Created by: wangkuiyi

This issue records an idea inspired by reading https://github.com/PaddlePaddle/Paddle/pull/2489/.

Majels' Style -- Extremely Deep Tree

It seems that Majel's philosophy of making each module a directory is pretty complex and makes it difficult to browse the source code. For example:

• majel/include/majel/contrib/support/pinning.h only declares two global functions pin_memory and unpin_memory, and whose implementations are in
• majel/include/majel/contrib/support/detail/pinning-inl.h, and pin_memory's body just calls mlock and unpin_memory calls munlock.

Caffe2 and MXNet's Style -- Extreme Shallow Tree

Most core concepts of Caffe2 are in https://github.com/caffe2/caffe2/tree/master/caffe2/core. There are 78 files in this directory -- too many to browse in a glance.

The header files of most core concepts of MXNet are in https://github.com/dmlc/mxnet/tree/master/include/mxnet. Implementations (.cc files) are in https://github.com/dmlc/mxnet/tree/master/src, which adopts TensorFlow's Style as described below:

TensorFlow's Style -- In-between Extremely Deep and Shallow

Comparing with Caffe2, TensorFlow has a second level directory:

• tensorflow/core/platform
• tensorflow/core/framework
• tensorflow/core/...

and no header/source files in tensorflow/core.

A Proposal for Paddle:

1. Represent each module by a combination of x.h, x.cc/cu, and xx_test.cc/cu, so we could
2. have two-level source tree like TensorFlow:
   1. paddle/core/platform -- class Place, malloc_
   2. paddle/core/memory -- management and allocation, tensor type, etc
3. In each directory, if there are some facilities that are called only by code within the same directory, we put it into detail. For example:
   1. paddle/core/memory/detail/buddy_allocator.{h,cc,_test.cc}