transfomer on a single device step time
    reduces from 0.157 to 0.125

* Fix a critical bug in softmax_with_cross_entropy_op, which will lead to the wrong gradients.

* Enhance unit testing.

* Fix old documentation for channel_recv

* Initial design of CSP select

* Redesign channel implementation for Select Op

* Remove unecessary header

* Initial checkin of select op, currently will read all the conditional_op in the cases block and also pull out all channels involved in the select.

* Init python select op API

* Python select bug fix when checking op creates block

* Add case_to_execute as (a) input to select, (b) into the passed inputs into the select op

* Add in addition code for select op

* Init fibonacci test from python

* implement fibonnaci sequence test

* update fib unit test

* Improve select test cases

* Shorten non-pep-8-ed lines

* Add methods on channel needed by select op

* Fix compile issues, finish implementation, still need to debug code

* Fix issue with fibonncci test, it works now!

* Change QueueMessage callback to take in an ChannelAction enum, fix select unit test

* Fix case attributes

* Fix issue with select control flow

* Make cases - previously on each selectcase conditional_block - attributes to select

* Use class constants for type of channel

* Change select op to take in "cases" attribute

* return boolean from select callback function to tell Channel if this RECV or SEND should be executed

* Improve attributes and inputs comments on select op

* Fix issues with python unit test

* Assert fibonacci final output

* Fix issue when channel name / channel var is null for "default" case in select op

* Assert base select test output

* Make QueueMessage use shared pointer and modify the order of the callback

* Fixing the order in which the callback is called

* Move channel utility methods to paddle/fluid/operators/concurrency/channel_util

* Create channel_util and move channel util methods

* Fix crash when calling select_op

* Fix deadlock

* Fix issue of channel destructor deadlock

* Fix precommit issues

* Accidentally checked in changes to beam_search_op, reverting change.

* Fix dependency issue in concurrency cmake

* add device_context dependency for concurrency target

* "add softmax cudnn functor support"

* "add testing"

* "refine cmakelist"

* "sequence softmax forward speed up"

* "add softmax grad"

* "fix sequence softmax test"

* "add double precision'

* "fix softmax test"

* "add softmax cudnn support"

* "fix softmax cudnn test"

* "add softmax to nn.py"

* "fix compile bug"

* "refine cmakelist"

* "fix ci"

* "fix based on comment"

* "fix based on comments"

* "fix ci"

* add fp16 mul op support

* small fix

* fix bug

* small fix

* fix PADDLE_WITH_CUDA compiling issue

* reorg code

* test for pybind

* treate as float16 as uint16_t in pybind

* bind np.float16 to paddle float16

* small fix

* clean code

* remove redundancy

* fix mul_op test

* address comments

* small fix

* add is_float16_supported func

* "exported scatter to python"

* Revert ""exported scatter to python""

This reverts commit 38745a62.

* "polish scatter and export to python"

* "Refine nccl op"

* "refine code "

* "refine nccl code"

* fix nccl op unit test

* fix build error

* format code

* refine nccl related unit test

* fix build error

* add setGPUData

* clean up

* follow comments

* rm test_nccl.cu

* follow comment

* rm wait

* add memopt level

* add opt level for image classification demo

* clean code

* add delete op

* clean code

* test machine translation demo

* clean code

* clean code

* skip fill constant with force cpu

* clean code

* clean code

* refine code

* clean code

* fix bug

* Clipping bbox in the mAP evaluator calculation.

* Fix bug in detection_output and mAP calculation in SSD.

* Fix bug in detection.py.

* Fix bug in test_detection_map_op.py.

* MKLDNN pool2d OP kernel added

* conv2d and pool2d MKLDNN kernels renamed

* MKLDNN conv2d kernel refactoring