* WIP

* Fix incorrect CK output adjustment

* Bug fix and enroce sanity check

* Change cycle search depth, and fix sanity check

* cpu_fusion.validate_fuse_gru_inputs passes.

* Fix as_single_output to be able to always create a GOE

* minor fix. style-apply

* Clean up debug msgs

* Switch to backward cycle check

* Enable failing test

* PR fixes

* Address feedback: Add fwd cycle checks. Make cycle checking depth configurable

* [MLIR] Bump MLIR repo 8/20/2019

MLIR_
commit 0cdb20a6add19bc96c20dad28589a1e54e4d8469
Author: Lei Zhang <antiagainst@google.com>
Date:   Tue Aug 20 13:33:41 2019 -0700

    Add spv.specConstant and spv._reference_of

LLVM:
commit 3b9a27b6908040881dad394022f8c472c15c0784
Author: Simon Pilgrim <llvm-dev@redking.me.uk>
Date:   Tue Aug 20 17:54:37 2019 +0000

    Fix typo in comment. NFCI.

* [MLIR] Disable three more tests with MLIR enabled.

This PR disables validate_fuse_gru_inputs, reshape_layout_optimizations4
and reshape_layout_optimizations5:
  1. trivial_in_place_relu_fail: It checks tensors pool offset. There is
     no memory pool in MLIR atm.
  2. validate_fuse_gru_inputs: It creates an infinite cycle in
     MLIR subgraph extraction pass (under investigation).
  3. reshape_layout_optimizations4/5: They fail due to CompiledKernel
     being not expected by CPULayout pass.

* Disable cpu_quant_fusion.qconcat

* - Add graph pass method for onnx lstmcell rewrite with lstm cpu op
- insert reshapes to keep the weights in ldigo format
- test case for onnx LstmCell to CPU Lstm

* fix typo

* - check LSTMCell for the fused op decomposistion in the backend

* - fix bug in onnx_lstm graph pass
- passes unit test

* style-fix

* - fix compilation error
- use IFCO gate ordering for bias

*  - Skip LSTMCell to LSTM CPU fusion for peephole

* - add comment && remove duplicate function

* -use dynamic_pointer_cast to check for constant

* - onnx bias will be of shape (2 * gates_count * hidden_size) bias of Wb and Rb are concatenated, we will split the bias, add and rearrange in order IFCO

* - Use most derived LSTM ctor for pattern matching

* - Style Fix

* style fix

* Address PR comments

* - add support for graph pass (MKLDNN version > 1) for mapping LSTMCell -> LSTM CPU op

* fix unit test failure for MKLDNN V1.0

CPU fusion is disabled in MLIR since fused ops are not supported in
nGraph dialect. CPU debug tracer test doesn't expect CompiledKernel ops
generated for MLIR.

* Convert CPU ops to new form

Remove obsolete sigmoid

* Fix export

* fix typo

* typo

Add comments.

Add more unit tests.

Add shape check in fusion.

Rename function.

* Modify QuantizedConvolutionBias(Signed)Add fusion.

* Add unit test.

* Initial implementation

* Added test case

* Bug fix; Dropout with 2 outputs, WIP

* Fixed in unit-testl; WIP for model

* Nothing is working

* Revert "Nothing is working"

This reverts commit d3ff09bb7a0d0519ab70ac85f2e7f30721afea96.

* Fixed unit-test; fusion with 2 outputs

* Fix style check, file permissions

* Changed input arg to Node

* Fix order of declaration

* Improved performance

* some cleanup

* Fixed CI error

* Fixed review comments

* Fix CI error

* Remove unused variable

* Fix other CI errors

* Changed type

* Fix style check

* Add codegen code for Dropout

* addressed PR feedback; will add codegen support later

* Cleanup; change variable name

* Support for use_seed

* Add setter for use_seed

* Add setter for use_seed

* Fix CI error

* Make use_seed as arg

* Fix CI error

* Fix CI error

* Change serializer disable so that it is enabled by default if you do nothing

* fix serializer disable

* fix disabled build

* add BatchMatMul unit test to all builds

* Check that interpreter is available

* fix error

* enable tests for features not yet implemented

We want to use ComputedKernel for any target to delimit sub-graphs to be
compiled and executed with MLIR.

* [Fused] LeakyRelu op

* Add LeakyRelu to serializer

* Add unit tests

* Fix merge branch 'master' into mkarzyns/fused_leaky_relu

* Change broadcasting rules to NumPy style

* Remove std:: and ngraph:: prefixes

* Rename CPU Runtime LeakyRelu to CPULeakyRelu

* Style apply

* Fix cpu_fusion.fuse_leaky_relu test

* Use eigen's tanh in the fused sigmoid multiply kernel (#2946)

* Merge branch 'master' into mkarzyns/fused_leaky_relu

* Add LeakyRelu to Intel GPU backend op list

* Add LeakyRelu to Intel GPU backend op list

* Move zero padded conv fusions from CPUFusion to CoreFusion.

* Address PR feedback: move unit tests to core_fusion.

* update visualize tree file extenstions and output formats

* fix runtime error

* constexpr ctor for EnumMask

* added pass properties to core passes.

* change fusion type to have better type safety.

* refactor to use enum mask.

* remove extra code.

* added constants for FusionType backward compatibility.

* spelling.

* grammar fix.

* Fix clang compiler warnings

* Remove unintended file.

* style

* Not part of PR

* Another extra closure ref

* More warnings from merges

* Lambda arg was used

* Moving GroupConvolution to ngraph fused opset

* style fix

* remove unused function

* IntelGPU backend: Add GroupConvolution operation into main switch

Added NGRAPH_JSON_ENABLE to conditionally enable code/features relying on external JSON libraries (#2821)

* deconv optimizations for dcgan

* Added test cases

* modified some tests, not working at this point

* Removed temp code

* fixes to get unit test to pass

* Added node validation checks

* Update mkldnn emitter to memory reuse design

* Code cleanup

* Fix to enable deconv select the right kernel

* Fix file permissions

* Disabled unit test cases

* Remove unused variable

* Address PR feedback

* Removed dead code

* Style check

* removed dead code

* - Moves some fused convolution ops to core FusedOps
- Adds support for decomposing and replacing multi-output FusedOps
- Adds query callbacks to FusedOpDecomposition to check if a FusedOp is
  supported by a backend
- Adds core fusion patterns for FusedOps
-

* style fix

* Added comments on FOP_FUSIONS

* gpu convolution 1d bug fix (#2741)

* Fix bug with dex-only compilation and addressed PR comments

* Add NodeInput and NodeOutput classes

* Deprecate Node::get_inputs, Node::get_outputs, Node::get_output_inputs. Remove Node::get_input_from and Node::get_output_from

* Privatize most fields of Node

* Make deprecation of descriptor-munching classes optional

* Review comments

* Adapt ReshapeSinking to use raw pointers for NodeInput

* Fix ZDTE (thought I had already done in this branch, weird); style

* wip

* Change get_node_outputs() and get_node_inputs() to return vectors

* Updates after merge

* Whoops, forgot to define these functions

* {NodeInput,NodeOutput} -> {Input,Output}

* Kill shared_ptr in Output

* Move Input and Output into node.hpp

* Templatize the underlying node (sub)type in Input and Output

* Eliminate some get_input_* and get_output_* functions

* Change get_outputs and get_inputs back to their original names; rename NGRAPH_DEPRECATE_IO_DESCRIPTORS to NGRAPH_DEPRECATE_OLD_NODE_APIS

* Miscellaneous cleanup

* More cleanup

* Unbreak CPU build

* Simplify unit tests

* Make Node less friendly

* Deprecate more get_output_* and get_input_* functions

* A couple of PR comments

* Make the deprecation stuff more generally available

* Better comment

* Be more consistent about [] vs. at

* batch dot WIP.

* cpu backend refactor and unit tests pass.

* WIP.

* batch dot interpreter impelementation.

* minor clean up.

* more clean up.

* patching the gpu backends.

* added more tests, fixes, etc.

* fixed compile error.

* renamed batch dot to batch matmul.

* refactor WIP.

* fixes some tests and formating.

* more fixes.

* Added extra case for conv bias affine folding

* Added unit test

* Check for fused ops in unit test

* Add support for negative padding

* Use std::bind in pad builder check

* Add support for negative padding in CPU backend

* Updated kernel to do pad+slice

* Remove type conversion warnings

* Fix review comments

* Remove interior padding from core op and interpreter stuff

* Update backends other than GPU for retirement of padding_interior

* Skeleton of support for edge/reflect padding

* Post-merge cleanup

* Attempt reference implementation for EDGE.

* Fix the edge-padding reference, and add some unit tests

* Implement REFLECT padding ref; add tests

* Fixes to the CPU stuff so it compiles now

* Fix test

* Add support for different pad modes

* Restore a stub get_padding_interior function, and tweak some stale comments

* Update ONNX importer to not supply interior padding value; add checks for padding-too-small for EDGE and REFLECT

* Typo

* Bop a warning

* Attempt fix to INTELGPU backend

* Attempt another fix to INTELGPU backend

* Fix pyapi

* Style apply

* Add support for padding modes

* Remove unnecesary node validation checks

* Remove tests for minimal reflect and edge pad

* Remove commented tests

* Remove unnecesary Asserts

* Little update of pad documentation

* Monospace for pad_mode options

* Revert "Remove tests for minimal reflect and edge pad"

This reverts commit 81e4787ea47195b832cab1452dde698bc05776fe.

* Revert "Remove unnecesary node validation checks"

This reverts commit 7e68db7564f3c9b1fd40e7db1d1bda4e0677cad9.

* Test only spatial dims

* axis -> spatial axis

* Fix typo

* Style check

* Update test

* Add CoordinateDiff include

* Remove pad_mode from tree visualization

* Convert padding into NVShape

* Skip failing tests on GPU

* Revert mode change

* Remove merge artifact

* Rename pad kernel into pad_ref

* change float comparisons from == to all_close_f

* style

* address a few more direct float comparisons

* add missing include

* specify tightest tolerance for Broadcast and Reshape tests

* Increased tightness of float testing

Increased tightness of float testing via MIN_FLOAT_TOLERANCE_BITS parameter

* style

* Add QuantizedConcat

* Remove unused variables and add check for size of mins and maxes vector

* Resolve conflicts

* Merged with master and addressed some PR feedback

* Maxpool and Avgpool fusions. Exclude Q from conv+relu fusion

* Remove single-user check from fusions

* Quantized concat fusion

* workaround: do reshape sinking by default

* style fix

* check scales for QuantizedConcat

* use compare_constants

* remove stale comment

* Handle all concat cases from arg size 2 to 6

* addressed feedback

* - Added reorder support for rnn weights_layer/iter

* i) fixed compilation issues ii) working but still observing precision error

* i) fixed failing rnn unit test for DEX ii) refactored workspace in RNN mkldnn emitter

* i) added support for src reorder to TNC from NTC

* reorder support for rnn output fron NTC to TNC

* - added support for rnn weight reorder ldgoi -> ldigo
- code refactor for lstm/rnn kernel in mkldnn emitter

* - refactor rnn mkldnnn kernel, change variable names

* fix RNN codegen kernel

* disbale layer rnn fusion pass, to test CI

* method to validate recurrent rnn inputs

* add correlated macthes for Recurrent RNN PM

* - simplify reorder logic for rnn_weights
- fix graph pattern for fusing rnn cell across time steps

* do weights reorders in rnn timesteps fusion

* refactored LSTM graph pass

* - Bug fix for finding the lstm inputs determenstically
- Refactored LSTM graph pass to single pass
- made changes to LSTM RNN time step fusion graph pass

* - use replace_node instead of replace_output in Lstm_step_wise fusion graph pass

* fix compilation error

* Fix GNMT rnn fusion

* check if the node is in use before replacing in RNN graph passes

*  i) fix style ii) fix topo sort issue in RNN graph pass

* style fix

* fix bug in simplify_concat pass

* replaces Lstm1 -> {GOE1, GOE2} -> {Slice1, Slice2} -> Concat -> Lstm2 with Lstm1 -> Lstm2

* cse for convert layout

* addressed PR comments

* - optimization pass to remove  Lstm1 -> {GOE1, GOE2} -> {Slice1, Slice2} -> Lstm2
- conditional fusing of LSTM cells only for the decoder

* made changes to multi layer RNN fusion callback

* fix asserts in RNN op

* - added support to fuse layers when slc=dlc for RNN cells
- bug fix on the sanity checks for RNN Op

* - support RNN layer fusion till slc = dlc
- bug fixes in multi layer rnn fusion call back

* capture reshape in the RNN weights

* Addressed PR comments

* - added comments in multi layer PM call back
- fuse only if slc == DLC across layers

* restore deleted 3_lstm_cell_forward.json file

* fix typo

* fix failing unit tets

* When processing in place slice, do not change the offset of the slice node if the argument pointer comes from function input.

* Address PR feedback: process in place slice after propagating in place input.

* Set INTERMEDIATE role before propagating in place input.

* Do not add temporaries to the variable name map before propagating in place input in codegen.

* Fix a bug in codegen.

* Fix a bug in codegen slice.

* reenable disabled rnn unit test

* fix compiler error

* - bug fix in the slicing logic for the layer fused rnn cell
- fix failing rnn unit test

* - Addressed PR comments
- removed redundant checks from the rnn graph pass
- simplified rnn call back replace node logic

* - added new multilayer rnn *.json file
- fix test case

* [PRIVATE BRANCH] Style fixes (#2080)

* Style fixes

* change order of lstm gates

* WIP bi rnn

* [PRIVATE BRANCH] Jbobba/rnn fusion review (#2113)

* Style fixes for single-layer RNN fusion

* Style fixes to multi-layer RNN

* added callback routine for bi-directional rnn

* fix rnn op ctor, rnn mkldnn emitter to accomodate bi directional rnn

* style fix

* added helper function for rnn's to query direction and cell_type

* fix clang error

* - unit test case for bi rnn fusion
- style fix

* - updated bi-rnn graph pass to handle reverse and reverse_seq ops in the predicate
- added bi-rnn inter v/s cpu unit test case
- add support to in mkldnn_utils to create_md with tnc/ntc format

* - added enum type to deduce rnn_type

* Addressed PR comments
    - handle reshapes from {t, n, c} to {n, t, c} in the graph pass

* fix style

* fix clang error

* fix style

* i) move enum specific to rnn to seperate header

* use calls for new backend API in unit tests

* fix compile error

* fix compile error

* CPUQuantFusion pass and some usions for converting mixed precision sub-graphs to int8 fused ops

* - Added unit tests and misc bug fixes for mixed-precision fusions
- Adjust fused sum_scale in quantization builders instead of mkldnn
  primitive creation