* update squared_mat_sub_fuse_pass ut

* update ut

* update ut

* rename cuda to gpu

* revert CMake change

* resolve conflit

* rename other cuda to gpu

* poish details

* relu forward opt

* add gelu functor

* optimize code

* add seqpool_cvm_concat_fuse_pass ut

* rename ut name

* fix when out_dtype is same with x.dtype and still transform type error

* fix spell error

* mean first version

* fix scalar mean

* add fp16 dtype for api

* remove eigen and blas dir

* fix declare error

* add mkldnn conv_transpose_bias fuse pass ut

* update conv_transpose_bias_mkldnn_fuse_pass ut

* update conv_transpose_bias_mkldnn_fuse_pass ut

* update conv_transpose_bias_mkldnn_fuse_pass ut

* restrict conv2d data_format in conv_transpose_bias_mkldnn_fuse_pass

* update ut timeout setting

* update ut

* add pten conj kernel

* modify conj_kernel file path

* add defined cuda macro to cuda/conj_kernel.h

* support FP16 for more ops

* add amp list tests

* refine reduce_mean_grad

* fix OP benchmark ci

* fix fp16 reduce_mean

* updat ut, but still have some problems

* remove mean/reduce_mean fp16 kernel

softmax_with_cross_entropy optimization with soft label. This PR includes optimization of
    "SoftmaxWithCrossEntropySoftLabel" : compute log_softmax and then compute loss.
    "CrossEntropySoftLabel" : compute loss with softmax as input.
These optimization includes following technics:
    read data to buffer with vectorization
    compute max and sum in warp
    fixed loop size with macro
Performance (computation time):
    softmax_with_cross_entropy_0 (forward) : -40.1%
    softmax_with_cross_entropy_0 (backward): -41%

* add matmul_scale matmul_v2_scale fuse pass

* add scaletensor judge

* modify var name

* add timeout notest;test=coverag

* fix error commit

* fix use_mkldnn attr

* fix use_mkldnn attr

* add multi_tensor for momentum and clear_grads for optimizer

* fix bug for dygraph

* add unittest

* refine comment

* add param_group

* refine regularizaiton logic

* del clear_grads

* add clear_grads

* add dispensable check of None

* refine clear_grad

* fix build bug

* refine code by comment

* refine code

* add multi tensor check

* refine param_group update

* add multi tensor for static mode

* refine comments

* delete useless comma for momentum

* refine comment for momentum

* refine code by commment

* fix bugs when run reshape

* fix ci bug

* add complex op and `paddle.complex`.

* support more eager tensor api

* support multiple constructor for eager tensor

* add place related code

* polish code

* specific randint with dtype of int64

* Support pure cpu test

* refine test in pure cpu

* refine test in pure cpu

* support multi precision update for LAMB

* hide some api

* fix ci uts

* fix lamb output of dygraph

* remove some changes to some PR

* try to fix Py3 CI compile error

* fix test_imperative_optimizer, add lars ut, add layer_norm ut

* fix ut, fix format

* fix ut

* fix windows ci

* modify sum mean args

* add GetExpectedPtenKernelArgs for redcue_op

* modify kernel args number

* modify kernel args number

* Rearranged Eager AutoCodeGen directory structure

* Removed USE_OP in Eager AutoCodeGen

* Enabled generation for Operators without Grad/Inputs/Outputs

* Resolved operators without input

* Fixed merge conflicts

* Enabled Eager AutoCodeGen for 10+ more operators

* Refactored Eager AutoCodeGen with more organized helper objects

* Enabled Eager AutoCodeGen for operators with multiple OpBases

* Adjusted Eager AutoCodeGen to Enable Passing Output Tensor as Input Argument

* Handled Dispensable Inputs/Outputs in Eager AutoCodeGen

* Adjusted function generation/call between Python-C API & Dygraph API

* Synchronized auto-generated Python-C API with Dygraph Forward Functions

* Generated CoreOpsInfos for potential use in append_op API

* Fixed CI problem

* Inspect the information inside a TRT engine.

* Follow up the google code style.

* Fix code error.

From --ptxas-options=-v, SegmentOpsKernel uses 66 registers in a block.
There are two ways to resolve this problem:
    Reduce the threads per block launch configuration
    add __launch_bound__ to give information to nvcc compiler for reducing registers usage
this PR chooses __launch_bound__ solution because changing gpu_launch_config may affect other ops.