* Cyphers/lastwarns (#3640)

* Build changes for -Wall

* Use properties

* Remove no-zero-as-null-pointer-constant

* add ##__VAR_ARGS__ workaround

* enable gnu-zero-variadic-macro-arguments warning

* fix gnu-zero-variadic-macro-arguments

* use PrintToDummyParamName

* remove ##__VA_ARGS__ workaround

* fix ##__VA_ARGS__ and enable gnu-zero-variadic-macro-arguments

* handle windows build

* use alternative fix to support VS compiler

* fix merge issue

* remove variadic macros from kernel_selectors

* replace remaining ##__VA_ARGS__

* fix missing braces

* Remove some clang warnings

* Warnings cleanup

* Switch op dynamic casts that crept back in

* PR for debugging purposes (#3718)

* back out

* transpose

* next try

* Another is_type

* Another is_type

* next is_type

* Revert, is_type remainder

* verify

* v1 and Dyn mixup

* Work around clang++-3.9 bug

* Implemented downgrade pass

* Using Pad:v1 in onnx_importer

* Downgrade transformation doc fixed

* Downgrade pass added for all backends

* Apply suggestions from code review

Changed pad_opset_pass to opset_downgrade
Co-Authored-By: Tomasz Socha <tomasz.socha@intel.com>

* Changed order of passes

* Changed downgrade pass order of CPU backend

* Code review remarks introduced

* add check for dynamic dimensions

* to_shape() -> rank()

* Delete unnecessary check

* Added in DeduceDynSlice suite to check dynamic dimension

* minor codestyle fix

* c-cast -> static_cast

Remove Intel GPU backend

* Added power constant folding

* Style apply

* Removed python operator pow

* Hotfix for negative axes in unsqueeze op (#3705)

* Hotfix for negative axes in unsqueeze op

* Review fix I

* Enable Gather with negative indices. (#3701)

* Enable Gather with negative indices.

* Address PR feedback.

* Remove GOE from Gather unit tests.

* merge error

* Disable test for MLIR

* More expressive error messages

* style

Enable in place rotation for Reshape for blocked data format.

* Added ReduceMean, ReduceMin, ReduceMax; Updated ReduceProd, ReduceSum

* Removed redundant method from arithmetic_reductions_keep_dims.cpp

* Fixed code style issues

* Removed generate_adjoints

* Fixed cpu_emitter

* [ONNX] Remove broadcasting from ONNX importer.

* [ONNX] Used Xor op instead of And,OR and NOT ops.

An optional optimization level was added to ExecutionEngine, initialized
to None by default. This caused fast-isel to be used in LLVM CG,
overriding all the flags related to not using fast-isel.

In this PR we pass the right optimization level to ExecutionEngine.

* Fix ops, cleanup ToC, add latest illustrations

* PR feedback

* Support non-mkldnn fallback for Batchnorm training bprop in CPU backend

* Skip unit test on PlaidML backend

* Change method of splitting data and filters for groups

* Add support for groups included in filters shape

* Add UT for groups in weights shape

* Add helper function to check if groups are in filters shape

* Review fix I

* Remove unused variables

* Skip test on PlaidML

* Review Fix II

* Internaly store groups as dimension instead of size_t

* Review fix III

* Add updating of m_groups in pre_validate_and_infer_types()

* Provenance for builders (#3644)

* Provenance for builders

* Tests

* Tests

* Update src/ngraph/node.hpp
Co-Authored-By: Sayantan Sarkar <sayantan.sarkar@intel.com>

* style

Finish migration

* Migration to master

* Handle op constructors that construct ops

* Update provenance for topk

* Add missing input (#3691)

* Add missing input

* If builder does nothing, don't dive into the inputs

* Reverse test

* simplify, add test

* Add test

* add test

* Add RandomUniform op

* Missing files

* Make CPU compile again (RandomUniform tests still don't pass)

* Bop users of existing RNGState class and reorder some comment junk

* Try to bop RNGState in gcpu

* Add RandomUniform to CPU manifest

* Add RandomUniform to PlaidML manifest

* Add .rst for random_uniform

* Clean up junk in the .rst

* Change UniformRNGState to always use double internally

* Change weird test failure message

* RandomUniform implementation for CPU backend

* Add missing file

* Add codegen support

* Code cleanup

* Address review comments

* Fix CI failure

* Fixed CI error

* Fixed CI error

* [MLIR] Add support for parsing nGraph tensor type

Initial commit that enables nGraph parsing. It's needed for testing.

* [MLIR] Enable nGraph dialect in ngraph-opt

This PR registers nGraph dialect in ngraph-opt and prepares
nGraph lowering pass for LIT testing, fixing all the related issues.
Among other things, lowering pass has to be turned into a function pass,
dead argument in constructor was removed and `convert-ngraph-to-affine`
flag was added.

* Fix issue with function name and multiple functions

* Extend module_function.mlir lit test

* [MLIR] Add support for parsing nGraph element types

It introduces initial support for parsing nGraph signed/unsigned
integer and floating point data types.

* [MLIR] Improve gather op lowering

This PR interchanges indices and param loops in gather lowering so
that a better memory access patter is generated. Fusion of gather
with other ops is also observed with this change.

* LayerNorm (#3630)

* Constructors.

Type prop.

Decompose LayerNorm.

Add serialize.

* Add dummy test case.

* Add dummy type prop test.

* Fix some build errors.

* Remove build errors.

* Update decompose for bprop.

* Change begin_norm_axis default value to 1.

* Style.

* Reorder class members.

* Add actual type prop tests.

* Add fprop test.

* Working on bprop test.

* Bprop tests.

* Allow flattened scale and bias.

* Add support for flattened scale and bias.

* Fix incorrect type_name.

* PlaidML: Decompose fused_op LayerNorm

* Update Backprop constructors.

* PlaidML: Add missing header file.

* Remove doc about removed param.

* Fix type prop tests.

* PlaidML: Disable unit test.

* Fix stats flattening axes bug.

* Upgrade description to type_info.

Aligned with contrib/mlir directory ownership. This prevents
that Scott is added as reviewer by default.

* [MLIR] Add support for parsing nGraph tensor type

Initial commit that enables nGraph parsing. It's needed for testing.

* [MLIR] Enable nGraph dialect in ngraph-opt

This PR registers nGraph dialect in ngraph-opt and prepares
nGraph lowering pass for LIT testing, fixing all the related issues.
Among other things, lowering pass has to be turned into a function pass,
dead argument in constructor was removed and `convert-ngraph-to-affine`
flag was added.

* Fix issue with function name and multiple functions

* Extend module_function.mlir lit test

* Improve module_function.mlir test

Remove ngraph to affine dialect conversion since we just need to verify
that we can parse and print modules and functions.
Add verification for parsing the printed code.

* [MLIR] Add support for parsing nGraph element types (#3665)

* [MLIR] Add support for parsing nGraph element types

It introduces initial support for parsing nGraph signed/unsigned
integer and floating point data types.

* Improve LIT tests

Test parsing and printing of types separately from lowering to affine
since these tests will evolve differently, particularly for tensor
types.

* Missed file

I left this file behind in the previous commit

* Instantiate only valid cases for binary op constant folding kernels

* removed unused parameter

* latest PlaidML-GPU updates for 0.26

* More theme updates

* Cleaner theme to match site

* Update ngraph_theme

* Update Release Notes page for 0.26

* update theme WIP

* Update theme w/feedback

* Add top navigation menu

* more theme customization and cleanup

* Refine more details on new ngraph theme

* Refine some xsys font-kerning

* Theme updates

* Add latest release notes

* Update HE transformer doc

* Fix rendering

* Resolve doc build warnings

* Add doc versions

* test code removed

* remove test code

LIT creates "Output" folders for each directory with LIT tests.
Before this change, LIT tests were executed in source directory and
"Output" folders had to be removed manually. Git also showed them as
folders to be potentially tracked. This PR changes LIT execution to
build directory to avoid this problem. Similar approach is used in MLIR.