* remove caching of ordered_ops

* graph_util logging msgs

* small cleanup

* remove files for the TopologicalSort pass

* remove NGRAPH_DEBUG from graph_util.hpp

* much faster compile time
* Remove all variables and just directly access inputs, output, and temps.
* compare layouts when checking if two ops are equal
* make performance counters available to all backends

* general cleanup

* remove runtime::Value

* more cleanup

* more cleanup

* cleanup

* remove arg_index

* remove argno from Input

* uncleanup

* Simple boilerplate for GPU runtime files

  - GPUBackend
  - GPU ExternalFunction
  - GPUManager
  - GPUCallFrame

* Test for construction all GPU runtime classes

* Comment out calls, constructors haven't been defined

* Clang CUDA source example to later test compiling

Clang cuda example from:
https://gist.github.com/anonymous/855e277884eb6b388cd2f00d956c2fd4

* Initial nvptx compiler copied from CPU compiler sources

* Define FunctionMap and Instruction for gpu external function

* Rename Compiler -> NVPTXCompiler for gpu compile. Add call to compile for test

* Rename StaticCompiler -> NVPTXStaticCompiler for GPU code gen

* CAdd nvptx_compiler and nvptx_execution_engine to gpu sources

* Compiling source unit test using hardcoded PTX

* (a+b)*c test for GPU

* WIP Fix compile

* rmed accidentally included file

* Fix compile, and LLVM link errosr from nvptx_compiler.cpp

* Stub out parts needed for GPU manager

* Test GPU runtime method stubs

* Cleanup

* Add GPU runtime to same cmake block as GPU, include CUDA headers if GPU enabled

* Kill reflexive assertion

* change GPU naming convention to match CPU

* Snake case functions and identifiers in test case

* Change element type to match changes in master

* Make CUDA headers accessible for codegen with GPU transformer

* clang-format

* apply-code-format

* updated the sphinx version using pip install in Dockerfile.ngraph_cpp

added a make target to build the docs to the contrib/docker/Makefile

* avoid upgrade pip message during build

* Definition and type checking for max pool

* Implement kernel, integrate into INTERPRETER, add a few unit tests, make function result type mismatch error message more informative (still need to update tests to reflect that)

* Temporarily delete unit tests to ease merge

* Temporarily delete unit tests to ease merge

* Restore deleted unit tests

* Fix a broken error message check in the unit tests

* Update to handle various TensorViewType-related things going away; add NGVM support

* Add codegen case

* Change various get_blah_shape methods to return const refs, and while we're here, make a similar change where it should have been done in convolution

* Use NDArray for max-pool tests

* recreate ops

* style

* recompute ops

* style

* fix

* recreate ops

* style

* recompute ops

* style

* fix

* some

* more

* style

* remove a line

* const

* style

* NodeMap was using non-standard operator[] behavior.

* Missing include

* Function can have multiple results
Remove external use of ValueType, TupleType, Tuple
Remove many external uses of Output and Input

* corresponding CPU backend changes

* Update master changes.

* Remove type arg from Function, add changes.md

* Merge changes.

* Move bodies to .cpp, add brief doc

* Merge CPU changes.

* Remove xla includes from non-xla files

* Remove xla from tests

* First part of xla tuple support

* change fprop_cache to assume multi-output bprop functions

* New wrappers for handling tuples with XLA

* Review comments

* remove old xla files

* fix merge errors

* hand edit models to use multi output instead of tuples

* remove llvm/clang dependency in headers

* copy elision

* add larger test models

This avoids bloating .data and clears the path
for code model fixes later

* wip

* constants as globals

* const emitter rewrite

* CMake: TBB integration placeholder

* CMake: Integrate TBB

* CMake: Indent

* CMake: Rewrite TBB integration

* CMake: More TBB integration changes

* CMake: Install TBB headers and DSOs

* CMake: Don't install the TBB debug DSO

* CMake: Propagate ngraph's configured compiler setting over to MKL-DNN

* CMake: Restore TBB debug DSO installation

* CMake: Add installed headers to search path.
This needs to be cleaned up along with other header search cleanup

* CPU: Build and execute TBB flowgraphs

* CPU: TBB fixes

* CPU: More TBB fixes

* CPU: Allow both TBB and serial codegen for now

* TBB: get_arguments -> get_input_ops

* CPU: Use node methods

* CPU: Add TBB headers in the build directory to the search path

* TBB: Incorporate various changes from master

* CMake: Indentation fix

* CMake: Indentation fix

* CMake: TBB is mandatory so remove additional predicates

* TBB: Add a test

* CMake: Fix linker flags with GCC

* in progress

* working cache_fprop, no tests

* style fix

* all inputs to bprop (except adjoints) are cached from fprop

* fix typos, make sure to check count == 0

* fix code format

* cleanup

* cleanup

* expand

* wip

* undo

* enable -O3 optimization

* add flags to support release/debug builds

* nan unit test

* fix NAN issue

* add INFINITY support

This reverts commit 39383029.

It looks like the commit actually suppressed parallel makes
of MKL-DNN, at least in the case where ngraph itself was being
built with parallel make.  It also introduced problems with
make jobserver warnings.

* add resource file generator and store all headers used by codegen in memory.

* cleanup

* cleanup

* update serializer to emit small, simple element_type. backwards compatible.

* allow for selecting indenting when serializing

* remove ngvm

* remove NGVM from cmake

* fix autodiff on non-NGVM backends. NGVM initializes all tensors to zero on allocation while the other backends do not. Had to initialize vector before use.

* change autodiff tests to use INTERPRETER

set code model back to default as medium is causing the CPU.divide_by_zero_int32 unit test to sefault when it throws an exception from the generated code (#328)

* CPU: Optimize Eigen based rowwise vector broadcast

* CPU: Remove the need for transposing the broadcast vector

* CPU: Optimize to a replicate expression

* CPU: Change code model to medium and compile for the host CPU
instead of hardcoding BDW

* remove ParameterizedConstant
* use simpler element Type definition
* Move TraitedType to NGVM directory

* Adds CMake variables `MKLDNN_BUILD_COMMAND_EXTRA_FLAGS`
  and `MKLDNN_CMAKE_EXTRA_FLAGS`.

* aliased output unit test
* add support for aliased outputs to INTERPRETER and CPU