In multi-thread condition, EnableProfiler can
be called after RecordEvent is constructed. In this
case, RecordEvent constructor will not init anything,
but RecordEvent destructor will do something since EnableProfiler
was called.
This PR fixes it.

"fix dtype ci error"

* test_patch_mkldnn

* use patch file

* follow comments

* fix patch

Reusing of softmax mkldnn primitives

Fix order of prior boxes.

Balance parameter_opt between cards

matmul support float16/double

Parallel Executor revised feeder

* Update the mnist test cases with the new API syntax.

* Turn on the tests for MNIST

* delete the test files got merged accidently.

* Enable the mnist tests. ready for test driven development.

* Comment out the infer first
This is to confirm that the Trainer.train is working

* Add CMake file to include the tests

* Make the train program only return avg_cost for now

* Update the tests to use the latest syntax

typo fixed

* add Inference.infer

* optimize code

* update no_test_word2vec_new_api.py

* update trainer

* split check_and_get_place

* use inference_program to save inference model in Trainer

* update demo

* update save_inference_model

* clean code

- Added hash function inside of MKLDNN softmax op to be used as handle for primitives stroing in a
context

- Style fixes to softmax mkldnn op

- Fixes after review

- Coding style

- Fix to style

- style fixes

- style fix

- style fixes

- Fix to cody style check

- Rephrasing a comment

update mklml version, disable building tests and examples when install mkldnn

fix unittest-error: test_network_with_dtype

Fix two deadlinks in paddle/README.md

Fix compile on mac

Clean OpProtoAndCheckerMaker

disable building tests and examples when install mkldnn
skip -Werror=unused-result

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/matmul_support_float16_double

* a draft of trainer.test()

* polish trainer.test()

* polish trainer.test()

* update code format

* update

* polish code

* polish code

* polish code

* Make trainer.test follow the rule of returning [loss, metric, metric, ..]

* Add FP16 option in save_combine_op

* Fix casting issue