Do not invoke GPU method when use_gpu=false

Refine ExpandConvLayer

Enhance unit testing framework for the operators with LoDTensor.

Add the check of inputs and outputs in all operators, to avoid SegFault.

unify functions of mkldnn_fc and refine comments

Delete unused `USE_OP` in combination op

Fix Clang compile error

Host and device transform API

Cond_op with dynamic if-else checked-in

Fix enforce test failed and demangle stack info

Using LoDTensor instead of Tensor in every operator.

Add MKLDNNConvLayer

auto generate USE_OP() in pybind.cc