* Optimize performance of depthwise_conv_bwd

* fix

* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* add bfloat16 to selu_grad to pass CI

* fix selu grad compilation error

* Move valid check from python to kernel

* fix error throw

* fix

* invalid label check

* fix

* Revert "fix"

This reverts commit 79fad6799cfa4b30423dbc84e67d7d843d22b84a.

* Revert "invalid label check"

This reverts commit 402a9707390ad5386b3222e85844b92d2e9b9fa4.

* Revert "fix"

This reverts commit 09ba3080ee0587447f875c19cdf060485f15ae3b.

* Revert "fix error throw"

This reverts commit a901bfcc2179d5c120ec29af766f392b122dab52.

* Revert "Move valid check from python to kernel"

This reverts commit baa03cc4ef82d8d45516c30dfb52bf5aead30748.

* final fix

* fix

* fix

* fix P40 topk: Make the optimized topk compatible with P40.

* fix P40 topk: Make the optimized topk compatible with P40.

* fix P40 topk: Make the optimized topk compatible with P40.

* remove needless using tensor

* remove needless using tensor

* resolve conflict

* replace tensor using

* fix format error

* revert needless changing

* fix rocm and npu compile error

* fix cinn compile error

* fix format error

* fix mkldnn format error

* fix mkldnn format error

* fix cinn compile error

* fix cinn compile error

* fix cinn compile error

* resolve conflict

* Optimize performance of depthwise_conv_fwd

* fix

* Optimize topk's performance when k is small and input_width is large

* 修改blockdim设置逻辑

* Update top_k_function_cuda.h

* add fp16 support

* update

* update half

* code format

* fix unittest

* fix rocm compile error

* code format

* code format

* fix rocm compile error

* fix rocm compile error

* use cinn in the paddle inference

* fix some cmake errors

* Avoid division by zero in the arange_kernel.

* Avoid dynamic ops.

* Remove some useless codes.

* Use OpTransInfo to encapsulate some codes used in the build_cinn_pass.

* fix linspace error in amp

* fix log

* fix amp error

* fix ocr error which caused by amp

* add more check

* rename dtype ns

* polish code comments

* polish data_device_transform.cc

* move sum

* fix ci bugs

* fix ci bugs

* fix set_lod bugs

* fix infershape bugs

* fix ci bugs

* fix ci unittest bug

* fix ci bugs

* perfect code

* update code according comment

* add unittest

* fix ci bugs

This reverts commit c252b1de.

* fix return order error and duplicate results with specific inputs

* perfect softmax functor

* fix compile bugs

* fix ci bugs

* support int64 non-broadcast

* support broadcast case for int64 index

* fix bug

* support more Arity

* remove some codes

* upgrade patchelf to v0.15.0 to pass CI build

* fix bug

* fix patchelf installation

* add debug flags

* remove useless codes

* fix viterbi_decode and set_value op uts

* remove always enable int64

* support bfloat16 for amp_decorate

* add check_finite for bf16

* fix bug

* add ut

* add ut

* refine code

* add softmax infer kernel

paddle::platform::CudaAtomicAdd
https://github.com/PaddlePaddle/Paddle/issues/45881

* simplify size op

* trans to cuda manuly

* fix copy error

Adapt tensor output_size for conv2d_transpose and depthwise_conv2d_transpose

* fix amp kernel

* update to remove PADDLE_WITH_XPU macro

* fix grad error of grounorm op when cuda version==11.7

* add dist cuda kernel

* reuse some funcs in phi

* 使用pnorm

* fix code style - explicit

* fix code sytle

* fix bug

* remove unused headers