* refine OpKernelKey

* refine codes

* fix code style

* follow comments

* "change operator interface"

* "move devicepool to device_context"

* "fix operator test"

* "fix op_registry Run interface"

* "net op passed. Need to fix nccl multi-Context"

* "add nccl group function"

* "add nccl group function"

* "fix gpu count exceed 32 error"

* "fix recurrent op, nccl op"

* "change the other operators interface with Place"

* "fix typo"

* "fix pybind"

* "fix device in python side"

* "fix pybind failed"

* "add init for test"

* "fix CI"

fix doc.

Update annotations of layers.py.

fix doc.

fix doc.

Refine fluid recommender test

remove unused usage_stat script

add python wrapper for sequence_pool

Add ReorderLoDTensorByRank

* add data layout

* fix ci

Add python wrapper for reduce_mean

fix cmake require docs

[Done] API for dist train

fix the latex equation in fluid fc layer.

Enforce drop_empty_grad=false When the input of an op is duplicable.

* "remove GPU Sync Interface"

* "fix typo"

* "fix type cast error"

* "fix related Copy with stream"

* "fix failed tests with DevicePool"

* "fix stupid removed position error"

enable alexnet benchmark

use small samples to infer openblas

Fix lr setting of param_attr in Fluid

For input argument with a list of variables, drop_empty_grad is not allowed because it makes the correspondence bewteen a variable and its gradient ambiguous. Use REGISTER_OP_EX to register the op or call InputGrad(?,false) in GradOpDescMaker.

Refine the doc for mul_op and fully connected layer.