Add Majel's Place concept to Paddle

[wangkuiyi]

Fixes #2090

Note: Please review and merge #2040 before this. In order to build the ported Majel code, I am using the development Docker image.

[gangliao]

@wangkuiyi It's better to replace boost using some light-weight implementation https://github.com/mapbox/variant since it's too heavy.

[gangliao]

https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/external/glog.cmake#L29 It's a good way to integrate external project via CMake.

[wangkuiyi]

About boost

I listed all boost packages that Majel depends on:

$ for i in $(du -a | grep '\.h$' | cut -f 2); do \
    grep 'boost::' $i; \
  done | \
 grep -o 'boost::[a-zA-Z_]*' | sort | uniq
boost::apply_visitor
boost::bad_get
boost::get
boost::python
boost::shared_ptr
boost::static_visitor
boost::variant

I am afraid that we cannot replace all of them using some lighter weighted versions?

[wangkuiyi]

I understand and agree that we should use CMake and its external projects. I didn't do it in this PR because I want to figure out what is the command line we need to build Majel.

[gangliao]

After this PR is merged, I will add a CMakeLists.txt.

[gangliao]

LGTM

[helinwang]

Why the above "==" operation check equal as device == o.device, but here checks if it's the same instance?

[wangkuiyi]

You are right! Is Greg nearby and can you ask him about the problem?

[helinwang]

My bad, I thought *this is this. In this case, it's taking the neg of result from == operator.

[gangliao]

@wangkuiyi @helinwang I think it's correct. Because *this == o will trigger inline bool operator==(const GpuPlace& o) So it's still compare device id.