* add cpu version, using set: sum, min, max

* add cpu version: mean

* improve cpu code and fix dynamic memory allcation problem

* fix arg error, add index judge, delete fp16

* fix bug in CudaAtomicMax and CudaAtomicMin

* add CUDA version

* fix grad_op bug for index

* add op test, add correct cpu grad op

* Add correct CUDA Mean grad

* [Add] Successful MEAN and SUM

* [Add] Successful MIN and MAX in CPU

* [Add] Successful MIN and MAX in CUDA

* fix windows dtype ci

* fix ROCM ci by adding HIP flag

* rename fused_gather_scatter to send_recv

* unify name as send and recv

* change zero index return time

* add send_recv incubate api

* fix index data type, add unittest case for API

* delete redundant input tensor

* fix en example and docs, add default value in pool_type

* add shape judge and max grid judge

* fix comment

* fix index type bug

* add const &

* fix en docs

* delete numpy in examples

* add unittest for int input

* fix send_recv comment

* change send_recv to graph_send_recv

* add reshape method for Tensor

* fix typo

* fix typo

* fix conflit with develop

* fix

* ci rerun

* ci rerun

* ci Rerun

* optimize the data structure from c++ to python to speed up sampling in graph engine

* update test

* fix bug to support dropout eval grad computing.

* Remove useless code.

* elementwise_add kernel refactor

* fix compile bugs in elementwise_add refactor

* fix compile bugs when run in npu/xpu

* fix bugs when run unit test

* fix bugs when run ci-windows

* modify code as recommended

* code format adjust

* fix bugs when run ci

* fix compile bug when run in ci-windwos

* elementwise_sub refactor

* add PD_DLL_DECL for elementwise_sub

* fix bugs when compilei

* Add the `GetFetchNames` method in CinnGraphSymbolization.

* Use unordered_set instead vector as the type of fetch_var_names.

* Reuse the definition of kCompilationKey.

* Use CompileOptions to set fetch_var_ids.

* Update the argument passing of GraphCompiler.Build.

* Fix some bugs in CinnGraphSymbolization::GetFetchIds.

topk中有cub和手写kernel两种实现，而cub是通过排序来获取topk，通过多组数据发现只有当input_width>=128且k超过input_width 75%的时候性能会比手写的更好。

* Use oneDNN reorder instead of custom one

* Fix whitespace typo

* Fix Code format error

* Incorporating feedback

* Remove unncessary reorder

* Support GIOHW format

* Fix code format error

* init

* add feed ops in python side

* import LRScheduler

* update_feed

* refine code format

* Change first batch of mkldnn headers and namespace names to dnnl

* Revert changes to tensor.h, which require approval

* Format changes with pre-commit

* Add int32 tests

* Fix int32 tests and call GetDataFromTensor for int32

* Fix test

* Modify reduce_op.op.h for xpu2 with kernel primitive api

* add slice api impl of Tensor

* fix test slice error

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* refactor heter trainer. test=develop

* fix. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix ut. test=develop

* fix ut. test=develop

* fix ut. test=develop

* copy beta pow to same place when skip_update=1

* fix xpu

* add

* add BuildOperatorDependences

* fix bug

* add unittest for write after write

* fix merge bug

* fix

* Added BF16 Pool2d grad

* upstream pulled

* fix for CI

* fixes after review

* reshape kernel refactor

* fix compile bugs when run ci

* support xpu for reshape

* fix bugs when run unittest in kunlun ci

* fix compile bugs when run kunlun

* perfect code according to suggestion

* add api and unit test for reshape

Add pure fp16 support for fused transformer.

* Make FLAGS_determinstic effective in conv2d forward.

* Add call of SetCinnCudnnDeterministic in cinn_launch op.

fused_attention_op的实现中，使用了bias_add，且其实现是通过使用kernel primitive来实现的，之后kernel primitive的WriteData api接口及函数内部实现发生了更改，将判断越界的逻辑移到了template的参数中，使得调用的分支有错误，产生了越界赋值操作，污染了别的显存空间的内容。具体表现为：test_fused_attention_op_api.py 单次执行基本上不会报错，多次循环执行不同shape的输入，结果计算不对，具有偶发性，bug不易察觉。

* move extension into pten [no-verify]

* append tensor methods by ext_tensor [no-verify]

* append other tensor methods [no-verify]

* ext related files tidy [no-verify]

* include relation tidy [no-verify]

* add pten tensor test [no-verify]

* replace tensor in custom op & compile success

* refine tensor constructor for unittest

* custom relu jit run success

* fix all custom op unittests

* add inference cmake adapt [no-verify]

* fix failed unittests

* fix windows failed unittests

* try to fix kunlun and inference failed

* fix test_elementwise_api error

* try to fix win compile failed

* fix kunlun fp16 type error

* remove useless haddle error macro

* add custom linear op test

* fix compile failed & add win symbols

* fix non pten kernel cast failed

* add dll decl for api

* polish several deetails

* polish details by review comment

* add dll_decl for register