* wip multi process multi gpu dist training

* workable for p2p

* update test=develop

* change back env name test=develop

* fix alloc init

* fix cpu build test=devlop

* fix mac tests test=develop

* refine code

* refine test=develop

* AsyncExecutor: C++ side

* Google naming conventions

* Rename MultiExecutor to AsyncExecutor

* pybind with async_executor

* Naming convention

* remove some flags and unused code

* add refactored file of async_executor and data_feed

* clear async executor interface and add data feed factory

* split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto

* Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch

* refine async_executor_refactor.cc

* add some files about datafeed

* Revert "add some files about datafeed"

This reverts commit 8ee8133a.

* Interface rework

* add MultiSlotDataFeed

* Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side

* update data_feed for add MultiSlotDataFeed

* update datafeed and async_executor to run bow_net demo

* fix bug that finish_set_filelist failed in multithread

* delete finish_binding_memory_(flag), because it can not be marked under the current interface

* Fix bug

* update async_executor.py for support set_use_slots

* update async_executor.py for support set_use_slots and set set_dense_slots

* fix bug that when the number of files is less than the number of threads, it will fetch nan

* remove redundant code, and make executor exit when set a illegal queue size

* add batch_size check

* add MultiSlotDesc

* Revert "add MultiSlotDesc"

This reverts commit 2e72ebfa.

* add some checkpoint in DataFeedDesc

* add CheckFile function in MultiSlotDataFeed

* update something error info

* fix deaded lock bug

* Fix fetch variable

* Merge error

* fix code style in async_executor

* using one lock blocking queue replace two lock blocking queue because of some bugs

* update code style

* add utest for data_feed

* Fix fetch var

* update utest for data_feed for multithread

* update SetFileList info

* fix bug in utest of data_feed

* Add comments for python

* Add comments for python code

* Fix pybind.cc with new pybind11 version

* add note for DataFeedDesc's set_use_slots function

* Add save_model

* update data_feed_test for multi-type

* add comment for executor_thread_worker

* Remove unused code

* update data_feed_test for generate test data file

* removed unnecessary interfaces and add comments

* c++ style check

* update data_feed.cc

* AsyncExecutor: C++ side

Google naming conventions

Rename MultiExecutor to AsyncExecutor

pybind with async_executor

Naming convention

remove some flags and unused code

add refactored file of async_executor and data_feed

clear async executor interface and add data feed factory

split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto

Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch

refine async_executor_refactor.cc

add some files about datafeed

Revert "add some files about datafeed"

This reverts commit 8ee8133a.

add MultiSlotDataFeed

Interface rework

Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side

update datafeed and async_executor to run bow_net demo

update async_executor.py for support set_use_slots

Fix bug

update async_executor.py for support set_use_slots and set set_dense_slots

fix bug that when the number of files is less than the number of threads, it will fetch nan

remove redundant code, and make executor exit when set a illegal queue size

add MultiSlotDesc

Revert "add MultiSlotDesc"

This reverts commit 2e72ebfa.

add some checkpoint in DataFeedDesc

Fix fetch variable

fix code style in async_executor

Fix fetch var

add utest for data_feed

Add comments for python

update utest for data_feed for multithread

fix bug in utest of data_feed

Add comments for python code

Fix pybind.cc with new pybind11 version

add note for DataFeedDesc's set_use_slots function

update data_feed_test for multi-type

Add save_model

update data_feed_test for generate test data file

removed unnecessary interfaces and add comments

add comment for executor_thread_worker

Remove unused code

update data_feed.cc

c++ style check

* commit for code style

* commit for code style

* commit for code style

* commit for code style

* Comment away __init__ in async_executor.py

* clang-format fix test=develop

* use PADDLE_THROW instead of exit(-1); use unique_ptr to manage scope var in data_feed_test.cc

* commit for update code style

* commit for update code style

* Add async_executor demo; Remove some methods
test=develop

* commit for update code style

* commit for update code style

* commit for update code style

* update API.spec

* AsyncExecutor
test=develop

* AsyncExecutor
test=develop

* AsyncExecutor
test=develop

* AsyncExecutor
test=develop

* Fix API.spec
test=develop

* Fix API.spec
test=develop

* Fix windows build error
test=develop

* FIx windows build error
test=develop

* FIx windows build error
test=develop

* FIx windows build error
test=develop

* Fix Windows Build
test=develop

* Fix Windows Build
test=develop

* Fix Windows Build
test=develop

* Fix code style
test=develop

* Fix code style
test=develop

* update datafeed

* Fix code style
test=develop

* update data_feed_test for test Tensor test=develop

* Fix code style
test=develop

* Fix windows build failure
test=develop

* Fix code style and windows build failure
test=develop

* Fix PYTHON3.5 build failure
test=develop

* AsyncExecutor API
test=develop

test=develop

test=develop

* add recordio support

* disable the openblas multi-thread on windows since no support
adjust the python script

* code style

* code style
test=develop

* add create_recordio_file_reader back

* fix code style
test=develop

* fix the gtest.cmake on windows

* fix cc_test on windows

* fix the win build
test=develop

* remove fused compile support on windows
test=develop

* add the jit support
test=develop

* add the jit support, test=develop

* add the jit support, test=develop

* add the jit back
fix compile error on windows

* rollback test=develop

* test case fix

* disable DSO by default on windows

* exclude warpctc_op on windows

* exclude the dynload_warpctc out on windows
test=develop

* fix the scripts error
test=develop

* disable avx on windows by default
test=develop

* re-organize the cmake file

* disable mkl on windows by default

* add warp_ctc back

* fix the dependency

* fix the dependency

* fix the build issue on windows

* remove unsupported flag on windows

* code style

* code style
test=develop

* fix issue

* add profiler, parallel_executor back

* clean up the pre-definitions on windows

* fix build issue

* test=develop

test=develop

test=develop

test=develop

* fix input<tensor>
test=develop

* fix split_ids
test=develop

* ElementwiseMul should not support SelectedRows

* fix scale op
test=develop

* change GetTensorFromVar() method to GetTensorOrSelectedRowsFromVar()

* fix operator

* refine MultiOutput

* fix MultiOutput
test=develop

* disable test_dist_save_load
test=develop

* fix elementwise_op
test=develop

* add get_sparse_as_op
test=develop

* add info for check
test=develop

* rename get_sparse_as_op with extract_rows_as_op.
test=develop

* elementwise doesn't support selected_rows

* fix regularizer

* remove extract_rows_as
test=develop

* fix ci
test=develop

* add test for sum_op

* fix regularizer
test=develop

*  test=develop

* fix pserver weight decay multi inputs test=develop

Fix code to support cpplint syntax check

test=develop

test=develop

Dry run tests will skip `Op.Run` and just perform job scheduling. It helps to analysis dead lock in PE.

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

* wip

* add ref_by_trainer_id op

* ready to test

* fix ref inputs

* refine rpc_op_handle

* fix merge bug

It seems that the blocking queue might be destroyed early than Run
method complete. It might because the Run method throw some unhandled
exception. However, it should be shared_ptr when multthread access an
resource. So change BlockingQueue as a shared_ptr.

test=develop

test=develop

* fix nccl2 lars dist support

* put lars in momentum op

* add tests lars

* fix ci

* fix cpu kernel

* soft warning

* remove lars in test_recognize_digits.py

* move to another op

* add file

* update api.spec test=develop

* update test=develop

* fix api.spec test=develop

* wip

* wip, finish grad merge ops

* wip, finish graph build

* wip test running

* work on 1 gpu

* workable version

* update

* fix tests

* fuse broadcast op

* fix compile failed

* refine

* add batch merge test mnist

* fix CI test=develop

* fix build

* use independent bn params for batch merge test=develop

* update api.spec

* follow comments and for test

* wip

* refine tests test=develop

* follow comments test=develop

* remove startup bn modify test=develop

* follow comments test=develop

* fix merge test=develop

test=develop