* adding constant propagation pass

* adding test/constant_propagation.cpp

* template make_constant_reshape function

* code review feedback

* add missing files

* serialize logic for reverse_sequence

* Added serializer support for Softmax

* Fix bug with concat for 0-size tensors

* Simplify test for zero-length axes, per PR comments

* [doc] Fix code snippet in derive-for-training

* Fix another code snippet in derive-for-training

* add assert to make sure we don't replace unreachable nodes

* fix unittest failures

* sparsity fix

* fix mkldnn rpath

* fix compile warning

* close backends when exiting

* set backend output directory of backends to the ngraph output directory

* Aprocter/patch patch (#1119)

* Move more rpath stuff inside if(NOT APPLE)

* fix repatch problem with mkldnn library

* add updated patch command for older versions of cmake

* loop kernel + tests

* remove commented out code

* remove commented code; add comments

* copy_with_new_args +test

* add comment

* fix comp errors

* Move to depth-first serialization of graph for better cache behavior

* Added comment

* Force 64 byte stack alignment to avoid crashes from unaligned AVX loads/stores

* Revert "Force 64 byte stack alignment to avoid crashes from unaligned AVX loads/stores"

This reverts commit 84346420fbd0fbd5d05a4a1e8f5fae12bdc7348b.

* revert to breadth-first serialization

DEX Part 2

* enable more gpu ops test

* add reverse_sequence

* fix bugs, compiled

* fix index bug

* fix bug and clang format

* correct function name

* clang format

* remove extra ;

* remove tests from skip list

* add backward support, skip tests

* add back template<> line

* remove unecessary lines in kernel

* optimized strided convolutions

* clean up debug messages

* format fixes

* more tests

* even more tests

* adapt to resnet-50.v1

* fix format errors; remove changes from diff PRs

* enable cse for reduction ops

* reduction tests

* - Added graph pass for fusing RNN op across layer
- Added test case for inter v/s cpu for verifying layer fused RNN
- more sanity checks in the RNN fusion graph pass
- added support to replace the recurrent cell state correctly in the fused RNN op

* Fixed multi layer rnn fusion unit test failure

* Addressed PR comments

* enable tests

* add funciton call

* working version

* remove test from ski list

* remove comments within body of function

* change to emit each op exactly once

* update code per review comments

Also, formatting fixes

* backend libraries now found in tree

dynamically read header search paths

fix running from install

*  group conv init

* add GroupConvolution op; refine checks in fusion logic

* add an emitter, cpu assigment

* cpu_layout

* add checks to algebraic simplification

* updating emitter logic for groupconvolution

* working before refactoring

* moving primitive creation logic to mkldnn_emitter

* group convolution graph test

* rename an opt

* address jbobba's feedback

* add pad_dilation function

* add dilation to gpu_emitter

* add CoordinateDiff constructor to GPUShape

* remove unecessary cast

* working version for forward

* forward working

* forward test all pass

* deconvolution forward

* backward data dilation

* forward test passed

* initial to 0

* fix bug for get_padded_shape and clang format

* code style, change variable names

* refactor convolution conditions

* fix bug padding_below_diff

* change pad_dilation to pad_dynamic, compare to pad

* remove passed convolution test from skip list, clang format

* change pad to use GPUShape