* fix numeric stability bug in autodiff of divide

* add a test for divide autodiff stability

* Basic assignment pass for CPU backend

* Change CPU convolution emitter to check op annotations

* Queries MKLDNN for optimal layout on convolutions

* Added layout conversions through ConverLayout ops and explicit layout conversion on CPU tensor view objects

* Added layout conversions for non-MKLDNN ops

* - Style fixes
- Removed unused variables to avoid clang errors
- Added more mkldnn format types to utility functions

* Move ConvertLayout back to runtime::cpu::op namespace

* Added more mkldnn memory formats

* Moved op annotations to Op class

* Style changes

* Minor fix

* Minor fix to keep clang happy

* Use ngraph element type instead of c_type_string in MKLDNN utility functions

* Addressed PR(#502) comments

* fix a segfault due to an unhandled op

* fix a missing new line

* generalize Relu

* format fixes

GPU codegen framework refactor

* added block_begin and block_end util to CodeWriter.

* refactored convolution backprop to use block_begin() and block_end()

* relu for interpreter

* relu in serializer

* core fusion

* relu backprop

* relu backprop and test interpreter

* core fusion for CPU

* COREFusion -> CoreFusion

* relu MKL dnn

Merge pull request #493

MKLDNN Emitter

* Add product op

* Add Max (max reduce) and Min (min reduce) ops

* Refactor arithmetic reduction ops to a common base class

* Fix PREFER_EIGEN codepaths in cpu_emitter

* fixed bn weights initialization to correct size

* style fix

mkldnn MaxPoolBprop implementation in CPU emitter