* use iwyu clean include second time, test=develop

* unify calling cudaSetDevice

* fix compile

* Refine PADDLE_ENFORCE in paddle/fluid/platform
test=develop

* refine PADDLE_ENFORCE
test=develop

* Add a StatValue class in the backend to represent a stat.
* Add a singleton StatRegistry to maintain the collection of stats.
* For the sake of code neatness, we only support type of int and float, which can cover most of the scenarios.

* Optimize the error messages of paddle CUDA API, test=develop

* fix the error messages of paddle CUDA API, test=develop

* Refactoring PADDLE_ENFORCE_CUDA_SUCCESS, and apply to curand/cudnn/cublas/NCCL,test=develop

* remove build_ex_string,test=develop

* merge conflict,test=develop

* add recorded cuda memory apis, fix typo, test=develop

* add more ut, test=develop

* follow comments, test=develop

* fix py35 incompatible issues, test=develop

stack's wait cost a lot of cpu time, use cuda kernel to do memory copy
will reduce cpu time.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* Improve topk performance.

give 200000 data to compute topk,
before opt: cost 1s
after opt: cost 0.0028s.

* Refine return value.
* Add cuda util funtions.
* Fix ComputeBlockSize bug & refine comments.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

enable cpu machine to run paddle model in gpu lib

* fix gpu_info, test=develop

* fix reserving gpu memory calculation bug, add fraction=1 unittest, test=develop

* fix bug again for reserving size, test=develop

GPU allocation uses fraction of available memory, also fix the GetUsed without lock

Optimize the content of error reporting information, print error code and official document web sites (#18671)

 optimize the error reporting information of cuda related API
 index on develop: 130ac177 Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into develop

test=develop

modify allocator strategy
remove changes of legacy buddy_allocator
test=develop

test=develop

* Refine the beam_search op and test.

* A basic CUDA implementation of beam_search for small batch_size.

* Implement CUDA kernel for beam_search_op.

* Use multiple CUDA threads in the same block to select the top beam.

* Update the python api of beam_search op.

* Enable extend function in CPU kernel of beam_search op.

* Unify the CUDA codes.
test=develop

* Unify the CPU kernel of beam_search op.

* Ensure the seletced items of beam_search_op's CPU kernel sorted by scores.

* Update the description of beam_search in API.spec.

* Enable the use of CUDA kernel in beam_search op.

* Exclude the beam_search's CUDA unittest when there is no CUDA gpu, and delete some debuging statements.
test=develop

* Follow comments.
test=develop

* Call the CPU kernel for beam_search op when batch_size > 4.
test=develop

* Remove the except of is_empty op in PrepareData.
test=develop

test=develop

* wip multi process multi gpu dist training

* workable for p2p

* update test=develop

* change back env name test=develop

* fix alloc init

* fix cpu build test=devlop

* fix mac tests test=develop

* refine code

* refine test=develop

test=develop

* refine cublase
test=develop

* code refine

* refine cublas

* add GEMME_EX

* add enable_cublas_tensor_op_math doc and add cublasCall
test=develop

* fix CublasCall for cuda version
test=develop

* fix error
test=develop

* fix GEMM_EX to be compatible with gcc 4.8
test=develop

* add GEMM_EX
test=develop

* to compatiable with gcc4.8
test=develop

Fix code to support cpplint syntax check

test=develop

test=develop

test=develop

This reverts commit 1d91a49d.

* some trivial opt

* remove the fix of lod_tensor and shrink_rnn_memory_op

* refine ShrinkRNNMemoryOp

test=develop