* add reducer

* refine envent for memorycopy

* add concat&split for allreduce

* apply concat & split for fuse tensor

* fix nccl dep

* fix the untest, compile problem and ddp initialize problem

* fix untest for mac & add some comments & solve the repeated param in sublayers

* fix untest for windows & fix document

* remove -DSUPPORTS_CUDA_FP16 in cuda.cmake

* comile with cuda9

* add some unittest

* notest;test=coverage

* add unittest for trt plugin swish && split

* update ernie unittest

* fix some error message

* remove repeated judgement of CUDA version in mbEltwiseLayerNormOpConverter

* fix comile errror when CUDA_ARCH_NAME < Pascal"

* fix comile error

* update unittest timeout

* compile with cuda9

* update error msg

* fix code style

* add some comments

* add define IF_CUDA_ARCH_SUPPORT_FP16

* rename IF_CUDA_ARCH_SUPPORT_FP16 to CUDA_ARCH_FP16_SUPPORTED

* default not show cpp statck & add hint

* fix failed unittest

* fix failed unittests

remove eigen threadpool for the speed up

* change avg pooling and global pooling to trt layer

* add support for static shape global pooling

* modify trt errmsg

* adjust kunlun header file
*test=kunlun

* update kunlun unittest
*test=kunlun

* update xpu unitest
* test = kunlun

* update xpu unittest
* test=kunlun

* update xpu unitest
* test=kunlun

* fp16 result ok

* change -DWITH_NVINFER_PLUGIN toconfig.EnableTensorRtOSS

* auto detect special slice op converter for ernie with trt oss

* ernie oss only support fp16

* fix special_slice_plugin serialize bug

* matmul in tensorrt ok

* ernie unittest ok

* add matmul tensorrt unittest

* remove demo code

* Add rnn_op.
test=develop

* Fix rnn_op grad maker's drop_empty_grad.
test=develop

This PR is follow up of #28213. On that PR we tried to decrease GPU usage, however the CI still randomly failed. So I added retry logic for the initialization of nccl and cusolver. If the initialization failed, we can retry to avoid the random failure.

* add multiple exception type

* define all exception & polish compile pystack

* mapping paddle error to python exception

* polish static mode error format

* fix failed unittests

* fix dytostatic test_error

* fix check_nan_inf failed

* add unittest for coverage

* revert some code try to solve compile error

* refactor enforce & error change

* polish code & add unittest

* add compile limit for paddle enforce

* polish elementwise_op_function.cu.h

* fix failed unittest

* fix windows compile failed

* detail polish

* revert no type constructor

* conv dilated mkldnn support: forward and backward pass

* add mkldnn conv_transpose dilation UT
test=develop

* remove unnecessary PADDLE_ENFORCE

* add int8 and bf16 dilated conv UT

* update according to reviews

* use exhaustive_search for float16

* tune algo only when dtype is float16

* - condidate fix to issue #25537

test=develop

* - UT for transpose NHWC

test=develop

* Add avx instructions check

* Small fix

* Change function name

* Change uint to unsigned int

* test=develop, optimize geo communicator 

* add gloo initializer, test=develop

* Add support for mkldnn ops types selection with FLAGS in dygraph

* use regex to match DNNL verbose

* python3 encoding fix

* include ncclRecv and ncclSend, test=develop