* add reduce_sum_op by add self-kernel

* set all ReduceKernel MPType for accuracy

* add float16 test script which input is integer number

* solve reduce sum float16 check_grad problem

* solve conflict and change test script for CI

* change kernel register for CI

* remove all useless template

* Add digamma_op and unittest

* add digamma_op api

* remove special DigammaCudaKernel and correct some docs

* remove unused headers

* fix api doc error

This reverts commit b2afc8df.

* add warning log for DataLoader output format imcompatible upgrade. test=develop

* Eliminate numerical differences of LayerNorm; fix LayerNorm Nan Bug while large data input

* fix bug while large shape of data input

* add support for trt dynamic shape flatten op

* add version restriction

* add ut input dynamic shape

* Small changes related to BF16 fusion_gru and fusion_lstm

* Correct to pass arg by value

* Add conditions to rnn op

* Correct the spelling mistake

* Improving the test with checking activation

* Trigger CI

* fix gather shape bug

* fix None

* fix topo

* fc_elementwise_layer_fuse_pass

* fc_ele_layernorm_pass

* fc_elementwise_layernorm_pass

* fc_elementwise_layernorm_pass_amend

* - Second fix

- fix

* - fix

* update spawn doc for xpu, test=ducument_fix

* add note for gpu and xpu, test=document_fix

* use PYTHON_C_API in dygraph, test=develop

* update 2.0 public api in all left files

* reverse device.py all list;
fix some flake8 errors

* update 2.0 public api in vision

* fix some flake8 errors

For python, if users write `a, b = var`, the `__getitem__` method will iterate through 0, 1, 2 ... until `__getitem__` throws an IndexError, then stop. The var[0], var[1] will be given to a, b respectively. If more values are given, the unpack size would cause error.

We didn't raise the IndexError in the past and we add statement in `__getitem__` to raise IndexError here to support grammar like `a, b = var` in this PR.

* fix Undefined variables

* fix Undefined variables

* 增加方法获取和执行全量api的示例代码进行测试

* start the sample code test for gpu

* should import paddle separately

* add a stdout handler, the default is stderr. the paddle_build.sh will catch the stdout content.

* add RUN_ON_DEVICE into the requires set

* if codeblok['required'] is empty, use the RUN_ON_DEVICE instead

* set the threads to 16

http://agroup.baidu.com/paddlepaddle-org-cn/md/article/4036225

* 设置默认日志级别为INFO级别

* using the logic from gen_doc.py

* using modulelist to get the all apis

* as we don't care which name is the shorttest, so fetch the first name in the all_names list

* the new list from project

* 先不启用gpu测试，先把print_signature获取不到全部API的问题解决了

* add wget option in download

Change-Id: I4e09e7710f6693bff5388983270781a4ef70519e

* add check nan of inf for dygraph

* add unittest for dygraph

* revert error change

As the title

* support diff dataset tensor place in single process dataloader

* fix unittest failed

* fuse L2Decay and momentum when param.regularizer is set

* add unittest

* refine

* refine _create_regularization_of_grad of momentum

* improve append_optimizer_op