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

Enable P2P memory copy

On k40 with 4 devices, time reduces from ~4.0 to ~3.8+, should be
more obvious on better hardware

move two images

Limit the symbol table of fluid shared library.

Add distributed lookup table design

Remove unnecessary clone of program in C++ Executor.Run

Add float16 support for pool 2d operator

Add float16 support for cast op

add gserver to capi dep

rm unused private field in profiler

fix regularizer when gradient is None

Update cpp reader doc

delete unused images

Add float16 support for cudnn conv2d

remove legacy design image

Fix a program copy regression.

fix AttributeError: 'module' object has no attribute 'framework_pb2'

add math_function to selected_rows_functor dependency list

change the dir of docs

Single device on se-resnet reduce from 0.56 to 0.50

* New test for testing channels / concurrency

* Formatting fix