"
@@ -74,7 +74,7 @@ runtime::Handle runtime::hybrid::HybridBackend::compile(shared_ptr<Function> fun
 
         // Run placement pass
         ngraph::pass::Manager pass_manager;
-        pass_manager.register_pass<runtime::hybrid::pass::AssignPlacement>(m_backend_list);
+        pass_manager.register_pass<runtime::hybrid::pass::DefaultPlacement>(m_backend_list);
         pass_manager.register_pass<runtime::hybrid::pass::FixGetOutputElement>();
         pass_manager.register_pass<runtime::hybrid::pass::Liveness>();
         pass_manager.register_pass<runtime::hybrid::pass::Dump>("graph.dump");
@@ -94,7 +94,7 @@ runtime::Handle runtime::hybrid::HybridBackend::compile(shared_ptr<Function> fun
         size_t subfunction_number = 0;
         for (shared_ptr<Function>& sub_function : instance.m_sub_functions)
         {
-            size_t placement = runtime::hybrid::get_colocated_function_placement(sub_function);
+            size_t placement = sub_function->get_placement();
             if (m_debug_enabled)
             {
                 string name = "subfunction_" + to_string(subfunction_number++);
@@ -149,7 +149,7 @@ bool runtime::hybrid::HybridBackend::call(shared_ptr<Function> func,
     for (const shared_ptr<Function>& sub_function : instance.m_sub_functions)
     {
         // Init backend
-        size_t placement = runtime::hybrid::get_colocated_function_placement(sub_function);
+        size_t placement = sub_function->get_placement();
         auto backend = m_backend_list[placement];
 
         // Prepare parameter Tensors

src/ngraph/runtime/hybrid/hybrid_util.cpp

@@ -15,6 +15,7 @@
 //*****************************************************************************
 
 #include "ngraph/runtime/hybrid/hybrid_util.hpp"
+#include "ngraph/log.hpp"
 #include "ngraph/pass/manager.hpp"
 #include "ngraph/pass/visualize_tree.hpp"
 
@@ -25,20 +26,20 @@ static Node* take_independent_node_with_placement_priority(
     map<size_t, deque<Node*>>& independent_nodes_by_placement, size_t placement)
 {
     Node* selected_node = nullptr;
-    if (independent_nodes_by_placement.find(placement) != independent_nodes_by_placement.end() &&
-        independent_nodes_by_placement.at(placement).size() != 0)
+    auto it = independent_nodes_by_placement.find(placement);
+    if (it != independent_nodes_by_placement.end() && it->second.size() != 0)
     {
-        selected_node = independent_nodes_by_placement.at(placement).front();
-        independent_nodes_by_placement.at(placement).pop_front();
+        selected_node = it->second.front();
+        it->second.pop_front();
     }
     else
     {
-        for (auto& it : independent_nodes_by_placement)
+        for (auto& p : independent_nodes_by_placement)
         {
-            if (it.second.size() > 0)
+            if (p.second.size() > 0)
             {
-                selected_node = it.second.front();
-                it.second.pop_front();
+                selected_node = p.second.front();
+                p.second.pop_front();
                 break;
             }
         }
@@ -238,8 +239,10 @@ pair<vector<shared_ptr<Function>>, unordered_map<shared_ptr<op::Parameter>, shar
     {
         ParameterVector par_vector;
         ResultVector res_vector;
+        size_t placement = -1;
         for (auto node : cluster)
         {
+            placement = node->get_placement_index();
             if (auto res_node = dynamic_pointer_cast<op::Result>(node))
             {
                 res_vector.push_back(res_node);
@@ -250,6 +253,7 @@ pair<vector<shared_ptr<Function>>, unordered_map<shared_ptr<op::Parameter>, shar
             }
         }
         auto sub_function = make_shared<Function>(res_vector, par_vector);
+        sub_function->set_placement(placement);
         sub_functions.push_back(sub_function);
 #ifdef HYBRID_DEBUG
         ngraph::pass::Manager pass_manager;
@@ -261,26 +265,3 @@ pair<vector<shared_ptr<Function>>, unordered_map<shared_ptr<op::Parameter>, shar
 
     return make_pair(sub_functions, map_parameter_to_result);
 }
-
-// Assert that nodes in the function is colocated and return that placement
-size_t runtime::hybrid::get_colocated_function_placement(shared_ptr<Function> func)
-{
-    auto ops = func->get_ops();
-
-    //it's okay to not do Placement::DEFAULT check; the same node will be checked in the loop below
-    size_t function_placement = ops.front()->get_placement_index();
-    for (auto op : ops)
-    {
-        size_t node_placement = op->get_placement_index();
-        if (node_placement == Node::placement_invalid)
-        {
-            throw ngraph_error("Node " + op->get_name() + " should have a device placement");
-        }
-        if (function_placement != node_placement)
-        {
-            throw ngraph_error("Function contains nodes of two different placements");
-        }
-    }
-
-    return function_placement;
-}

src/ngraph/runtime/hybrid/hybrid_util.hpp

@@ -35,9 +35,6 @@ namespace ngraph
                 std::vector<std::shared_ptr<Function>>,
                 std::unordered_map<std::shared_ptr<op::Parameter>, std::shared_ptr<op::Result>>>
                 split_function_by_placement(const std::shared_ptr<Function>& f);
-
-            // Assert that nodes in the function is colocated and return that placement
-            size_t get_colocated_function_placement(std::shared_ptr<Function> func);
         }
     }
 }

src/ngraph/runtime/hybrid/pass/assign_placement.cpp → src/ngraph/runtime/hybrid/pass/default_placement.cpp

@@ -14,7 +14,7 @@
 // limitations under the License.
 //*****************************************************************************
 
-#include "ngraph/runtime/hybrid/pass/assign_placement.hpp"
+#include "ngraph/runtime/hybrid/pass/default_placement.hpp"
 #include "ngraph/log.hpp"
 #include "ngraph/node.hpp"
 #include "ngraph/placement.hpp"
@@ -23,13 +23,13 @@
 using namespace ngraph;
 using namespace std;
 
-runtime::hybrid::pass::AssignPlacement::AssignPlacement(
+runtime::hybrid::pass::DefaultPlacement::DefaultPlacement(
     const vector<shared_ptr<runtime::Backend>>& placement_backends)
     : m_placement_backends(placement_backends)
 {
 }
 
-bool runtime::hybrid::pass::AssignPlacement::run_on_node(shared_ptr<Node> node)
+bool runtime::hybrid::pass::DefaultPlacement::run_on_node(shared_ptr<Node> node)
 {
     size_t backend_index = 0;
     for (auto backend : m_placement_backends)

src/ngraph/runtime/hybrid/pass/assign_placement.hpp → src/ngraph/runtime/hybrid/pass/default_placement.hpp

@@ -30,16 +30,16 @@ namespace ngraph
         {
             namespace pass
             {
-                class AssignPlacement;
+                class DefaultPlacement;
             }
         }
     }
 }
 
-class ngraph::runtime::hybrid::pass::AssignPlacement : public ngraph::pass::NodePass
+class ngraph::runtime::hybrid::pass::DefaultPlacement : public ngraph::pass::NodePass
 {
 public:
-    AssignPlacement(
+    DefaultPlacement(
         const std::vector<std::shared_ptr<ngraph::runtime::Backend>>& placement_backends);
 
 private:

test/hybrid_backend.cpp

@@ -56,7 +56,7 @@ TEST(HYBRID, abc)
     auto t1 = A * B;
     auto t2 = t1 * D;
     auto C = make_shared<op::Parameter>(element::f32, shape);
-    auto f = make_shared<Function>((t2 + C) * t1, ParameterVector{A, B, C, D});
+    auto f = make_shared<Function>(((t2 + C) + A) * t1, ParameterVector{A, B, C, D});
 
     shared_ptr<runtime::Backend> backend = runtime::Backend::create("H1");
     static_pointer_cast<runtime::hybrid::HybridBackend>(backend)->set_debug_enabled(true);
@@ -75,5 +75,5 @@ TEST(HYBRID, abc)
 
     auto handle = backend->compile(f);
     backend->call_with_validate(handle, {result}, {a, b, c, d});
-    EXPECT_EQ(read_vector<float>(result), (vector<float>{145, 552, 1113, 1408}));
+    EXPECT_EQ(read_vector<float>(result), (vector<float>{150, 576, 1176, 1536}));
 }