* fix error message of l2_normalize, matmul, mean, etc. 
* add the test case for those ops

* Initial FP32 DNNL MatMul Implementation

* Implement int8 DNNL MatMul

* Unify in-kernel-naming, clean UTs

* MatmuL: Introduce op caching

* Final adjustments

test=develop

* Remove dy_graph disablement

test=develop

* Change dnnl header name to new one

test=develop

* Contrain multi head check to prevent fails

test=develop

* Resolve dnnl header problems on MAC CI

* Variable namings to kernel and skip_grad_ci added

test=develop

* Prevent MAC CI from failing

* Prevent windows build from failing

test=develop

* Modify UTs to conform to the rules

* Modify MatMul aux functions namings

test=develop

* delete invalid check inferface Ref & VectorRef, test=develop

* fix vector ref delete error, test=develop

* try the new check inferface, test=develop

* change all related code with new check macro, test=develop

* remove static assert, test=develop

* polish detail, test=develop

* skip coverage problem, test=develop

* add new check macro, test=develop

In the op of gemm, we use the gemm to replace batch gemm, speed up the matmul op 

* refine grad maker, test=develop

* refactor tracer stage 1, test=develop

* merge develop to solve conflict third times, test=develop

* refactor dygraph,test=develop

* fix failed unittest,test=develop

* polish code,test=develop

* check windows ci error,test=develop
try to fix windows ci error by np.allclose,test=develop

* polish vlog and profiler, test=develop

* try to fix preceding ops order,test=develop

* test transformer in windows ci, test=develop

* use python c-api to speed up tracer.trace,test=develop

* test=develop, fix docker with paddle nccl problem

* test=develop, add ut for debug string and gradient_accumulator

* test=develop, add tests for layer/gradient_accumulator/prepared_op

* test=develop, fix complie error for test_prepared_op

* test=develop, add more ut for dygraph

* test=develop, create API.spec for dygraph api change

* optimize grad maker; test=develop

* optimize grad maker

* test

* grad make optim; test=develop

* fix unittest bugs; test=develop

* add dygraph grad op maker and split_op

* grad op maker refactor; test=develop

* add dygraph grad maker; test=develop

* fix op deformable_conv_v1_op bug; test=develop

* fix deformable_conv prroi pool bugs;

* fix new op grad op maker bug; test=develop

* fix split by ref bug; test=develop

* fix dygraph auto prune bug; test=develop

* fix test_trace bug; test=develop

* fix fused emb seq pool bug; test=develop

* remove useless code in op_desc file; test=develop

* remove useless code, StrVarBaseNode; test=develop

* fix review issues; test=develop

* fix rank_loss grad maker; test=develop

* remove flag in VarBase; test=develop

* fix distributed_notify_op compile bug ; test=develop

* fix reshape op double grad; test=develop

* fix expand as op; test=develop

* add impertive type_defs.h for demo_train; test=develop

* fix inference lib cmake; test=develop

* fix inference lib; test=develop

* fix infernce_lib; test=develop

* fix inference cmake; test=develop

* fix inference lib; test=develop

* fix inference lib; test=develop

* remove condition dygraph grad maker, modify local name; test=develop

* fix split grad maker bug; test=develop

* fix pyramid_op bug; test=develop

* change travis time out limit; test=develop

* restore travis; test=develop

* change timeout limit; test=develop

* update the infer shape of matmul,  test=release/1.6

* add unittests of matmul, test=release/1.6

* change func names, test=develop

* add data type check, test=develop

* polish error messages, test=develop

* polish error messages, test=develop

* Remove support for the CPU architecture matmul, test=develop

* fix syntax bug, test=develop

* add support of matmul with multiple head even different width and height

Original matmul with multiple head supports only the mat_a.width == mat_b.height,
in that case, mat_b will be horizontally split. In this patch, we extend the
support when mat_a.width != mat_b.height but mat_a.width/head_number == mat_b.height,
in this case, mab_b will be vertically split.

One example is A is [3, 8], B is [2, 16], head_number is 4. In this
case, A will be split as [3, 2], B will be (vertically) split as
[2, 4]. The final result will be 4 matrix of 4 matrix of [3,4], i.e. [3, 16]

test=develop

* add support of matmul with multiple head even different width and height

Original matmul with multiple head supports only the mat_a.width == mat_b.height,
in that case, mat_b will be horizontally split. In this patch, we extend the
support when mat_a.width != mat_b.height but mat_a.width/head_number == mat_b.height,
in this case, mab_b will be vertically split.

One example is A is [3, 8], B is [2, 16], head_number is 4. In this
case, A will be split as [3, 2], B will be (vertically) split as
[2, 4]. The final result will be 4 matrix of 4 matrix of [3,4], i.e. [3, 16]

test=develop

* refactor the code of matmul with multiple head even different width and height

test=develop

* extend matmul op to support multiple head multiplication

With the support of multiple head, the multiplication of two big matrixes is
split into multiplication of several (head_number) small matrixes. e.g. if
Mat A is [3, 24] and Mat B is [24, 4], when multiple A and B with head_number
as 4, Mat A will be split as 4 matrix of [3, 6] and Mat B will be 4 matrix of
[6, 4]. The result of final matrix will be 4 matrix of [3, 4], i.e. [3, 16].

Do not use ctor

* Reduce line of codes.
* We can use virtual function for Maker now.
* The implementation does not care what maker holds, it is easier to
refactor later.

* script to add semicolon

* fix typo

* Fix cpplint errors, round2

* Fix pointer issue

* Move framework.proto to proto namespace

* Fix compile

* Fix compile

* Fix Compile

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`

* fix m_ops

* fix activation op

* initial matmul operator

Similar to np.matmul, but also has transpose_X and transpose_Y flags,
and only supports tensors from rank 1 to 3 inclusive.

For GPU, uses cublas?gemmStridedBatched. For CPU, uses
cblas_?gemm_batch if available via MKL; otherwise a simple serial
implementation that loops over the batch dimension is employed for now.