这个PR只包含fused_feedforward前向的代码。

相关kernel实现：fused_dropout_act_bias, fused_residual_dropout_bias, fused_layernorm_residual_dropout_bias

fused_feedforward是一个融合算子，该算子对transformer模型的feed forward层的算子进行融合和封装，使得前端只呈现一个接口，通过融合减少部分访存和kernel launch的时间，以此提升性能。

* slice

* add UT

功能：本PR的目标是提高attention模块的计算性能。
为了减少框架层对op的调度开销，本PR通过在C++层手动实现attention模块，对外提供attention 大op；
为了减少防存开销，本PR采取了两种优化方法：
（1）在q,k,v计算时通过共享输入X，将该处的gemm，transpose和bias add从三次调用减少为一次；
（2）使用kernel融合优化技术，在不同cuda kernel之间通过寄存器传输数据；

* [hapi] support dygrapg amp O2

* fix problem of static pure fp16 in hapi

* fix bug

* fix format

* fix ut

* follow comments

* update ut

* update amp save/load

* fix ut

* refine code format

* add inceptionv3
Co-authored-by: Ainavo <ainavo@163.com>
Co-authored-by: Npithygit <pyg20200403@163.com>

* added base changes for matmul_v2+trans+resh fuse pass

* added full matmul_v2+transpose+reshape pass

* removed a file added by mistake

* added reviewers suggestions

* Changed ops type in checking capatibility version

* Deteled one statement

* add sync_batch_norm (support train, infer, and fp32, fp16, and NCHW, NHWC)

* [NPU] Delete debug codes

* [NPU] Remove FP16

* add viterbi decode cpu kernel

* add viterbi decoder api in paddle.text

* add a data buffer once to avoid create many small pieces of data buffer frequently

* fix viterbi max_seq_length bug

* fix seq_len=1 bug

* fix device context

* move split out of for loop

* remove INVERSE_SUB

* remove 2 GET_CAST_MASK

* remove 1 loop

* remove Functor

* add to_static deploy code

* use MAX_FUNC instead of ELE_MAX

* add MaxFunctor

* impl max_func

* remove MaxFunctor

* remove cast op

* use REGISTER_OP_WITHOUT_GRADIENT

* add viterbi cuda kernel

* add FIX_BLOCKDIM_CASE macro

* add MKL add, mul; add get data mask

* add arange mkl impl

* add CPU Argmax

* add cpu gather

* use EXECUTE_MKL_ELEMENT_BINARY_OP instead of some ADD, MUL

* use SameDimsBinaryOP instead of EXECUTE_MKL_ELEMENT_BINARY_OP

* use SAME_DIMS_ELEMENT_BINARY_OP

* add SimpleBroadcastBinaryOP

* use int instead of int64_t to accelerate

* optimize SimpleBroadcastBinaryOP

* optimize SimpleBroadcastBinaryOP

* optimize performance in both single thread and multithread situation

* remove useless line

* remove useless code

* add CREATE_TENSOR_BUFFER macro

* add INIT_REQUIRED_TENSOR macro

* add comment

* fix windows ci

* add viterbi unittest

* remove cuda add functor

* remove cuda equal

* remove a template function

* fix windows ci

* fix windows dtype

* remove some template instance

* remove useless header file

* remove some blockdim

* remove transpose impl

* accelerate cpu performance on single thread situation

* viterbi_decode->crf_decode

* rename crf params name

* add viterbi api test

* remove useless import

* add enable_static

* use viterbi decoder

* fix viterbi len=1

* fix  viterbi unittest

* remove useless comments

* reconstruct viterbi decode

* remove ADD,SUB,MUL structure

* fix coverage

* remove CREATE_TENSOR

* add name args

* crf.py->ops.py; with_start_stop_tag->include_start_end_tag

* update crf_decode en docs

* fix viterbi decode en docs

* fix some review comments

* add FIXED_BLOCK_DIM_CASE in cuda

* push_back->emplace_back

* crf_decode->viterbi_decode; include_start_end_tag->include_bos_eos_tag

* paddle.text.ops.viterbi_decode->paddle.text.viterbi_decode

* fix viterbi_decode en docs

* add some ops to train ssd on kunlun

* update test_fill_any_like_op_xpu.py

* fix bugs of ClipGradByGlobalNorm

* add unittests

* add unittests

* fix global gather and global scatter operators

Add Tokenizer related functionalities for Transformer model in order that the process of training and predicting is consistent.

* support the text string as an input Tensor
* support the "VOCAB"unordered_map<wstring, int> as an input Tensor to lookup tokens
* Tokenizer used for BERT. This tokenizer applies an end-to-end, text string to wordpiece tokenization.
* It first applies basic tokenization, followed by wordpiece tokenization.

* native commit for triple grad of sigmod

* Updated unittests files

* init functional jacobian api

* Updated trible_test func

* Updated gradient_checker & test_script

* finish test with dtype float32

* add float64 test case

* polish code

* use atol=1e-5 with dtype float64

* fix for ci

* set timeout for test_jacobian

* fix dygraph grad to support high differential

* polish API docstring

* Updated gradient checker and some related files

* fix double grad strip error for high differential

* fix double grad strip error for high differential

* Add Sigmoid triple grad tests

* fix dygraph double grad dtype error when calling for high differential senario

* Updated triple grad teses func

* Use np.random to initialize ddx

* Updated triple_grad_check func

* add todo for gradient checker and refine some comments

* remove additional code

* add test for warnging in backward.py

* add tanh triple grad

* format python code

* refine code

* make test_activation_nn_grad test time to 150s
Co-authored-by: Nveyron95 <veyron_wu@163.com>
Co-authored-by: Nlevi131 <limaolin01@baidu.com>

* default dist op

* add dist_attr for dist op

* add unitest

* update inputname

* update function name

* add unitest

* update CMakeLists.txt for CI

* fix dis_matmul

* fix compile error

* update matmul to matmul_v2

* unify api

* unify api

* todo

* update distop forward func

* update distop forward func

* auto parallel backward

* update dist op

* autoparallel backward

* add backward for embedding

* temp1

* temp2

* temp3

* temp4

* backward done1

* backward done2

* backward done3

* dist embedding remove mp mode

* dist matmul remove mp mode

* update dist embedding
『

* dist op init1

* dist op init 2

* update unitest

* context remove parallel mode

* partitioner remove parallel mode

* update unitest

* a more general method to support varying mesh in pipeline parallel

* support varying mesh in pipeline parallel

* embedding support varying mesh in pipeline parallel

* matmul support varying mesh in pipeline parallel

* default dist op support varying mesh in pipeline parallel

* dist attribute for startup program

* default dist op support varying mesh in pipeline parallel 2

* partitoner support varying mesh in pipeline parallel

* revise logic for auto compeletion

* revise framework.py

* revise reshard unitest

* revise unitest for parallelize

* chmod

* fixed bug for dist embedding name mapping
Co-authored-by: Nzhaoyingli <zhaoyingli@baidu.com>

* remove no_value using var.name

* fix unit test for CI

* fix unit test

* add test case

* fix test case

* add more test case

* Support elementwise_add triple grad Kernel

* Change code-format to follow CI std

* [NPU] Add iou_similarity op

* [NPU] Add iou_similarity op

* [NPU] Add iou_similarity op

* update

* fix ut error

* update ut

* add nearest_interp_v2 trt plugin

* fix replicate pad when input size is 0
* add unit test

* catch the generatorfunction and intercept it.

* add test generator

* add test case

* refine the testcase

* Add QR decomposition op

* Change codes to adapt to new svd_helper

* Update linalg.py

Restore the deleted comma

* Restore the deleted line

* Update linalg.py

* Update linalg.py

* Improve the qr code by reviews

* Update QR based on CI results

* Update qr doc, test=document_fix

* Change unsafe and ill-formed codes

* Add auto parallel cost model and unittests

* Fixed code styles.

* Fixed bugs and codes style

* fixed typo

* Improved code style: object encapsulation.

* Fixed codes.

* Refractored estimate_cost

* Fixed typo