* [MLU]add mlu op interface

* [MLU]fix alpha of activation op

 add pull gpups sparse op

* add new API/OP:paddle.poisson

* fix comment

* First set of fixes

* - Make more likely to GetBlob find a blobs

* - Lint

* add erfinv API, test=develop

* fix gradient accuracy error, test=develop

* fix cuda compilation error on Windows, test=develop

* fix M_2_SQRTPI undeclared identifier on Windows, test=develop

* add empty and empty_like kernel in pten

* add empty dev_api

* optimize gelu backward

* optimize gelu backward

* optimize code

* Number to expression

* Replacement number

* move flatten

* fix bugs of test

* modify header file

* add copy declare

* fix compile bugs

* rename cuda to gpu

* revert CMake change

* resolve conflit

* rename other cuda to gpu

* poish details

* relu forward opt

* add gelu functor

* optimize code

* mean first version

* fix scalar mean

* add fp16 dtype for api

* 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%

* fix bugs when run reshape

* fix ci bug

* add complex op and `paddle.complex`.

* 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

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.

* fix_bugs_for_elementwise_branch_selection

* fix merge_dims bugs

* fix all influenced file

* Faster implementation of CPU kernel for ROI_ALIGN Operator

* Add missing variable to CUDA roi_align_op

* Style

* Fix boundaries

* Rename variables for indexes calculation

* Remove unnecessary emplace

* Revert "Remove unnecessary emplace"

This reverts commit c10e87f7fb812f1a672fde32f2690a97d47e2f5a.

* Style

* add activation

* update activation_op

* add unitest for activation

* fix acosh for init, test=develop

Revert "modify the fix_seed attribute in dropout op is a def attribute.test=develop (#38100)" (#38127)

This reverts commit f44add7b.

* add register_ctx_kernel and move scale kernel

* polish details by reviewer comment

* fix xpu compile failed

* fix cmake error

* trainer_device fix and checknan tool for psgpu;test=develop

* disable show_one_table;test=develop