In cupti samples, only cuptiFlush is used.
I can't find any places calling cuptiFinalize and
this API can error out as not_implemented in some
cuda installation.

* Add MKLDNN layout support in Paddle

Add MKLDNN layout in Paddle so that MKLDNN friendly memory layout
can be used in MKLDNN enabled OP kernel. Before this commit, NCHW
is hardcode to be used in all MKLDNN op kernels. As a result,
non-optimized execution path is selected in MKLDNN primitive which
bring worse performance.
Besides framework change, three MKLDNN OP kernels were updated
for using new MKLDNN layout. They are conv/pool2d/batch_norm.
Other MKLDNN OP kernels need be also updated in similar way to
achieve best performance.

* Add MKLDNN layout support in activation OP

* Don't populate layout from input to output when kMKLDNN in

* Refine pool mkldnn op kernel

* MKLDNN layout

* Remove the inferitance from tensor file

* MKLDNN layout: refactoring

* Remove additional #define to register new operator

* Prepare mkldnn tests to work with layout

* Enable assertions in CUDA.

* Fix PADDLE_ASSERT.

* "fix deterministic"

* "fix ci"

* "fix init"

Sometimes dev_ctx is not available when RecordEvent.

    Experiment on vgg flower, 2 trainers, 1ps.
    more trainer could have more speedup.

    After:
    Pass = 0, Iters = 327, Speed = (7.52) img/s
    Before:
    Pass = 0, Iters = 385, Speed = (6.77) img/s

- Finished draft of pooling reusing of operators

- Using gethash in PoolGrad added

- Removed diagnostic

- Added pool mkldnn grad reusing of primitives

- Added diagnostic

- Removed diagnostic

- added dependency to mkldnn data type for pooling mkldnn

- Added mkldnn memory data type determining based on template type of op

- Compilation warning fix

- codying style fixes

In multi-thread condition, EnableProfiler can
be called after RecordEvent is constructed. In this
case, RecordEvent constructor will not init anything,
but RecordEvent destructor will do something since EnableProfiler
was called.
This PR fixes it.

* Fix the bug when a input variable of op is dispensable.

* Add HasInputs/Outputs interfaces to OperatorBase.

* Remove the unreferenced header file.

* fix __shfl_down_sync_ of cross_entropy

* use reduceSum

* "fix ci"