* Add Seq2BatchFunctor, which will be used in WarpCTCOp.

* Implement WrapCTCFunctor and WrapCTCKernel.

* Add unittest of warpctc_op.

* Modify the check_output inferface in python unittest framework to allow check a subset of outputs.

* Use absolute offset lod in warpctc_op and related functors.

* Refine the comments of warpctc_op.

* The new python unittest supports checking a subset of the outputs, so revoke the previous change.

* Rename the transform from LoDTensor to Tensor with shape [max_sequence_length, num_sequences, sequence_width] to PaddingSequenceFunctor.

* Update to the newest codes.

* Rename the PaddingSequenceFunctor to PaddingLoDTensorFunctor and remove the computation of dimensions out of the functos.

* "change operator interface"

* "move devicepool to device_context"

* "fix operator test"

* "fix op_registry Run interface"

* "net op passed. Need to fix nccl multi-Context"

* "add nccl group function"

* "add nccl group function"

* "fix gpu count exceed 32 error"

* "fix recurrent op, nccl op"

* "change the other operators interface with Place"

* "fix typo"

* "fix pybind"

* "fix device in python side"

* "fix pybind failed"

* "add init for test"

* "fix CI"

* Add HasCUDNN to detect if CUDNN is installed or not

* Fix CI

* fix compile on cudnn7

* update

* update

* make silent

* make enforce a target and dependent on nccl when gpu is enabled

* add some more dependency

It caused by a bug of std::call_once described in https://stackoverflow.com/questions/41717579/stdcall-once-hangs-on-second-call-after-callable-threw-on-first-call. It is likely caused by a deeper bug of pthread_once, which is discussed in https://patchwork.ozlabs.org/patch/482350/

* init cudnn batch norm op

* rename batch_norm_cudnn_op.cc batch_norm_op.cu

* correct name style

* add ExtractNCWHD, simplify code

* fix ExtractNCWHD

* use CUDNN_ENFORCE instead of PADDLE_ENFORCE

* "add nccl enforce"

* Dev

* Update comment

* Add nccl test

* Follow comments

* 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.

* Should always use `dynload::` for cuda function.
* Fix cublas.h without DSO load.