Skip to content

N
ngraph
Unverified Commit 166fb921 authored by Nagy Mostafa's avatar Nagy Mostafa Committed by GitHub
Browse files
Options

Merge branch 'master' into nmostafa/update_mlir

parents 285a24ae 27199cee
Expand all Show whitespace changes
Inline Side-by-side
Showing with 423 additions and 205 deletions
doc/sphinx/ngraph_theme/layout.html
View file @ 166fb921
......@@ -84,9 +84,9 @@
<body>
<div id="menu-float" class="menu-float">
<a href="https://www.ngraph.ai">Home</a>
<a href="https://www.youtube.com/embed/C9S0nmNS8bQ">Video</a>
<a href="https://www.ngraph.ai/ecosystem">Ecosystem</a>
<a href="https://www.ngraph.ai" target="_blank">Home</a>
<a href="https://www.youtube.com/embed/C9S0nmNS8bQ" target="_blank">Video</a>
<a href="https://www.ngraph.ai/ecosystem" target="_blank">Ecosystem</a>
<a href="https://ngraph.nervanasys.com/docs/latest">Docs</a>
<a href="https://www.ngraph.ai/tutorials">Tutorials</a>
<a href="https://ngraph.slack.com/"><img src="https://cdn.brandfolder.io/5H442O3W/as/pl546j-7le8zk-5h439l/Slack_Mark_Monochrome_White.png?width=35&height=35"></a>
......
doc/sphinx/ngraph_theme/ngversions.html
View file @ 166fb921
......@@ -10,13 +10,12 @@
<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/v0.26">0.26</a></li> -->
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.26.0">0.26.0</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.26.0">0.25.1</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.26.0-rc.3">Prerelease 0.26</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.25.1-rc.4">0.25.1</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.25.0">0.25.0</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.24.0">0.24.0</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/v0.22.2-rc.0">0.22.2</a></li>
<li><a href="https://github.com/NervanaSystems/ngraph/releases/tag/v0.22.1">0.22.1</a></li>
</ul></dd>
</dl>
<dl>
......
doc/sphinx/ngraph_theme/static/css/theme.css
View file @ 166fb921
......@@ -1645,6 +1645,7 @@ h1 {
h2 {
font-size: 133%;
text-decoration: underline 4px dotted #D3D3D3;
margin-top: -2px;
}
h2, .rst-content .toctree-wrapper p.caption {
......
doc/sphinx/source/backends/index.rst
View file @ 166fb921
......@@ -51,7 +51,7 @@ How to use?
#. A single iteration of the executable is executed by calling the ``call``
method on the ``Executable`` object.
.. figure:: ../graphics/execution-interface.png
.. figure:: ../graphics/ExecutionInterfaceRunGraphs.png
:width: 650px
The execution interface for nGraph
......@@ -69,6 +69,13 @@ interface; each backend implements the following five functions:
* And, finally, the ``call()`` method is used to invoke an nGraph function
against a particular set of tensors.
How to display ngraph-related passes executed during runtime?
-------------------------------------------------------------
One easy way to get info about passes is to set the environment variable
:envvar:`NGRAPH_PROFILE_PASS_ENABLE=1`. With this set, the pass manager
will dump the name and execution time of each pass.
.. _ngraph_bridge:
......
doc/sphinx/source/core/constructing-graphs/index.rst
View file @ 166fb921
.. howto/index:
.. core/constructing-graphs/index.rst:
Constructing graphs
.. _constructing_graphs:
Constructing Graphs
===================
.. toctree::
......
doc/sphinx/source/core/fusion/index.rst
View file @ 166fb921
.. fusion/index.rst:
Pattern matcher
Pattern Matcher
###############
.. toctree::
......
doc/sphinx/source/core/fusion/overview.rst
View file @ 166fb921
.. fusion/overview.rst
.. core/fusion/overview.rst
.. _fusion_overview:
Overview: Optimize graphs with nGraph Compiler fusions
-------------------------------------------------------
......
doc/sphinx/source/core/overview.rst
View file @ 166fb921
.. core/overview.rst:
Overview
========
Basic concepts
==============
.. figure:: ../graphics/whole-stack.png
.. figure:: ../graphics/nGraphCompilerstack.png
:alt: The whole stack
The whole nGraph Compiler stack
......@@ -13,7 +13,7 @@ The nGraph Compiler stack consists of bridges, core, and backends. We'll examine
each of these briefly to get started.
A framework bridge interfaces with the "frontend" Core API. A framework bridge
is a component that sits between a framework like TensorFlow or MXNet, and the
is a component that sits between a framework like TensorFlow or PaddlePaddle, and the
nGraph Core frontend API. A framework bridge does two things: first, it
translates a framework's operations into graphs in nGraph’s in-memory
:abbr:`Intermediary Representation (IR)`. Second, it executes the nGraph IR
......@@ -24,8 +24,9 @@ are some common patterns: a fairly typical example for a graph-based framework
is illustrated here, and consists of basically two phases: a **clustering**
phase and a **translation** phase.
.. figure:: ../graphics/translation-flow-to-ng-fofx.png
:alt: The whole stack
.. figure:: ../graphics/overview-translation-flow.svg
:width: 725px
:alt: Translation flow to an nGraph function graph
Translation flow to an nGraph function
......
doc/sphinx/source/core/passes/passes-that-use-matcher.rst
View file @ 166fb921
.. fusion/passes-that-use-matcher.rst:
.. core/passes/passes-that-use-matcher.rst:
Passes that use Matcher
......
doc/sphinx/source/core/passes/passes.rst
View file @ 166fb921
.. core/passes:
.. core/passes/passes.rst:
Compiler passes
.. _core_compiler_passes:
Compiler Passes
===============
.. toctree::
......@@ -11,9 +13,8 @@ Compiler passes
passes-that-use-matcher.rst
Overview
--------
Basic concepts
--------------
*Generic graph optimization passes*
......
doc/sphinx/source/frameworks/fw_overview.rst 0 → 100644
View file @ 166fb921
:orphan:
.. frameworks/fw_overview:
.. _fw_overview:
Overview
========
A framework is "supported" with a framework :term:`bridge` that can be written or
cloned and used to connect to nGraph device backends while maintaining the
framework's programmatic or user interface. There is a bridge for the
`TensorFlow framework`_. We also have a :doc:`paddle_integ` bridge. Intel
previously contributed work to an MXNet bridge; however, support for this
bridge is no longer active.
`ONNX`_ on its own is not a framework; however, it can be used with nGraph's
:doc:`../python_api/index` to import and execute ONNX models.
.. figure:: ../graphics/overview-framework-bridges.svg
:width: 960px
:alt: Framework bridge to a graph construction
Framework bridge to nGraph
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-modularized stack of components that
can be implemented or integrated in countless ways.
Please read this section if you are considering incorporating components from
the nGraph Compiler stack in your framework or neural network design. Contents
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.
.. figure:: ../graphics/overview-translation-flow.svg
:width: 725px
:alt: Translation flow to nGraph function graph
.. _TensorFlow framework: https://github.com/tensorflow/ngraph-bridge/README.md
.. _ONNX: http://onnx.ai/
.. _tune the workload to extract best performance: https://ai.intel.com/accelerating-deep-learning-training-inference-system-level-optimizations
.. _a few small: https://software.intel.com/en-us/articles/boosting-deep-learning-training-inference-performance-on-xeon-and-xeon-phi
doc/sphinx/source/frameworks/index.rst
View file @ 166fb921
.. frameworks/index.rst
Working with Frameworks
=======================
.. include:: overview.rst
#######################
.. toctree::
:maxdepth: 1
getting_started.rst
overview.rst
quickstart.rst
onnx_integ.rst
paddle_integ.rst
tensorflow_connect.rst
generic_configs.rst
other.rst
doc/sphinx/source/frameworks/onnx_integ.rst
View file @ 166fb921
.. frameworks/onnx_integ.rst:
ONNX Support
============
ONNX overview
=============
nGraph is able to import and execute ONNX models. Models are converted to
nGraph's :abbr:`Intermediate Representation (IR)` and converted to ``Function``
......@@ -78,7 +78,7 @@ data:
Find more information about nGraph and ONNX in the
`nGraph ONNX`_ GitHub\* repository.
`nGraph ONNX`_ GitHub repository.
.. _ngraph ONNX: https://github.com/NervanaSystems/ngraph-onnx
......
doc/sphinx/source/frameworks/generic_configs.rst → doc/sphinx/source/frameworks/other.rst
View file @ 166fb921
.. frameworks/generic_configs.rst:
.. frameworks/other.rst:
.. _generic_configs:
.. _fw_other:
Integrating new frameworks
==========================
Integrating other 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
......
doc/sphinx/source/frameworks/overview.rst
View file @ 166fb921
.. frameworks/overview.rst
.. _fw_overview:
Basic concepts
==============
Overview
========
.. figure:: ../graphics/overview-framework-bridges.svg
:width: 960px
:alt: Bridge to nGraph graph construction API
A framework is "supported" with a framework :term:`bridge` that can be written or
cloned and used to connect to nGraph device backends while maintaining the
framework's programmatic or user interface. A `bridge currently exists`_ for the
TensorFlow framework. We also have a bridge to do :doc:`paddle_integ`. Intel
previously contributed work to an MXNet bridge; however, support for this
bridge is no longer active.
A framework bridge connects to the nGraph graph construction API
`ONNX`_ on its own is not a framework; however, it can be used with nGraph's
:doc:`../python_api/index` to import and execute ONNX models.
To understand how a data science :term:`framework` (:doc:`TensorFlow <tensorflow_connect>`,
PyTorch, :doc:`paddle_integ`, and others) can unlock acceleration available in
the nGraph Compiler, it helps to familiarize yourself with some basic concepts.
.. figure:: ../graphics/overview-framework-bridges.svg
:width: 960px
:alt: JiT compiling of a computation
We use the term :term:`bridge` to describe code that connects to any nGraph
device backend(s) while maintaining the framework's programmatic or user
interface. We have a `bridge for the TensorFlow framework`_. We also have a
:doc:`paddle_integ` bridge. Intel previously :doc:`contributed work to an MXNet bridge <../project/extras/testing_latency>`;
however, support for the MXNet bridge is no longer active.
`ONNX`_ on its own is not a framework; it can be used with nGraph's
:doc:`../python_api/index` to import and execute ONNX models.
:abbr:`Just-in-Time (JiT)` Compiling for computation. nGraph `Core`
components are colored in blue.
Because it is framework agnostic (providing opportunities to optimize at the
graph level), nGraph can do the heavy lifting required by many popular
:doc:`workloads <validated/list>` without any additional effort of the framework user.
Optimizations that were previously available only after careful integration of
a kernel or hardware-specific library are exposed via the
:doc:`Core graph construction API <../core/constructing-graphs/index>`
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.
The illustration above shows how this works.
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-modularized stack of components that
can be implemented or integrated in countless ways.
While a :abbr:`Deep Learning (DL)` framework is ultimately meant for end-use by
data scientists, or for deployment in cloud container environments, nGraph's
:doc:`Core ops <../core/overview>` are designed for framework builders themselves.
We invite anyone working on new and novel frameworks or neural network designs
to explore our highly-modularized stack of components.
Please read this section if you are considering incorporating components from
the nGraph Compiler stack in your framework or neural network design. Contents
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.
Please read the :doc:`other` section for other framework-agnostic
configurations available to users of the nGraph Compiler stack.
.. figure:: ../graphics/overview-translation-flow.svg
:width: 725px
:alt: Translation flow to nGraph function graph
:alt: Translation flow to an nGraph function graph
.. _bridge currently exists: https://github.com/tensorflow/ngraph-bridge/README.md
.. _bridge for the TensorFlow framework: https://github.com/tensorflow/ngraph-bridge/README.md
.. _ONNX: http://onnx.ai/
.. _tune the workload to extract best performance: https://ai.intel.com/accelerating-deep-learning-training-inference-system-level-optimizations
.. _a few small: https://software.intel.com/en-us/articles/boosting-deep-learning-training-inference-performance-on-xeon-and-xeon-phi
\ No newline at end of file
doc/sphinx/source/frameworks/paddle_integ.rst
View file @ 166fb921
.. frameworks/paddle_integ.rst:
PaddlePaddle
============
nGraph PaddlePaddle integration overview
----------------------------------------
PaddlePaddle integration
========================
PaddlePaddle is an open source deep learning framework developed by Baidu. It
aims to enable performant large-scale distributed computation for deep learning.
......@@ -67,7 +64,7 @@ is organized in the following file structure:
.. _figure-paddle-dir:
.. figure:: ../graphics/paddlepaddle_directory.png
.. figure:: ../graphics/PaddlePaddleDir.svg
:width: 555px
:alt:
......@@ -135,8 +132,8 @@ and nGraph bridges are provided below:
are managed through a map.
- SetOutputNode: sets the constructed node to the map.
- Related code :
+ ``Paddle/fluid/operators/ngraph/ngraph_bridge.h`` `link to ngraph_bridge header code`_
+ ``Paddle/fluid/operators/ngraph/ngraph_bridge.cc`` `link to ngraph_bridge cpp code`_
- ``Paddle/fluid/operators/ngraph/ngraph_bridge.h`` `link to ngraph_bridge header code`_
- ``Paddle/fluid/operators/ngraph/ngraph_bridge.cc`` `link to ngraph_bridge cpp code`_
nGraph compilation control and trigger method
----------------------------------------------
......@@ -149,8 +146,9 @@ nGraph compilation control and trigger method
#. **Trigger Control** -- ``FLAGS_use_ngraph`` triggers nGraph. If this option
is set to ``true``, nGraph will be triggered by the PaddlePaddle executor
to convert and execute the supported subgraph. `Examples are provided`_ under
``paddle/benchmark/fluid/ngraph``.
to convert and execute the supported subgraph. Demos are provided under
``paddle/benchmark/fluid/train/demo`` (link `train_demo`_) and ``paddle/benchmark/fluid/train/imdb_demo``
(link `imdb_demo`_)
.. _link to ngraph_engine_op header code: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/ngraph/ngraph_engine_op.h
......@@ -160,5 +158,5 @@ nGraph compilation control and trigger method
.. _located in the ngraph ops: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/ngraph/ops
.. _link to ngraph_bridge header code: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/ngraph/ngraph_bridge.h
.. _link to ngraph_bridge cpp code: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/ngraph/ngraph_bridge.cc
.. _Examples are provided: https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/fluid
.. _train_demo: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/train/demo
.. _imdb_demo: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/train/imdb_demo
doc/sphinx/source/frameworks/getting_started.rst → doc/sphinx/source/frameworks/quickstart.rst
View file @ 166fb921
.. frameworks/getting_started.rst
.. frameworks/quickstart.rst
Getting Started
###############
.. _fw_quickstart:
Quick start
===========
No matter what your level of experience with :abbr:`Deep Learning (DL)` systems
may be, nGraph provides a path to start working with the DL stack. Let's begin
with the easiest and most straightforward options.
.. figure:: ../graphics/translation-flow-to-ng-fofx.png
:width: 725px
:alt: Translation flow to nGraph function graph
TensorFlow
----------
The easiest way to get started is to use the latest PyPI `ngraph-tensorflow-bridge`_,
which has instructions for Linux* systems, and tips for users of Mac OS X.
which has instructions for Linux\* systems, and tips for users of Mac OS X.
You can install TensorFlow\* and nGraph to a virtual environment; otherwise, the code
You can install TensorFlow and nGraph in a virtual environment; otherwise, the code
will install to a system location.
.. code-block:: console
......@@ -45,15 +46,15 @@ Output will look something like:
More detail in the `ngraph_bridge examples`_ directory.
See also the `diagnostic tools`_.
ONNX
====
----
Another easy way to get started working with the :abbr:`DL (Deep Learning)`
stack is to try the examples available via `nGraph ONNX`_.
Installation
------------
To prepare your environment to use nGraph and ONNX, install the Python packages
for nGraph, ONNX and NumPy:
......@@ -67,18 +68,18 @@ Now you can start exploring some of the :doc:`onnx_integ` examples.
See also nGraph's :doc:`../python_api/index`.
PlaidML
=======
-------
See the :ref:`ngraph_plaidml_backend` section on how to build the
nGraph-PlaidML.
Other integration paths
=======================
-----------------------
If you are considering incorporating components from the nGraph Compiler stack
in your framework or neural network design, another useful doc is the section
on :doc:`generic-configs`. Contents here are also useful if you are working on
on :doc:`other` . Contents 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.
......@@ -86,3 +87,4 @@ widely-supported than the popular frameworks like TensorFlow and PyTorch.
.. _ngraph-tensorflow-bridge: https://pypi.org/project/ngraph-tensorflow-bridge
.. _ngraph ONNX: https://github.com/NervanaSystems/ngraph-onnx
.. _ngraph_bridge examples: https://github.com/tensorflow/ngraph-bridge/blob/master/examples/README.md
.. _diagnostic tools: https://github.com/tensorflow/ngraph-bridge/blob/master/diagnostics/README.md
\ No newline at end of file
doc/sphinx/source/frameworks/tensorflow_connect.rst
View file @ 166fb921
.. frameworks/tensorflow_connect.rst:
Connect TensorFlow\*
====================
nGraph Bridge for TensorFlow
============================
See the `README`_ on the `ngraph_bridge repo`_ for the many ways to connect
......
doc/sphinx/source/frameworks/validated/list.rst
View file @ 166fb921
......@@ -3,7 +3,7 @@
.. _validated:
Validated workloads
Validated Workloads
###################
We have validated performance [#f1]_ for the following workloads:
......
doc/sphinx/source/graphics/PaddlePaddleDir.svg 0 → 100644
View file @ 166fb921
This diff is collapsed.
doc/sphinx/source/index.rst
View file @ 166fb921
......@@ -12,9 +12,9 @@
.. limitations under the License.
.. ---------------------------------------------------------------------------
######################
nGraph Compiler stack
######################
nGraph Compiler stack Documentation
###################################
.. _ngraph_home:
......@@ -23,36 +23,25 @@ nGraph Compiler stack
nGraph Compiler stack documentation for version |version|.
Documentation for the latest (master) development branch can be found
at https://ngraph.nervanasys.com/docs/latest
.. https://docs.ngraph.ai/
.. only:: (development or daily)
nGraph Compiler stack documentation for the master tree under development
(version |version|).
The nGraph Library and Compiler stack are provided under the `Apache 2.0 license`_
(found in the LICENSE file in the project's `repo`_). It may also import or reference
packages, scripts, and other files that use licensing.
.. _Apache 2.0 license: https://github.com/NervanaSystems/ngraph/blob/master/LICENSE
.. _repo: https://github.com/NervanaSystems/ngraph
.. toctree::
:name: mastertoctree
:titlesonly:
.. toctree::
:maxdepth: 1
:caption: Getting Started
introduction.rst
features.rst
project/release-notes.rst
.. toctree::
:maxdepth: 1
:maxdepth: 2
:caption: Framework Support
frameworks/index.rst
......@@ -89,15 +78,20 @@ packages, scripts, and other files that use licensing.
.. toctree::
:maxdepth: 1
:caption: Project Metadata
:caption: Contributing
project/release-notes.rst
project/contribution-guide.rst
project/index.rst
project/extras/index.rst
glossary.rst
.. toctree::
:maxdepth: 1
:hidden:
project/release-notes.rst
project/index.rst
project/extras/index.rst
.. only:: html
......
doc/sphinx/source/ops/index.rst
View file @ 166fb921
......@@ -160,15 +160,30 @@ Not currently a comprehensive list.
More about Core Ops
-------------------
An ``Op``'s primary role is to function as a node in a directed acyclic graph
dependency computation graph.
An ``Op``'s primary role is to function as a node in a ddirected acyclic
computation graph.
*Core ops* are ops that are available and generally useful to all framework
bridges and that can be compiled by all transformers. A framework bridge may
define framework-specific ops to simplify graph construction, provided that the
bridge can enable every transformer to replace all such ops with equivalent
clusters or subgraphs composed of core ops. In a similar manner, transformers may define
transformer-specific ops to represent kernels or other intermediate operations.
clusters or subgraphs composed of core ops. In a similar manner, transformers
may define transformer-specific ops to represent kernels or other intermediate
operations.
The input and output ports of ops are any of the functions which work with
``Output<Node>/Input<Node>``. Previous functions that worked at the level
of ops are deprecated, like::
Node::get_element_type()
as it does not take any input. This function has been replaced with new
functions like::
Node::get_output_element_type(index)
where there is no ambiguity.
If a framework supports extending the set of ops it offers, a bridge may even
expose transformer-specific ops to the framework user.
......
doc/sphinx/source/ops/max.rst
View file @ 166fb921
......@@ -44,6 +44,6 @@ Outputs
C++ Interface
=============
.. doxygenclass:: ngraph::op::Max
.. doxygenclass:: ngraph::op::v0::Max
:project: ngraph
:members: m_axes
doc/sphinx/source/ops/min.rst
View file @ 166fb921
......@@ -44,6 +44,6 @@ Outputs
C++ Interface
=============
.. doxygenclass:: ngraph::op::Min
.. doxygenclass:: ngraph::op::v0::Min
:project: ngraph
:members: m_axes
doc/sphinx/source/project/contribution-guide.rst
View file @ 166fb921
......@@ -4,7 +4,7 @@
.. _contribution_guide:
##################
Contribution guide
Contribution Guide
##################
......
doc/sphinx/source/project/release-notes.rst
View file @ 166fb921
......@@ -2,6 +2,8 @@
:orphan:
.. _release_notes:
Release Notes
#############
......@@ -18,19 +20,33 @@ We are pleased to announce the release of version |version|.
Core updates for |version|
--------------------------
+ All ops support ``Output<Node>`` arguments
+ Additional ops
+ ONNX handling unknown domains
+ Provenance works with builders and fused ops
+ ``RPATH`` for finding openmpi
+ Negative indices/axes fixes
+ Migrate some ``get_argument`` removals
+ Negative indices/axes fixes
+ Better support for MKL-DNN 1.0 (DNNL)
Latest documentation updates
----------------------------
Latest documentation updates for |version|
------------------------------------------
+ Note the only support for nGPU is now through PlaidML; nGraph support for nGPU
(via cuDNN) has been deprecated.
+ iGPU works only with nGraph version `0.24`.
+ Add new Sphinx-friendly theme (can be built natively for an alternative to ngraph.ai docs).
+ Update PaddlePaddle documentation to reflect demo directories instead of example directory.
+ Update doc regarding the validation of ``Sum`` op.
.. important:: Pre-releases (``-rc-0.*``) have newer features, and are less stable.
Changelog on Previous Releases
==============================
......@@ -41,7 +57,7 @@ Changelog on Previous Releases
+ Add rank id to trace file name
+ Allow provenance merging to be disabled
+ Remove some white-listed compiler warnings
+ Provenance on builders and fused op expansions
+ Provenance, builders, ops that make ops, and fused op expansions
0.25.0
......@@ -136,8 +152,8 @@ Changelog on Previous Releases
+ Provenance improvements
0.19
----
pre-0.20
--------
+ More dynamic shape preparation
+ Distributed interface factored out
......@@ -153,11 +169,7 @@ Changelog on Previous Releases
+ Add graph visualization tools to doc
+ Update doxygen to be friendlier to frontends
0.18
----
.. 0.18
+ Python formatting issue
+ mkl-dnn work-around
+ Event tracing improvements
......@@ -166,10 +178,7 @@ Changelog on Previous Releases
+ ONNX quantization
+ More fusions
0.17
----
.. 0.17
+ Allow negative padding in more places
+ Add code generation for some quantized ops
+ Preliminary dynamic shape support
......@@ -177,10 +186,7 @@ Changelog on Previous Releases
+ Pad op takes CoordinateDiff instead of Shape pad values to allow for negative
padding.
0.16
----
.. 0.16
+ NodeInput and NodeOutput classes prepare for simplifications of Node
+ Test improvements
+ Additional quantization ops
......
doc/sphinx/source/sitemap.rst deleted 100644 → 0
View file @ 285a24ae
:orphan:
.. toctree::
:includehidden:
frameworks/index
project/index
python_api/index
inspection/index
core/overview
backends/index
project/extras/index
src/contrib/mlir/compiler/lowerer.cpp
View file @ 166fb921
......@@ -589,10 +589,14 @@ namespace
IndexedValue iRes(result), iLhs(lhs), iRhs(rhs);
ValueHandle zeroInit(rewriter.create<ConstantOp>(loc, rewriter.getZeroAttr(elemTy)));
{
IndexHandle n, k;
LoopBuilder(&n, nLb, nUb, nStep)(
[&] { LoopBuilder(&k, kLb, kUb, kStep)([&] { iRes(n, k) = zeroInit; }); });
}
LoopBuilder(&n, nLb, nUb, nStep)([&] {
LoopBuilder(&k, kLb, kUb, kStep)([&] {
iRes(n, k) = zeroInit;
LoopBuilder(&m, mLb, mUb, mStep)([&] { iRes(n, k) += iLhs(n, m) * iRhs(m, k); });
LoopBuilder(&m, mLb, mUb, mStep)([&] {
LoopBuilder(&k, kLb, kUb, kStep)([&] { iRes(n, k) += iLhs(n, m) * iRhs(m, k); });
});
});
......
src/ngraph/node.hpp
View file @ 166fb921
......@@ -37,6 +37,7 @@
#include "ngraph/descriptor/output.hpp"
#include "ngraph/descriptor/tensor.hpp"
#include "ngraph/op/util/attr_types.hpp"
#include "ngraph/op/util/op_annotations.hpp"
#include "ngraph/placement.hpp"
#include "ngraph/strides.hpp"
#include "ngraph/type.hpp"
......@@ -485,6 +486,15 @@ namespace ngraph
/// \throw std::out_of_range if the node does not have at least `output_index+1` outputs.
Output<const Node> output(size_t output_index) const;
void set_op_annotations(std::shared_ptr<ngraph::op::util::OpAnnotations> op_annotations)
{
m_op_annotations = op_annotations;
}
std::shared_ptr<ngraph::op::util::OpAnnotations> get_op_annotations() const
{
return m_op_annotations;
}
private:
descriptor::Input& get_input_descriptor(size_t position);
descriptor::Output& get_output_descriptor(size_t position);
......@@ -504,6 +514,7 @@ namespace ngraph
std::unordered_map<Node*, autodiff::Adjoints> m_adjoint_map;
Placement m_placement = Placement::DEFAULT;
size_t m_placement_index = placement_invalid;
std::shared_ptr<ngraph::op::util::OpAnnotations> m_op_annotations;
};
/// \brief A handle for one of a node's inputs.
......
src/ngraph/op/avg_pool.cpp
View file @ 166fb921
......@@ -234,13 +234,13 @@ shared_ptr<Node> op::v0::AvgPool::get_default_value() const
}
op::v0::AvgPoolBackprop::AvgPoolBackprop(const Shape& forward_arg_shape,
const shared_ptr<Node>& delta,
const Output<Node>& delta,
const Shape& window_shape,
const Strides& window_movement_strides,
const Shape& padding_below,
const Shape& padding_above,
bool include_padding_in_avg_computation)
: Op(check_single_output_args({delta}))
: Op({delta})
, m_forward_arg_shape(forward_arg_shape)
, m_window_shape(window_shape)
, m_window_movement_strides(window_movement_strides)
......
src/ngraph/op/avg_pool.hpp
View file @ 166fb921
......@@ -179,7 +179,7 @@ namespace ngraph
const NodeTypeInfo& get_type_info() const override { return type_info; }
AvgPoolBackprop() = default;
AvgPoolBackprop(const Shape& forward_arg_shape,
const std::shared_ptr<Node>& delta,
const Output<Node>& delta,
const Shape& window_shape,
const Strides& window_movement_strides,
const Shape& padding_below,
......
src/ngraph/op/constant.hpp
View file @ 166fb921
......@@ -30,7 +30,7 @@ namespace ngraph
namespace op
{
/// \brief Class for constants.
class Constant : public Node
class Constant : public Op
{
public:
NGRAPH_API
......@@ -251,7 +251,7 @@ namespace ngraph
protected:
void* get_data_ptr_nc() { return (m_data ? m_data->get_ptr() : nullptr); }
Constant(const OutputVector& args)
: Node(args)
: Op(args)
, m_shape({})
{
}
......
src/ngraph/op/op.hpp
View file @ 166fb921
......@@ -18,9 +18,7 @@
#include <string>
#include "ngraph/autodiff/adjoints.hpp"
#include "ngraph/node.hpp"
#include "ngraph/op/util/op_annotations.hpp"
namespace ngraph
{
......@@ -30,15 +28,6 @@ namespace ngraph
class Op : public Node
{
public:
void set_op_annotations(std::shared_ptr<ngraph::op::util::OpAnnotations> op_annotations)
{
m_op_annotations = op_annotations;
}
std::shared_ptr<ngraph::op::util::OpAnnotations> get_op_annotations() const
{
return m_op_annotations;
}
virtual bool is_op() const override { return true; }
protected:
Op()
......@@ -48,9 +37,6 @@ namespace ngraph
Op(const NodeVector& arguments);
Op(const OutputVector& arguments);
Op(const std::string& node_type, const NodeVector& arguments);
private:
std::shared_ptr<ngraph::op::util::OpAnnotations> m_op_annotations;
};
}
}
src/ngraph/op/reverse.cpp
View file @ 166fb921
......@@ -24,16 +24,16 @@
using namespace std;
using namespace ngraph;
constexpr NodeTypeInfo op::Reverse::type_info;
constexpr NodeTypeInfo op::v0::Reverse::type_info;
op::Reverse::Reverse(const Output<Node>& arg, const AxisSet& reversed_axes)
op::v0::Reverse::Reverse(const Output<Node>& arg, const AxisSet& reversed_axes)
: Op({arg})
, m_reversed_axes(reversed_axes)
{
constructor_validate_and_infer_types();
}
void op::Reverse::validate_and_infer_types()
void op::v0::Reverse::validate_and_infer_types()
{
const auto input_shape = get_input_partial_shape(0);
const Dimension input_rank = input_shape.rank();
......@@ -56,13 +56,13 @@ void op::Reverse::validate_and_infer_types()
set_output_type(0, get_input_element_type(0), input_shape);
}
shared_ptr<Node> op::Reverse::copy_with_new_args(const NodeVector& new_args) const
shared_ptr<Node> op::v0::Reverse::copy_with_new_args(const NodeVector& new_args) const
{
check_new_args_count(this, new_args);
return make_shared<Reverse>(new_args.at(0), m_reversed_axes);
return make_shared<v0::Reverse>(new_args.at(0), m_reversed_axes);
}
void op::Reverse::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas)
void op::v0::Reverse::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas)
{
auto delta = deltas.at(0);
......
src/ngraph/op/reverse.hpp
View file @ 166fb921
......@@ -21,6 +21,8 @@
namespace ngraph
{
namespace op
{
namespace v0
{
// clang-format off
/// \brief Axis-reverse operation.
......@@ -77,6 +79,7 @@ namespace ngraph
AxisSet m_reversed_axes;
};
}
namespace v1
{
......@@ -131,5 +134,7 @@ namespace ngraph
Mode m_mode;
};
}
// default opset version
using v0::Reverse;
}
}
src/ngraph/op/softmax.cpp
View file @ 166fb921
......@@ -53,14 +53,6 @@ op::v0::Softmax::Softmax(const Output<Node>& arg, const AxisSet& axes)
input_shape,
").");
}
if (input_shape.is_static())
{
set_output_type(0, get_input_element_type(0), input_shape.to_shape());
}
else
{
set_output_type(0, get_input_element_type(0), PartialShape::dynamic());
}
// empty axes == all axes
if (m_axes.size() == 0)
......@@ -72,6 +64,19 @@ op::v0::Softmax::Softmax(const Output<Node>& arg, const AxisSet& axes)
}
}
void op::v0::Softmax::validate_and_infer_types()
{
const PartialShape& input_shape = get_input_partial_shape(0);
if (input_shape.is_static())
{
set_output_type(0, get_input_element_type(0), input_shape.to_shape());
}
else
{
set_output_type(0, get_input_element_type(0), PartialShape::dynamic());
}
}
shared_ptr<Node> op::v0::Softmax::copy_with_new_args(const NodeVector& new_args) const
{
check_new_args_count(this, new_args);
......@@ -128,7 +133,11 @@ op::v1::Softmax::Softmax(const Output<Node>& arg, const size_t axis)
") is out of bounds (argument shape: ",
input_shape,
").");
}
void op::v1::Softmax::validate_and_infer_types()
{
const PartialShape& input_shape = get_input_partial_shape(0);
if (input_shape.is_static())
set_output_type(0, get_input_element_type(0), input_shape.to_shape());
else
......
src/ngraph/op/softmax.hpp
View file @ 166fb921
......@@ -43,6 +43,8 @@ namespace ngraph
///
Softmax(const Output<Node>& arg, const AxisSet& axes);
void validate_and_infer_types() override;
virtual std::shared_ptr<Node>
copy_with_new_args(const NodeVector& new_args) const override;
......@@ -79,6 +81,8 @@ namespace ngraph
///
Softmax(const Output<Node>& arg, const size_t axis);
void validate_and_infer_types() override;
size_t get_version() const override { return 1; }
virtual std::shared_ptr<Node>
copy_with_new_args(const NodeVector& new_args) const override;
......
src/ngraph/pass/opset0_downgrade.cpp
View file @ 166fb921
......@@ -15,8 +15,10 @@
//*****************************************************************************
#include "ngraph/pass/opset0_downgrade.hpp"
#include "ngraph/graph_util.hpp"
#include "ngraph/op/constant.hpp"
#include "ngraph/op/get_output_element.hpp"
#include "ngraph/op/pad.hpp"
#include "ngraph/op/reverse.hpp"
using namespace std;
using namespace ngraph;
......@@ -87,6 +89,38 @@ bool pass::Opset0Downgrade::run_on_node(shared_ptr<Node> node)
modified = true;
break;
}
case OP_TYPEID::Reverse:
{
auto tmp = as_type_ptr<op::v1::Reverse>(node);
auto axes_node = tmp->input_value(1).get_node_shared_ptr();
NGRAPH_CHECK(axes_node->is_constant(),
"Unable to convert Reverse:v1 to Reverse:v0 "
"if reduction axes are not constant. Node: ",
*node);
const auto axes_node_const = as_type_ptr<op::Constant>(axes_node);
AxisSet axes{};
if (tmp->get_mode() == op::v1::Reverse::Mode::INDEX)
{
axes = axes_node_const->get_axis_vector_val();
}
else // Mode::MASK
{
auto axes_mask = axes_node_const->get_vector<bool>();
for (size_t i = 0; i < axes_mask.size(); ++i)
{
if (axes_mask[i])
{
axes.emplace(i);
}
}
}
auto replacement_node =
make_shared<op::v0::Reverse>(node->input(0).get_source_output(), axes);
replace_node(node, replacement_node);
modified = true;
break;
}
default: break;
}
#if defined(__clang__)
......
src/ngraph/runtime/cpu/cpu_emitter.hpp
View file @ 166fb921
......@@ -29,6 +29,7 @@
#include "ngraph/op/min.hpp"
#include "ngraph/op/pad.hpp"
#include "ngraph/op/product.hpp"
#include "ngraph/op/reverse.hpp"
#include "ngraph/op/sum.hpp"
#include "ngraph/runtime/cpu/cpu_external_function.hpp"
#include "ngraph/runtime/cpu/cpu_tensor_view_wrapper.hpp"
......@@ -127,7 +128,6 @@ namespace ngraph
class QuantizedMaxPool;
class QuantizedAvgPool;
class MaxPoolWithIndices;
class Reverse;
class ReverseSequence;
class MaxPoolWithIndicesBackprop;
class Erf;
......
src/ngraph/runtime/cpu/kernel/gather.hpp
View file @ 166fb921
......@@ -85,7 +85,6 @@ namespace ngraph
static_cast<ElementType*>(inputs), in_dims);
auto indices_ptr = static_cast<IndicesType*>(indices);
IndicesType index_value;
auto indices_rank = indices_shape.size();
auto outer_loop_num = 1;
for (size_t i = 0; i < axis; i++)
......@@ -125,7 +124,7 @@ namespace ngraph
// at axis
in_extents[axis] = 1;
// at axis, get the value from indices arg
index_value = indices_ptr[0];
IndicesType index_value = indices_ptr[0];
// take care of negative indices
in_offsets[axis] =
index_value >= 0 ? index_value : index_value + axis_length;
......@@ -200,7 +199,7 @@ namespace ngraph
}
// at axis, get the value from indices arg
int k = i % num_indices;
index_value = indices_ptr[k];
IndicesType index_value = indices_ptr[k];
// take care of negative indices
in_offsets[axis] =
index_value >= 0 ? index_value : index_value + axis_length;
......
src/ngraph/serializer.cpp
View file @ 166fb921
This diff is collapsed.
test/mlir/affine_conversion/core_ops.mlir
View file @ 166fb921
......@@ -5,13 +5,34 @@
// -----
// Gather Op
// CHECK: affine.for [[i:%.*]] = 0 to 16 {
// CHECK: [[L0:%.*]] = affine.load %{{.*\[}}[[i]]{{\]}}
// CHECK: [[GATHER_IDX:%.*]] = index_cast [[L0]]
// CHECK: affine.for [[j:%.*]] = 0 to 32 {
// CHECK: [[VALUE:%.*]] = load %{{.*\[}}[[GATHER_IDX]], [[j]]{{\]}}
// CHECK: affine.store [[VALUE]], %{{.*\[}}[[i]], [[j]]{{\]}}
// CHECK: affine.for %[[I:.*]] = 0 to 16 {
// CHECK: %[[L0:.*]] = affine.load %{{.*}}[%[[I]]]
// CHECK: %[[GATHER_IDX:.*]] = index_cast %[[L0]]
// CHECK: affine.for %[[J:.*]] = 0 to 32 {
// CHECK: %[[VALUE:.*]] = load %{{.*}}[%[[GATHER_IDX]], %[[J]]]
// CHECK: affine.store %[[VALUE]], {{.*}}[%[[I]], %[[J]]]
func @simple_gather(%arg0: !ng.tensor<16x!ng.i64>, %arg1: !ng.tensor<512x32xf32>) -> !ng.tensor<16x32xf32> {
%0 = "ng.gather"(%arg1, %arg0) {axis = 0 : i64} : (!ng.tensor<512x32xf32>, !ng.tensor<16x!ng.i64>) -> !ng.tensor<16x32xf32>
"ng.return"(%0) : (!ng.tensor<16x32xf32>) -> ()
}
// -----
// Dot Op
// CHECK: affine.for %[[I:.*]] = 0 to 16
// CHECK-NEXT: affine.for %[[J:.*]] = 0 to 32
// CHECK-NEXT: affine.store %{{.*}}, %[[RESULT:.*]][%[[I]], %[[J]]]
// CHECK: }
// CHECK-NEXT: }
// CHECK: affine.for %[[K:.*]] = 0 to 16
// CHECK-NEXT: affine.for {{%.*}} = 0 to 8
// CHECK-NEXT: affine.for %[[M:.*]] = 0 to 32
// CHECK: affine.load
// CHECK: affine.load
// CHECK: mulf
// CHECK: %[[R:.*]] = addf
// CHECK: affine.store %[[R]], %[[RESULT]][%[[K]], %[[M]]]
func @simple_dot(%arg0: !ng.tensor<16x8xf32>, %arg1: !ng.tensor<8x32xf32>) -> !ng.tensor<16x32xf32> {
%0 = "ng.dot"(%arg0, %arg1) : (!ng.tensor<16x8xf32>, !ng.tensor<8x32xf32>) -> !ng.tensor<16x32xf32>
"ng.return"(%0) : (!ng.tensor<16x32xf32>) -> ()
}
test/opset_pass/reverse_opset_pass.cpp
View file @ 166fb921
......@@ -19,18 +19,19 @@
#include "ngraph/ngraph.hpp"
#include "ngraph/pass/manager.hpp"
#include "ngraph/pass/opset0_downgrade.hpp"
#include "ngraph/pass/opset1_upgrade.hpp"
#include "util/type_prop.hpp"
using namespace std;
using namespace ngraph;
TEST(serialize, opset1_reverse_upgrade)
TEST(opset_upgrade, opset1_reverse_upgrade)
{
const auto data = make_shared<op::Parameter>(element::f32, Shape{2, 2, 2});
const AxisSet reverse_axes{1, 2};
const auto reverse_v0 = make_shared<op::Reverse>(data, reverse_axes);
const auto reverse_v0 = make_shared<op::v0::Reverse>(data, reverse_axes);
const auto result = make_shared<op::Result>(reverse_v0);
auto f = make_shared<Function>(ResultVector{result}, ParameterVector{data});
......@@ -50,3 +51,78 @@ TEST(serialize, opset1_reverse_upgrade)
// should match the number of elements of v0::Reverse reverse_axes attribute
EXPECT_EQ(rev_axes_input_shape, Shape{2});
}
TEST(opset_downgrade, opset0_reverse_downgrade_index_mode)
{
const auto data = make_shared<op::Parameter>(element::f32, Shape{2, 2, 2});
const auto reverse_axes =
make_shared<op::Constant>(element::i64, Shape{2}, vector<int64_t>{1, 2});
auto mode = op::v1::Reverse::Mode::INDEX;
const auto reverse_v1 = make_shared<op::v1::Reverse>(data, reverse_axes, mode);
const auto result = make_shared<op::Result>(reverse_v1);
auto f = make_shared<Function>(ResultVector{result}, ParameterVector{data});
ngraph::pass::Manager pass_manager;
pass_manager.register_pass<pass::Opset0Downgrade>();
pass_manager.run_passes(f);
const auto pass_replacement_node =
f->get_result()->input(0).get_source_output().get_node_shared_ptr();
const auto reverse_v0 = static_pointer_cast<op::v0::Reverse>(pass_replacement_node);
EXPECT_EQ(reverse_v0->description(), "Reverse");
EXPECT_EQ(reverse_v0->get_version(), 0);
EXPECT_EQ(reverse_v0->get_reversed_axes(), AxisSet({1, 2}));
}
TEST(opset_downgrade, opset0_reverse_downgrade_mask_mode)
{
const auto data = make_shared<op::Parameter>(element::f32, Shape{2, 2, 2});
const auto reverse_axes =
make_shared<op::Constant>(element::boolean, Shape{3}, vector<bool>{true, false, true});
auto mode = op::v1::Reverse::Mode::MASK;
const auto reverse_v1 = make_shared<op::v1::Reverse>(data, reverse_axes, mode);
const auto result = make_shared<op::Result>(reverse_v1);
auto f = make_shared<Function>(ResultVector{result}, ParameterVector{data});
ngraph::pass::Manager pass_manager;
pass_manager.register_pass<pass::Opset0Downgrade>();
pass_manager.run_passes(f);
const auto pass_replacement_node =
f->get_result()->input(0).get_source_output().get_node_shared_ptr();
const auto reverse_v0 = static_pointer_cast<op::v0::Reverse>(pass_replacement_node);
EXPECT_EQ(reverse_v0->description(), "Reverse");
EXPECT_EQ(reverse_v0->get_version(), 0);
EXPECT_EQ(reverse_v0->get_reversed_axes(), AxisSet({0, 2}));
}
TEST(opset_downgrade, opset0_reverse_downgrade_axes_not_constant)
{
const auto data = make_shared<op::Parameter>(element::f32, Shape{2, 2, 2});
const auto axes = make_shared<op::Parameter>(element::boolean, Shape{3});
const auto reverse_v1 = make_shared<op::v1::Reverse>(data, axes, op::v1::Reverse::Mode::MASK);
const auto result = make_shared<op::Result>(reverse_v1);
auto f = make_shared<Function>(ResultVector{result}, ParameterVector{data, axes});
ngraph::pass::Manager pass_manager;
pass_manager.register_pass<pass::Opset0Downgrade>();
try
{
pass_manager.run_passes(f);
FAIL() << "Exception after Opset0Downgrade pass was not thrown.";
}
catch (const ngraph_error& error)
{
EXPECT_HAS_SUBSTRING(error.what(),
std::string("Unable to convert Reverse:v1 to Reverse:v0"));
}
catch (...)
{
FAIL() << "Reverse:v1 pass failed for unexpected reason";
}
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment