* test=develop, add gradient sort backward strategy

* test=develop, fix test by add FLAGS_cudnn_deterministic on new tests

* test=develop, fix memory leak in dygraph mode

* test=develop, fix memory leak in dygraph mode

* test=develop, polish code

* test=develop, polish code

* test=develop, polish code

* add set_not_owned function for graph

* add scope set. test=develop

* add scope_ptr enforce not null before setting.test=develop

* add save/load model, shrink table, cvm, config file & fix pull dense bug
test=develop

* fix global shuffle bug, fix pull dense bug, fix release memeory bug, fix shrink error
add client flush, add get data size
test=develop

* fix global shuffle bug
test=develop

* fix global shuffle bug
test=develop

* fix code style
test=develop

* fix code style & modify pslib cmake
test=develop

* fix error of _role_maker
test=develop

* fix code style
test=develop

* fix code style
test=develop

* fix code style
test=develop

* fix code style
test=develop

* fix code style
test=develop

* fix windows compile error of fleet
test=develop

* fix global shuffle bug

* add comment
test=develop

* update pslib.cmake
test=develop

* fix fill sparse bug
test=develop

* fix push sparse bug
test=develop

* make parallel_executor support FLAGS_use_mkldnn

test=develop

* add warning when set mkldnn_enabled_op_types_ in non-mkldnn env

test=develop

* test=develop, add gradient sort backward strategy

* test=develop, fix test by add FLAGS_cudnn_deterministic on new tests

* reset drop local scope counter
test=develop

* Add conv2d_grad_grad_op
* Extracte the cuDNN conv algo searching code in conv_cudnn_helper.h.
    - Now use it in conv2d_grad_grad.
    - Will simply the searching code in conv2d and conv2d_grad in next PR.
* Enhance and fix bug in unit testing of gradient_checker.
* Support to fetch empty variables，return None in Python.

Fix the following API examples:

paddle.fluid.scope_guard
paddle.fluid.backward.append_backward
paddle.fluid.cpu_places
paddle.fluid.cuda_pinned_places
paddle.fluid.cuda_places
paddle.fluid.in_dygraph_mode
paddle.fluid.CUDAPlace
paddle.fluid.CPUPlace
paddle.fluid.CUDAPinnedPlace

* move pass to ir

* polish code
test=develop

* fix dependency
test=develop

* cherry-pick commit from 88770542

* cherry-pick commit from 3f0b97df

* cherry-pick from 16691:Anakin subgraph support yolo_v3 and faster-rcnn

(cherry picked from commit 8643dbc2)

* Cherry-Pick from 16662 : Anakin subgraph cpu support

(cherry picked from commit 7ad182e1)

* Cherry-pick from 1662, 16797.. : add anakin int8 support

(cherry picked from commit e14ab180)

* Cherry-pick from 16813 : change singleton to graph RegistBlock
test=release/1.4

(cherry picked from commit 4b9fa423)

* Cherry Pick : 16837 Support ShuffleNet and MobileNet-v2

Support ShuffleNet and MobileNet-v2, test=release/1.4

(cherry picked from commit a6fb066f)

* Cherry-pick : anakin subgraph add opt config layout argument #16846
test=release/1.4

(cherry picked from commit 8121b3ec)

* 1. add shuffle_channel_detect

(cherry picked from commit 6efdea89)

* update shuffle_channel op convert, test=release/1.4

(cherry picked from commit e4726a06)

* Modify symbol export rules

test=develop

* fix tensor_py,test=develop

* change class name,test=develop

* fix mem leak when converting Tensor to numpy array
test=develop

* remove unused unittest,test=develop

* follow comments, test=develop

* fix dygraph bug,test=develop

implement dygraph.parallel.DataParallel to hook reduce op.

* add doc for memory_optimize, test=develop

* update doc, test=develop

* doc update, test=develop

* Speed affine_channel_op unit testing
* Add check in tensor_py
* Fix ONLY_CPU Compiling

test=develop

* speedup gc and inplace softmax_with_cross_entropy_grad
test=develop

* refine models gpu mem
Merge skip vars and warning messages of mem opt
remove relu mem opt
test=develop

* follow comments
test=develop

test=develop

test=develop

* Add an option to enable the cache of expected kernel in train phase.
test=develop

* Change the default value of cache_expected_kernel to true.

* test fuse all_reduce

* Try to enable the runtime_context_cache pass in train phase.

* Put the append of runtime_context_cache pass ahead of multi_dev passes.
test=develop

test=develop

add NCCLParallelContext for parallel dygraph

test=develop

* expose fuse broadcast ops

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop