* Fix softmax cuda bug

* Refine multihead log and softmax logic

* yolo_box OP add Attr(clip_bbox). test=develop

test=develop

* fix elementwise_pow bug on integer, test=develop

* use llrint to support elementwise_pow_grad, test=develop

* add some tests, test=develop

* revert grad functor, test=develop

* fc-dequantize squash
test=develop

* change according to reviews
test=develop

* change PADDLE_ENFORCE
test=develop

* add second test when fc-dequant do not fuse
test=develop

* change all related PADDLE_ENFORCE
test=develop

* rename paddle throw error macro, test=develop

* fix new error use case, test=develop

* polish cmake, test=develop

* add current directory to LD_LIBRARY_PATH, test=develop

* Add reshape int8 op

test=develop

* Change test to CPUPlace

test=develop

* Correct tests

test=develop

* add fake init for the trainer, fix large memory hold in the trainer
* do not merge recv vars from a remote endpoint, test=develop
* add recv and save op, merge slice var in one op, save memory
* remove hsigmoid with pull sparse, test=develop

test=develop

* make OperatorWithKernel::InferShape virtual, test=develop

* fix test_prepare_op by relu, test=develop

* add no_need_buffer_slots interface to pybind

test=develop

* refine dygraph dataloader & polish related code, test=develop

* refine code based review comment, test=develop

* add seed op

* MKLDNN v1.0 rebase to Paddle 1.6
test=develop

* Add hacky paddle::string::to_string() implementation

* vectorize<int64-t>() -> vectorize() cleanup
test=develop

* PADDLE_ENFORCE and void_cast fixes
test=develop

* Rebase changes
test=develop

* Cosmetics
test=develop

* Delete MKL from mkldnn.cmake
test=develop

* CMake debug commands
test=develop

* Delete MKLDNN_VERBOSE and rebase fixes
test=develop

* Rebase fixes
test=develop

* Temporarily disable int8 resnet101 vgg16 and vgg19 tests
test=develop

* Add libmkldnn.so.1 to python setup
test=develop

* Add libmkldnn.so.1 to inference_lib cmake after rebase
test=develop

* Post rebase fixes + FC int8 changes
test=develop

* Fix LRN NHWC
test=develop

* Fix NHWC conv3d
test=develop

* Windows build fix + next conv3d fix
test=develop

* Fix conv2d on AVX2 machines
test=develop

* ZeroCopyTensor::mutable_data in the right device, test=develop

* add unittest for zerocopy, test=develop

* fix master patch when slot is dense
* test=develop

* fix code style of fleet_wrapper
* test=develop

* accelerate mean op test=develop

* add op function generator, test=develop

* add unittest, test=develop

* follow comments, test=develop

* fix windows compilation problem, test=develop

* update benchmark for int8v2, QAT1, QAT2 accuracy and performance
test=document_fix

* change according to reviews
test=develop test=document_fix

* improve some descriptions and some models
test=develop test=document_fix

* update models benchmark data
test=develop test=document_fix

* update int8v2 and qat2 performance
test=develop test=document_fix

* refine init function, test=develop

* add tests, test=develop

* remove extern, which may cause symbol error in gcc-4.8, test=develop

* dygraph_grad_maker supports varbase without grad_var, test=develop

* fix compile, test=develop

* fix test_tracer, test=develop

* follow comments, test=develop

* remove optimize_for in framework.proto
* test=develop

* polish infer shape registry, test=develop

* modify some operators registry, test=develop

Add tests to use dy/dx to make sure the gradient values calculated by the control flow backward is correct. Also fixed bugs detected by those tests.

Fix bugs:

1. Unlike sum_op, optimizer ops don't allow uninitialized input tensor. But in conditional_block_grad_op, since the conditional_block may not run, the output gradient tensor may be uninitialized, which will cause the optimizer op error. To fix it, we should let optimizer ops support uninitialized input like sum_op or assign the uninitialized gradient to 0 when the conditional_block_grad_op doesn't run. I found there are about 10+ optimizer ops. **To be simpler, I just assign output gradient of the conditional_block_grad_op to 0 in this PR**. But it can be further explored whether we can make optimizer ops like sum_op to support uninitialized input tensor because theoretically we can speed up without the assigning in conditional_block_grad_op.

2. Infer parameter shapes during append_backward. I didn't know that all our parameters are in global block. When op_desc is inferring shapes at the sub-block, it may not know the shape of gradients of parameters whose shape information is at global block. I fixed it by inferring shapes of gradients from forward var.

This PR also did some code clean up:
1. Print the var name when sgd_op catches shape error so that it is easier to debug
2. Fix a typo: dicta -> dict

* Add a single_process_multi_thread transpiler.
* Add some UTs.
* Fix some API description.

* add file check_op_desc.py and add interface to get default value. test=develop

* add test for c++ coverage rate. test=develop

* Correct typo. test=develop

test=develop

* - BAtch norm mkl-dnn NHWC

test=develop

- compilation fix

test=develop

- UT fix

- cosmetics

test=develop

- Fix to Batch Norm MKL-DNN NHWC UT

test=develop

Conflicts:
	paddle/fluid/operators/batch_norm_op.h

* - Lint fixes

test=develop