* add anakin subgraph engine

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* add initial op converter

* update

* update

* fix op register compile error

* update
test=develop

* update

test=develop

test=develop

test=develop

* staged.

* polish code

* polish code. test=develop

* polish code. test=develop

* api change. test=develop

* fix default value. test=develop

* fix default value. test=develop

* Rewrite is_empty op to avoid unnecessary data transform.
test=develop

* Add the implementation of InferShape and InferVarType for is_empty op.
test=develop

* Rewrite is_empty op to avoid directly inherit OperatorBase.
test=develop

* Optimize key creation of INT8 pool kernel to improve the peformance of ResNet-50 and MobileNet, especially for latency.
test=develop

* Optimize key creation of pool fp32 grad.
test=develop

test=develop

test=develop

- MKLDNN ops revisited

- disabled softmax modifications

- disabled elementwise_add

- reverted LRN modifications

- reverted SUM primitive

- Partial reviing of softmax

- Enable softmax

- Softmax changes

- LRN is back

- LRN partially disabled

- LRN is back

- LRN fix

- compilation fixes

- Sum fixed(hopefully)

- Enabling (partially) elementwise_add

- Fixes to elemenwise_add

- Lint fixes

quantize fix

- compilation fix

test=develop

Disabling pooling

- Disabled quantize op

test=develop

* loosly check in cross_entropy_op when soft_label is True
* Add Runtime assertion in backward infer_shape check.
* Skip InferShape check when un-know the input dimensions

test=develop

test=develop

Optimize the CUDA implementation of sequence_expand op by reduce the times of copying lod data from CPU to GPU. (#15493)

* Optimize the CUDA implementation of sequence_expand op by reduce the times of copying lod data from CPU to GPU.
test=develop

* Refine the op benchmark to support setting lod in config.
test=develop

* This PR improve performance of prior_box op about 1.25x faster on CPU.

* Test Env:SKX 8180 with fake data on 28 threads(bs=1).
* The below table shows the ~25% improvement which generated by [eval_tp_fake_data.py](https://github.com/PaddlePaddle/Paddle/issues/15618#issuecomment-464613976).

| Type |Event | Calls |   Total     |  Min.    |   Max.      |  Ave.      |  Ratio.|
| ---------------- | ------------------ | ---- | ------- | -------- | -------- | ------------ | -------- |
| w/ optimization  | thread0::prior_box | 6000 | 921.201 | 0.110572 | 0.383402 | **0.153533** | 0.084585 |
| w/o optimization | thread0::prior_box | 6000 | 1151.85 | 0.102276 | 0.426702 | **0.191976** | 0.103337 |

test=develop

* Fix the style issue.

test=develop

* add alloc_continuous_space_op
test=develop

* Polish code
test=develop

* follow comment
test=develop

test=develop

test=develop

test=develop

test=develop

* refile profiler
test=develop

* follow comment
test=develop

* Refine doc of uniform_random and fix dtype
* Update defaule value in the arguments

* - Implemented draft of primitive desc keeping in Tensor

test=develop

- TransposeMKLDNNHandler::AcquireSrcMemory was reimplemented

- Added nchw and nc formats setting for sake of compatiblity

Fixed unit tests

- Worakaround to problem with 5D data in conv

- Added 3D and 1D MKL-DNN formats for name handles for tensor

test=develop

- Fix to UTs

test=develop

- Conv fp32 op was updated

Cosmetic fixes

test=develop

- tensor mkldnn cosmetics

test=develop

- Moved most of mkl-dnn specific code from Tensor to mkl-dnn utils

* - Lint fixes

test=develop

* - setting prim dec in Tensor , sets also layout to kMKLDNN

test=develop

* - Moved creation of prim desc totally out of Tensor

test=develop

* - Cosmetic fixes adter review

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop